Compartment models, pipeline attrition, and capacity nowcasts for bioeconomic measurement

28 June 2026 · Revised 4 September 2026

Abstract

We treat disease incidence, biotech pipelines, and hospital occupancy as three clocks with different time scales. The epidemiological object is a SEIRD system with a death sink and, when counts are small, demographic noise. The pipeline object is a product of stage survivals, optionally replaced by a Cox model on public stop and go dates. The capacity object is a stock of occupied staffed beds driven by a lagged incidence map. The paper states the next-generation matrix, the basic reproduction number, the product formula for expected new molecular entities, and the nowcast recursion. It does not diagnose, dose, or file an IND.

The 4 September 2026 revision removes register language that had been written as if it were a measured 30-day skill, a national dollar impact, or a binding-affinity headline. Those sentences are not results of this paper.

Keywords. SEIRD, next-generation matrix, pipeline survival, Cox model, capacity nowcast

1. Introduction

A finance house holds this stack so an equity residual does not invent a trial curve and so a cash nowcast does not replace a count with a polarity. Incidence moves in days. A pipeline moves in years. A hospital budget moves in payroll cycles. Joining them is allowed only when a dated edge exists: a public trial readout that changes a regional load, a variant that changes length of stay. The failure mode is a single dashboard number.

Inputs are public or licensed and dated: WHO situational counts, CDC influenza-like series, ClinicalTrials registrations, CMS cost reports, company trial-stop filings¹²³. Outputs that are refused: badge files, unnamed hospital feeds, a generated molecule presented as intelligence, any sentence that reads as a prescription. County is the finest geographic grain in any public figure.

2. The SEIRD skeleton

Let $S,E,I,R,D$ be susceptible, exposed, infectious, recovered, and dead, normalized by a closed population when that approximation is tolerable. Mass-action incidence with transmission $\beta$ , latent rate $\sigma$ , recovery $\gamma$ , and fatality $\mu$ gives

\dot S=-\beta SI,\ \dot E=\beta SI-\sigma E,\ \dot I=\sigma E-(\gamma+\mu)I,\ \dot R=\gamma I,\ \dot D=\mu I

Recovered and dead are different sinks. $\beta$ is fit to a dated incidence series and frozen for the forecast window. If the series is revised, the fit is revised. If testing policy changes, $\beta$ is marked broken and the nowcast uses a shorter memory.

Figure 1. SEIRD flow. Recovered and dead are separate exits.

When a ward can move the path, we put demographic noise on $I$ :

dI_t=\big(\sigma E_t-(\gamma+\mu)I_t\big)\,dt+\eta\sqrt{I_t}\,dW_t

$\eta$ is estimated from the residual of the deterministic fit. If it wants to be huge, the horizon is shortened. We do not decorate the SDE to keep a long path.

3. Next-generation matrix and R₀

Linearize the infected subsystem $(E,I)$ at the disease-free point $(S,E,I,R,D)=(1,0,0,0,0)$ . Write the linearized infected dynamics as $\dot x=(T+\Sigma)x$ where $T$ is transmission and $\Sigma$ is transition.

T=\begin{pmatrix}0&\beta\\0&0\end{pmatrix},\qquad \Sigma=\begin{pmatrix}-\sigma&0\\sigma&-(\gamma+\mu)\end{pmatrix}

Theorem 1.

The basic reproduction number is the spectral radius of the next-generation matrix $-T\Sigma^{-1}$ , and equals $R_0=\beta/(\gamma+\mu)$ .

Proof.

$\Sigma^{-1}$ exists because $\sigma>0$ and $\gamma+\mu>0$ . Direct inversion yields

-\Sigma^{-1}=\begin{pmatrix}\sigma^{-1}&0\\(\gamma+\mu)^{-1}&(\gamma+\mu)^{-1}\end{pmatrix}

hence $-T\Sigma^{-1}$ has a single nonzero eigenvalue $\beta/(\gamma+\mu)$ . Van den Driessche and Watmough identify that radius with $R_0$ for this compartmental class. Local stability of the disease-free point holds when $R_0<1$ .

A time-varying $\beta_t$ produces an instantaneous reproduction number $R_t=\beta_t/(\gamma+\mu)$ . We do not treat $R_t$ as a national score. It is a local diagnostic on a named series.

4. Fitting and broken β

Parameters are estimated on a sliding window, not on a global pandemic. A global fit hides the week the testing policy changed. The objective is a Poisson or negative-binomial likelihood on new counts, with $(\sigma,\gamma,\mu)$ given weakly informative priors from the literature for the pathogen class and $\beta$ free. A parameter without a series identifier is deleted in review.

Algorithm 1 — Windowed SEIRD fit

Input: dated counts C_t, window length w, priors on (σ, γ, μ).
For each t, form the window [t−w, t] and maximize the count likelihood in β (and optionally η).
If a documented testing-policy change sits in the window, mark β broken and shrink w.
Freeze the last unbroken β for the nowcast horizon. Recalibrate when new counts arrive.

5. Pipeline attrition

Let $N_0$ be a starting set of assets and $\pi_s$ the survival probability through stage $s$ . Under independent stage risks, expected late-stage count is the product

\mathbb E[N_{\mathrm{NME}}]=N_0\prod_s\pi_s

$\pi_s$ is taken from published industry bands for the modality and then stressed. We do not invent a house $\pi$ that fills a fund. A Cox sleeve on public stop and go dates is the alternative when the question is time-to-exit of a named asset:

h(t\mid x)=h_0(t)\exp(x^\top\theta)

Covariates are stage, indication class, and sponsor size. They are not a tweet. If the Cox residual is noise, the desk reports the stage table.

Figure 2. Attrition as narrowing stage bars.

Figure 3. Kaplan–Meier style survival of trial assets by stage. Steps, not a smooth marketing curve.

Default attrition bands, to be replaced by indication-specific literature.
Stage	Keep π	Stress π
Target to hit	0.40	0.25
Hit to lead	0.50	0.35
Lead to IND	0.35	0.22
Phase I	0.60	0.45
Phase II	0.30	0.18
Phase III	0.55	0.40
NME	0.85	0.70

Proposition 2.

If stage indicators are independent Bernoulli with parameters $\pi_s$ , the variance of the terminal count is $N_0\bar\pi(1-\bar\pi)$ with $\bar\pi=\prod_s\pi_s$ . Dependence across assets in the same indication increases that variance.

Proof.

A single asset is Bernoulli with success probability $\bar\pi$ . Summing $N_0$ independent copies gives the binomial variance. Positive intra-class correlation replaces the binomial factor by $N_0\bar\pi(1-\bar\pi)\{1+(N_0-1)\rho\}$ .

6. Capacity as a stock

Occupied staffed beds $H_t$ rise with admissions and fall with discharge. A linear nowcast is

H_{t+1}=H_t+a\,I_t^{\mathrm{lag}}-d\,H_t

Coefficients are hospital-class specific. A children’s hospital is not a VA ward. If class is unknown, we do not nowcast that name. An LSTM on occupancy is a competitor and must win or lose on a later winter, not on the winter it was trained on.

Figure 4. Occupancy and a short nowcast. The dashed stroke is the forecast.

Proposition 3.

If $|1-d|<1$ and the incidence drive is bounded, $H_t$ remains bounded. The unique steady state for constant drive $I$ is $H_\star=aI/d$ .

Proof.

The homogeneous recurrence contracts at rate $1-d$ . Variation of constants gives a geometrically weighted sum of past drives, which is bounded when the drive is. Setting $H_{t+1}=H_t$ produces the displayed steady state.

7. A health shock in an economy

A computable general-equilibrium sleeve is used only when the question is sectoral. Labor supply $L(I)$ falls when infectious prevalence is high. That is the only epidemiological door into the economic model. We do not put a language model inside a utility function. No dollar impact is quoted without a named region, window, and counterfactual. The older card’s billions are not a result of this paper.

8. Algorithms and horizons

Algorithm 2 — Daily board

Pull dated public series only.
Mark β broken if testing policy moved.
Nowcast I and H at 7 and 14 days; H at 30 days as a nowcast, not a season.
Read overnight trial stops and completions from public registries.
Hand one residualized column to the equity stack only if the edge is dated.

What this paper will quote.
Horizon	Object	Quote
7 days	I, H	Yes
14 days	I, H	Yes, with a band
30 days	H	Nowcast only
90 days	Pipeline events	Event list, not a path
5 years	CGE	Scenario, not a forecast

9. What is refused

Molecular generation, a single binding-accuracy headline, person-level rows, maps that point at a building, and any clinical sentence. Variational autoencoders for SMILES strings stay in a chemistry notebook until a wet lab says otherwise. Reinforcement learning for trial sites is a research sleeve; live allocation is a constrained integer program.

Equity research may receive one residualized column after a public, dated trial event. A forecast stack may receive a dated $I_t$ as a regressor, not a refit SEIRD. Drift in a pathogen residual is an alarm, not a new compartment.

10. Limitations and conclusion

Testing changes, holidays, a silent hospital, a trial that is listed active and is dead. The note will look precise on a quiet flu week and late on the first week of a new variant. Lateness is preferred to a fake percentage. Three clocks, dated doors, no prescription.

11. Next-generation algebra

For a general infected block the next-generation operator is $-T\Sigma^{-1}$ . We record the two-by-two inverse used in Theorem 1 so a later desk can change the latent graph without guessing. If a hospitalized compartment is added, $T$ and $\Sigma$ grow and $R_0$ is recomputed; we do not keep the two-compartment formula by habit.

12. Negative-binomial overdispersion

If counts are over-dispersed relative to Poisson, the likelihood uses a negative-binomial with mean equal to the SEIRD incidence and dispersion $\phi$ . The score for $\beta$ is a weighted residual. $\phi$ is estimated, not set to one to make a path look tight.

\ell(\beta,\phi)=\sum_t\log\mathrm{NB}\big(C_t;\,\mu_t(\beta),\phi\big)

13. Sources

World Health Organization. CDC weekly influenza surveillance. ClinicalTrials.gov. CMS. van den Driessche and Watmough, Reproduction numbers and sub-threshold endemic equilibria (2002). Diekmann, Heesterbeek and Metz, On the definition and computation of R0 (1990). Cox, Regression models and life-tables (1972).

14. Final size of an epidemic

In an SIR reduction (no latent class, no death), the final attack size $z=1-S_\infty$ solves $z=1-e^{-R_0 z}$ . The same transcendental equation appears as a first approximation when the latent period is short relative to infectiousness. We do not use it as a national forecast. We use it to check that a fitted $R_0$ is not claiming an attack size the series cannot support.

Proposition 4.

For $R_0>1$ the equation $z=1-e^{-R_0 z}$ has a unique root in $(0,1)$ . For $R_0\le 1$ the only root in $[0,1]$ is zero.

Proof.

Consider $f(z)=1-e^{-R_0 z}-z$ . Then $f(0)=0$ , $f'(0)=R_0-1$ , and $f(1)<0$ . If $R_0>1$ the function becomes positive then negative, hence another root. Strict concavity gives uniqueness. If $R_0\le 1$ , $f$ is non-positive on the unit interval.

15. Discrete generation interval

When counts are weekly, a renewal representation $C_t=R_t\sum_{s}\omega_s C_{t-s}$ with a generation-interval mass $\omega$ is sometimes more stable than integrating the ODE at a daily step and aggregating. We allow it as a competitor to Algorithm 1. The generation interval is taken from the pathogen class, not from a price series.

16. Identifiability

$\beta$ and reporting probability are not separately identified from counts alone. A change in testing that is not marked as a broken $\beta$ will be eaten by the transmission parameter. That is why the broken-β rule exists. Adding a seroprevalence point, when one is public and dated, anchors reporting; we do not invent one.

Proposition 5.

If observed counts equal $\rho\beta SI$ with unknown reporting $\rho\in(0,1]$ and unknown $\beta$ , the product $\rho\beta$ is identified from the early exponential growth rate together with an assumed generation interval, but not the pair.

Proof.

Early growth of a SEIRD linearization depends on the spectrum of $T+\Sigma$ , which scales with $\beta$ not with $\rho$ . The observed level scales with $\rho$ . Without an external level anchor the two cannot be separated. The product that maps to observed incidence is identified; the split is not.

17. Site allocation as an integer program

Live trial-site allocation, when we do it at all, is

\max_{a\in\{0,1\}^K} \sum_k u_k a_k\quad\mathrm{s.t.}\quad \sum_k c_k a_k\le B,\ \sum_k a_k\ge m

with site utilities $u_k$ from inclusion criteria and historical enrollment, costs $c_k$ , budget $B$ , and a minimum site count $m$ . An agent that “optimizes a trial” in a slide is not this program.

Algorithm 3 — Greedy knapsack for sites

Sort sites by u_k / c_k descending.
Add a site if it fits the remaining budget and the remaining minimum count is still feasible.
Stop when the budget or the list is exhausted. Record the bound versus the LP relaxation.

18. Length of stay

The discharge rate $d$ in the bed stock is the reciprocal of mean length of stay for that class, when a public LOS series exists. If LOS lengthens in a variant wave and we keep last year’s $d$ , the nowcast will be optimistic. A trip from the pattern paper on the occupancy residual is a reason to refit $d$ , not to invent a new disease.

19. Next-generation matrix, with the inverse written out

Diekmann, Heesterbeek and Metz defined $R_0$ as the spectral radius of a next-generation operator. van den Driessche and Watmough gave the compartmental Jacobian form used here⁴⁵. Infected compartments are $(E,I)$ . New infections and transitions at the disease-free equilibrium $(S,E,I,R,D)=(1,0,0,0,0)$ are

T=\begin{pmatrix}0&\beta\\0&0\end{pmatrix},\qquad \Sigma=\begin{pmatrix}-\sigma&0\\sigma&-(\gamma+\mu)\end{pmatrix}.

Only the $(1,2)$ entry of $T$ is a new infection: one infectious person produces $\beta$ latent infections per unit time in a fully susceptible population. The matrix $\Sigma$ holds progressions and removals, not new infections. That split is the whole method. Putting recovery into $T$ produces a meaningless radius.

The inverse of $\Sigma$ is obtained by solving $\Sigma v=e_k$ . From the first row, $-\sigma v_1=1$ when $k=1$ , so $v_1=-1/\sigma$ is wrong-signed until we remember we need $-\Sigma^{-1}$ for mean sojourn times. Direct inversion:

\Sigma^{-1}=\frac{1}{\sigma(\gamma+\mu)}\begin{pmatrix}-(\gamma+\mu)&0\\-\sigma&-\sigma\end{pmatrix} =-\begin{pmatrix}\sigma^{-1}&0\\(\gamma+\mu)^{-1}&(\gamma+\mu)^{-1}\end{pmatrix}.

Hence $-\Sigma^{-1}$ has non-negative entries, as required of an M-matrix inverse, and

-T\Sigma^{-1}=\begin{pmatrix}\beta/(\gamma+\mu)&\beta/(\gamma+\mu)\\0&0\end{pmatrix},\qquad R_0=\rho(-T\Sigma^{-1})=\frac{\beta}{\gamma+\mu}.

Theorem 6.

For the SEIRD system of Section 2, the disease-free equilibrium is locally asymptotically stable if $R_0<1$ and unstable if $R_0>1$ .

Proof.

van den Driessche–Watmough, Theorem 2: under their (A1)–(A5), which this mass-action SEIRD satisfies near the disease-free state (non-negative new infections, a non-singular M-matrix of transitions, and no new infection in uninfected compartments), the sign of the spectral abscissa of the infected Jacobian $T+\Sigma$ equals the sign of $R_0-1$ . The uninfected block is already Hurwitz (conservation into $R$ and $D$ ). Therefore local stability of the full linearization is equivalent to $R_0<1$ .

The two-by-two formula is not a relic to keep after a hospitalized compartment is added. If $H$ is inserted between $I$ and $D$ , both $T$ and $\Sigma$ grow and $R_0$ is recomputed as a spectral radius, not as $\beta/(\gamma+\mu)$ by habit.

20. Linearization and the early growth rate

Near the disease-free state, $S\approx 1$ and the infected block is $(\dot E,\dot I)^\top=(T+\Sigma)(E,I)^\top$ . The characteristic polynomial of

T+\Sigma=\begin{pmatrix}-\sigma&\beta\\sigma&-(\gamma+\mu)\end{pmatrix}

is $\xi^2+(\sigma+\gamma+\mu)\xi+\sigma(\gamma+\mu-\beta)=0$ . The constant term is negative precisely when $R_0>1$ , so there is a positive real eigenvalue. That eigenvalue is the early exponential growth rate $r$ of incidence. It is identified from a log-linear segment of a dated count series only if reporting is stable on that segment. A testing-policy break that is not marked will be eaten by $r$ and then by $\beta$ .

Proposition 7.

If observed counts equal $\rho\cdot\sigma E$ with unknown reporting $\rho\in(0,1]$ and the generation interval is treated as known, the early growth rate identifies a function of $\beta$ but not the pair $(\rho,\beta)$ .

Proof.

The spectrum of $T+\Sigma$ depends on $\beta$ and the progression rates, not on $\rho$ . The observed level scales with $\rho$ . Without an external level anchor (a seroprevalence point, a death-based IFR argument with a stated IFR), the split is not identified. The product that maps to observed incidence is. That is why the broken- $\beta$ rule exists: a testing change is a change in $\rho$ , and treating it as a change in $\beta$ is a different model.

21. Final-size equation, with uniqueness

In an SIR reduction (no latent class, no death), $\dot S=-\beta SI$ , $\dot I=\beta SI-\gamma I$ , and $R_0=\beta/\gamma$ . Dividing the two equations on $I>0$ gives $dI/dS=-1+1/(R_0 S)$ . Integrating from $(S,I)=(1,I_0)$ with $I_0\downarrow 0$ to $I=0$ produces the attack-size equation $z=1-e^{-R_0 z}$ for $z=1-S_\infty$ .

Proposition 8.

Let $f(z)=1-e^{-R_0 z}-z$ . If $R_0\le 1$ , the only root in $[0,1]$ is $z=0$ . If $R_0>1$ , there is a unique root in $(0,1)$ .

Proof.

$f(0)=0$ , $f(1)=1-e^{-R_0}-1<0$ , and $f'(z)=R_0 e^{-R_0 z}-1$ , so $f'(0)=R_0-1$ . Also $f''(z)=-R_0^2 e^{-R_0 z}<0$ , so $f$ is strictly concave. If $R_0\le 1$ , $f'$ is non-positive at zero and decreasing, hence $f\le 0$ on the unit interval with equality only at zero. If $R_0>1$ , $f$ rises above zero and must cross zero again before $z=1$ ; strict concavity forbids a third root.

We use the equation as a consistency check on a fitted $R_0$ , not as a national forecast. A fitted $R_0$ that implies an attack size the series cannot support is a broken fit, not a scenario.

22. Negative-binomial score

If $C_t\mid\mu_t\sim\mathrm{NB}(\mu_t,\phi)$ in the mean-dispersion parameterization $\mathbb E[C_t]=\mu_t$ , $\mathrm{Var}(C_t)=\mu_t+\mu_t^2/\phi$ , the log-likelihood is

\ell(\beta,\phi)=\sum_t\log\mathrm{NB}\big(C_t;\,\mu_t(\beta),\phi\big).

The mean $\mu_t(\beta)$ is the SEIRD incidence map (or a discrete renewal mean) on the window. The score for $\beta$ is a weighted residual $\sum_t w_t(\phi)(C_t-\mu_t)\partial_\beta\mu_t$ with weights that shrink when $\mu_t$ is large, which is the point of overdispersion. Setting $\phi=\infty$ recovers the Poisson score and a path that looks tighter than the counts. We estimate $\phi$ . We do not set it to one to make a figure pretty.

A common closed form is $\log\mathrm{NB}=\log\Gamma(C+\phi)-\log\Gamma(\phi)-\log C!+\phi\log\phi-(C+\phi)\log(\phi+\mu)+C\log\mu$ . We use a library implementation of that expression, not a handwritten $\log$ that overflows at large $C$ .

23. Cox partial likelihood

When the question is time-to-exit of a named trial asset, a Cox model on public stop and go dates is the alternative to the stage-product⁶. The hazard is $h(t\mid x)=h_0(t)\exp(x^\top\theta)$ . The partial likelihood on observed exit times $t_{(1)}<\cdots<t_{(m)}$ with risk sets $R_j$ is

L(\theta)=\prod_{j=1}^m\frac{\exp(x_{(j)}^\top\theta)}{\sum_{\ell\in R_j}\exp(x_\ell^\top\theta)}.

The baseline $h_0$ cancels. Covariates are stage, indication class, and sponsor size, all public. A tweet is not a covariate. If the Cox residual is noise, the desk reports the stage table and stops. Ties are handled by Efron’s approximation; we do not invent a continuous clock for a date that is only a day.

24. Binomial pipeline variance and intra-class correlation

Under independent Bernoulli stage risks the terminal indicator for one asset is Bernoulli with success probability $\bar\pi=\prod_s\pi_s$ . For $N_0$ independent assets the terminal count is binomial and $\mathrm{Var}=N_0\bar\pi(1-\bar\pi)$ . If assets in the same indication share a common shock $U$ with $\mathrm{Corr}(1_i,1_j)=\rho$ for $i\neq j$ , the variance becomes

\mathrm{Var}(N_{\mathrm{NME}})=N_0\bar\pi(1-\bar\pi)\big(1+(N_0-1)\rho\big).

A pipeline board that quotes only the product mean is quoting the least interesting moment. Intra-class correlation in an indication is why a single failed Phase III can move the whole bar. We do not estimate a house $\rho$ from three names. We stress $\pi_s$ and we report the product, not a fund.

Product of default keep-rates, for arithmetic only.
Band	∏ π	E[NME] if N0 = 100
Default	0.40×0.50×0.35×0.60×0.30×0.55×0.85 ≈ 0.0059	0.59
Stress	0.25×0.35×0.22×0.45×0.18×0.40×0.70 ≈ 0.00049	0.05

Those products are the reason a starting set of one hundred discovery assets is not one hundred late-stage assets. They are literature-band arithmetic. They are not a portfolio forecast.

25. Capacity recursion, variation of constants

The stock $H_{t+1}=(1-d)H_t+a I_t^{\mathrm{lag}}$ is a linear non-homogeneous recurrence. The homogeneous solution contracts at rate $|1-d|$ when $0<d<2$ . Variation of constants gives

H_t=(1-d)^t H_0+a\sum_{k=0}^{t-1}(1-d)^{t-1-k}I_k^{\mathrm{lag}}.

If the drive is bounded, so is $H_t$ . If the drive is constantly $I$ , the unique steady state is $H_\star=aI/d$ . The discharge rate $d$ is the reciprocal of mean length of stay for that hospital class when a public LOS series exists. Using last year’s $d$ in a variant wave that lengthens stay makes the nowcast optimistic. That is a broken coefficient, not a new disease.

Proposition 9.

If $|1-d|<1$ and $\sup_t|I_t^{\mathrm{lag}}|\le M$ , then $\limsup_t|H_t|\le aM/d$ when $d>0$ .

Proof.

Take absolute values in the variation-of-constants formula and pass to the limit. The geometric series sums to $1/d$ when $0<d<2$ and the initial term vanishes.

26. Discrete renewal competitor

When counts are weekly, a renewal representation $C_t=R_t\sum_s\omega_s C_{t-s}$ with a generation-interval mass $\omega$ taken from the pathogen class is sometimes more stable than integrating the ODE at a daily step and aggregating. The renewal is a competitor to Algorithm 1, not a replacement. The generation interval is not estimated from a price series. If the two competitors disagree on the sign of $R_t-1$ in a quiet week, we report both and we do not average them into a slogan.

Algorithm 4 — Broken-β gate

Fit the SEIRD (or renewal) mean on the current window.
Mark the window broken if a dated testing-policy change, a holiday, or a case-definition change sits inside it.
If broken, freeze β from the last clean window and refit only reporting, or refuse the nowcast.
Never absorb a testing break into β and then quote R0.

27. What a health-economy door is allowed to be

A computable general-equilibrium sleeve, when used at all, has one epidemiological door: labor supply $L(I)$ that falls when infectious prevalence is high. There is no language-model polarity inside a utility function. There is no dollar impact without a named region, a named window, and a named counterfactual. The older register’s billions are not a result of this paper and they are not restated here as a scenario.

28. Sources, continued

van den Driessche and Watmough (2002), Mathematical Biosciences. Diekmann, Heesterbeek and Metz (1990), Journal of Mathematical Biology. Cox (1972), Journal of the Royal Statistical Society B. World Health Organization. CDC FluView. ClinicalTrials.gov. Centers for Medicare & Medicaid Services.

29. Worked next-generation numbers

Take $\beta=0.40$ , $\sigma=0.20$ , $\gamma=0.18$ , $\mu=0.02$ in reciprocal-days. Then $R_0=\beta/(\gamma+\mu)=0.40/0.20=2$ . The infected Jacobian is

T+\Sigma=\begin{pmatrix}-0.20&0.40\\0.20&-0.20\end{pmatrix}.

The characteristic polynomial is $\xi^2+0.40\xi-0.04=0$ . The positive root is $\xi=(-0.40+\sqrt{0.32})/2\approx 0.083$ per day, a doubling time of about 8.4 days. That is the early-growth identification in Section 20, on numbers. If reporting halves on day 10 and we do not mark $\beta$ broken, a refit on days 1–20 will report $R_0$ closer to 1.4 and a doubling time that is a testing artefact. The broken- $\beta$ gate exists for that reason.

Spectrum of T+Σ at the stated parameters.
Object	Value
R0 = β/(γ+μ)	2.00
Positive eigenvalue of T+Σ	0.083 / day
Doubling time ln(2)/ξ	8.4 days
Mean latent sojourn 1/σ	5 days
Mean infectious sojourn 1/(γ+μ)	5 days

30. Final size at R0 = 2

The attack-size equation $z=1-e^{-2z}$ is solved by fixed-point iteration from $z_0=0.8$ : $0.798,\ 0.797,\ 0.797$ . The unique root in $(0,1)$ is approximately 0.797. An SIR reduction with $R_0=2$ therefore claims that about four fifths of a closed population are eventually infected. A fitted $R_0=2$ on a city series that never exceeds a 15 percent attack is a broken fit, not a scenario. That is the only use of the transcendental equation.

Proof.

Let $f(z)=1-e^{-2z}-z$ . Then $f(0.79)\approx 0.0016>0$ and $f(0.80)\approx -0.0032<0$ , so a root sits between. Strict concavity (Section 21) forbids a second root in the interval.

31. Negative-binomial weights on a short count series

Suppose weekly counts $(12, 18, 31, 44, 39, 28)$ and a SEIRD mean $\mu_t$ of $(14, 20, 29, 40, 42, 30)$ . A Poisson score treats the week-5 residual $39-42=-3$ as comparable to the week-1 residual $12-14=-2$ . The negative-binomial weight $w_t=\phi/(\phi+\mu_t)$ at $\phi=8$ is 0.36 in week 1 and 0.16 in week 5. The later week, where the mean is larger, is down-weighted. That is the overdispersion. Setting $\phi=\infty$ restores equal weights and a path that looks tighter than the counts.

w_t=\frac{\phi}{\phi+\mu_t},\qquad \partial_\beta\ell=\sum_t w_t(C_t-\mu_t)\,\partial_\beta\mu_t.

The likelihood in Section 12 is this score, integrated. The figure that previously rendered as the letters “ell” and “mathrmNB” is this expression, in display form:

\ell(\beta,\phi)=\sum_t\log\mathrm{NB}\big(C_t;\,\mu_t(\beta),\phi\big).

32. Pipeline product, written out

Default keep-rates $0.40,\ 0.50,\ 0.35,\ 0.60,\ 0.30,\ 0.55,\ 0.85$ multiply to

0.40\times 0.50=0.20,\ 0.20\times 0.35=0.070,\ 0.070\times 0.60=0.042,\ 0.042\times 0.30=0.0126,\ 0.0126\times 0.55=0.00693,\ 0.00693\times 0.85=0.00589.

One hundred discovery assets have expected terminal count 0.59 under independence. The stress product in Section 24 is 0.00049, expected count 0.05. Those two numbers are why a starting set is not a late-stage set. They are not a fund.

Proposition 10.

If any one stage keep-rate is replaced by half its default, the product halves. The map $(\pi_s)\mapsto\prod\pi_s$ is multilinear, so there is no hidden compensation between stages.

Proof.

Immediate from the definition of a product. The sentence is recorded because a board that “makes up” a later stage after an early-stage miss is inventing a dependence the independent-Bernoulli model does not have. Dependence is the intra-class correlation of Section 24, and it increases variance, it does not restore a killed stage.

33. Capacity path on a constant drive

Take $d=0.20$ (mean stay five days), $a=0.15$ , $H_0=100$ , and a constant lagged incidence $I=80$ . The recursion $H_{t+1}=0.80 H_t+12$ has closed form

H_t=0.80^t\cdot 100+12\sum_{k=0}^{t-1}0.80^k=0.80^t\cdot 100+60(1-0.80^t).

The path is 100, 92, 85.6, 80.5, … and tends to $H_\star=60$ . If stay lengthens to ten days ( $d=0.10$ ) and we keep last year’s $d=0.20$ , the nowcast still aims at 60 while the true steady state is 120. That is an optimistic nowcast from a stale discharge rate, not a new pathogen.

34. What these numbers refuse to be

They are not a city’s forecast. They are not a binding-affinity headline. They are not a national dollar impact. They are the arithmetic of the objects the paper is willing to name: a next-generation radius, a final-size check, a negative-binomial weight, a stage product, and a bed stock. If a later desk wants a city, it will have to name the series, the window, and the broken- $\beta$ marks.

35. What a compartment is allowed to mean

A compartment is a count of people in a state that the model is willing to name: susceptible, exposed, infectious, recovered, dead. It is not a hospital building. It is not a person. County is the finest geographic grain in any public figure. A map that points at a building is a different product and is refused.

Recovered and dead are different sinks. Mixing them into a single “removed” class is a modeling choice that this paper does not make when a death series exists. When a death series does not exist, the fatality rate is not invented to fill the sink. The sink is closed and the model is an SEIR, not a SEIRD that pretends to know μ.

Mass-action incidence assumes homogeneous mixing in the closed population that the normalization uses. A city that is not mixing that way will produce a β that is a mixture. The mixture is not a new pathogen. It is a broken mixing assumption. Age structure, if used at all, is a later note with a named contact matrix. This paper does not invent a contact matrix from a mobility tweet.

36. Testing policy, reporting, and the broken-β rule in practice

β is fit to a dated incidence series and frozen for the forecast window. If the series is revised, the fit is revised. If testing policy moves, β is marked broken. A testing change is a change in reporting. Treating it as a change in transmission is how a nowcast becomes a story about a variant that did not happen.

Holidays, case-definition changes, and silent hospitals are the same class of break. The gate is: mark the window, freeze β from the last clean window, refit only reporting, or refuse the nowcast. Absorbing the break into β and then quoting R0 is refused.

Identifiability is the reason the gate exists. Early growth identifies a function of β given a generation interval. The observed level scales with reporting. Without a seroprevalence point or another dated level anchor, the pair (reporting, β) is not identified. We do not invent a seroprevalence point.

37. Incidence clocks, pipeline clocks, and payroll clocks

Incidence moves in days. A pipeline moves in years. A hospital budget moves in payroll cycles. Joining them is allowed only when a dated edge exists: a public trial readout that changes a regional load, a variant that changes length of stay. The failure mode is a single dashboard number that pretends the three clocks are one.

A cash nowcast that replaces a count with a polarity is the other failure mode. This stack exists so an equity residual does not invent a trial curve and so a forecast stack does not refit a compartment model inside an ensemble. Equity research may receive one residualized column after a public, dated trial event. A forecast stack may receive a dated incidence path as a regressor, not a refit SEIRD.

Drift in a pathogen residual is an alarm, not a new compartment. The alarm is a boolean. The receiving desk already has a written action: mark β broken.

38. Public inputs and the refusals that keep them public

Inputs are public or licensed and dated: WHO situational counts, CDC influenza-like series, ClinicalTrials registrations, CMS cost reports, company trial-stop filings. Outputs that are refused: badge files, unnamed hospital feeds, a generated molecule presented as intelligence, any sentence that reads as a prescription.

Molecular generation, a single binding-accuracy headline, person-level rows, and maps that point at a building stay out. Variational autoencoders for SMILES strings stay in a chemistry notebook until a wet lab says otherwise. Reinforcement learning for trial sites is a research sleeve. Live allocation, when it happens at all, is a constrained integer program with site utilities from inclusion criteria and historical enrollment, not an agent that “optimizes a trial” on a slide.

County is the finest geographic grain. A finer grain is a different legal object and a different paper.

39. Reading R0 without turning it into a national forecast

R0 is the spectral radius of the next-generation matrix at the disease-free equilibrium. For the two-compartment infected block used here it equals β over γ plus μ. That formula is not a relic to keep after a hospitalized compartment is added. If H is inserted, T and Σ grow and R0 is recomputed as a spectral radius.

The disease-free equilibrium is locally stable if R0 is less than one and unstable if R0 is greater than one. That is van den Driessche and Watmough, not a slogan. The early exponential growth rate is the positive eigenvalue of T plus Σ when R0 exceeds one. It is identified from a log-linear segment of a dated count series only if reporting is stable on that segment.

The final-size equation in an SIR reduction is a consistency check on a fitted R0, not a national forecast. A fitted R0 that implies an attack size the series cannot support is a broken fit. At R0 equals 2 the unique attack size in the unit interval is about 0.80. A city series that never exceeds 15 percent attack and a fitted R0 of 2 cannot both be true under the closed SIR reduction.

40. Overdispersion, in words a desk can use

If counts are over-dispersed relative to Poisson, the likelihood uses a negative binomial with mean equal to the SEIRD incidence and a dispersion φ that is estimated. The score for β is a weighted residual. The weights shrink when the mean is large. That is the point. Setting φ to infinity recovers the Poisson score and a path that looks tighter than the counts.

We use a library implementation of the log-gamma form, not a handwritten log that overflows at large counts. φ is not set to one to make a figure pretty.

A short weekly series can make the Poisson and negative-binomial nowcasts disagree on the sign of a one-week change. When they disagree, both are reported. They are not averaged into a slogan.

41. Pipeline attrition as a product, not as a fund

Expected late-stage count is the product of stage survivals times the starting set, under independent stage risks. The default keep-rates in the table multiply to about 0.0059. One hundred discovery assets have expected terminal count 0.59. The stress product is about 0.00049. Those two numbers are why a starting set is not a late-stage set.

Intra-class correlation in an indication increases the variance of the terminal count. It does not restore a killed stage. A board that “makes up” a later stage after an early-stage miss is inventing a dependence the independent-Bernoulli model does not have.

π_s is taken from published industry bands for the modality and then stressed. We do not invent a house π that fills a fund. If the Cox residual on public stop and go dates is noise, the desk reports the stage table and stops.

42. Cox on public dates, and what a tweet is not

When the question is time-to-exit of a named trial asset, a Cox model on public stop and go dates is the alternative to the stage product. Covariates are stage, indication class, and sponsor size, all public. A tweet is not a covariate. Ties are handled by Efron’s approximation. We do not invent a continuous clock for a date that is only a day.

The partial likelihood cancels the baseline. That is the point of Cox. A notebook that estimates a parametric baseline to make a survival curve look smooth is a different model and must say so.

If the Cox residual is noise, the stage table is the report. A smooth Kaplan–Meier that has been drawn through three names is a marketing curve. The figure in this paper is steps.

43. Capacity as a stock of staffed beds

Occupied staffed beds rise with admissions and fall with discharge. The linear nowcast is a contraction plus a lagged incidence drive. Coefficients are hospital-class specific. A children’s hospital is not a VA ward. If class is unknown, we do not nowcast that name.

The discharge rate is the reciprocal of mean length of stay for that class when a public LOS series exists. If stay lengthens in a variant wave and we keep last year’s discharge rate, the nowcast is optimistic. That is a broken coefficient, not a new disease.

Seven and fourteen days are horizons we will quote for incidence and occupancy. Thirty days is a nowcast for occupancy, not a season. Ninety days is an event list for the pipeline, not a path. Five years is a CGE scenario, not a forecast. An LSTM on occupancy is a competitor and must win or lose on a later winter, not on the winter it was trained on.

44. A health shock in an economy, with one door

A computable general-equilibrium sleeve, when used at all, has one epidemiological door: labor supply that falls when infectious prevalence is high. There is no language-model polarity inside a utility function. There is no dollar impact without a named region, a named window, and a named counterfactual.

The older register’s billions are not a result of this paper. They are not restated here as a scenario. A sectoral question can use the sleeve. A headline cannot.

The door is dated. A labor-supply path that moves before the incidence path is a different model, and it is probably a polarity.

45. What the daily board does, and what it does not say

The daily board pulls dated public series only. It marks β broken if testing policy moved. It nowcasts incidence and occupancy at seven and fourteen days, and occupancy at thirty days as a nowcast. It reads overnight trial stops and completions from public registries. It hands one residualized column to the equity stack only if the edge is dated.

It does not diagnose. It does not dose. It does not file an IND. It does not quote a binding-affinity headline. It does not quote a national dollar impact. Lateness is preferred to a fake percentage.

The note will look precise on a quiet flu week and late on the first week of a new variant. That is expected. The first week of a new variant is why the broken-β rule exists.

46. Generation intervals and the renewal competitor

When counts are weekly, a renewal representation with a generation-interval mass taken from the pathogen class is sometimes more stable than integrating the ODE at a daily step and aggregating. The renewal is a competitor, not a replacement. The generation interval is not estimated from a price series.

If the two competitors disagree on the sign of Rt minus one in a quiet week, we report both. We do not average them into a slogan. Averaging a compartment and a renewal is how a dashboard number is born.

The generation-interval mass is taken from the literature of the pathogen class and frozen for the window. Refitting the mass on the same counts that produce Rt is leaking. It is refused.

47. Length of stay, class, and the optimistic nowcast

Length of stay is a class object. A children’s hospital, a VA ward, and a large urban teaching hospital do not share a discharge rate. A public LOS series, when it exists, sets d as the reciprocal of the mean stay for that class. When it does not exist, we do not nowcast that class.

A variant wave that lengthens stay and a discharge rate from last year produce an optimistic occupancy path. The arithmetic is in Section 33. The true steady state doubles when d halves and the drive is unchanged. The nowcast still aims at last year’s stock.

A trip on the occupancy residual is a reason to refit d, not a reason to invent a new disease. The trip is a boolean from the pattern note. This paper does not invent the trip.

48. Site allocation as an integer program, if it is done at all

Live trial-site allocation, when we do it at all, is a binary knapsack with a minimum site count: utilities from inclusion criteria and historical enrollment, costs, a budget, and a floor on the number of sites. An agent that “optimizes a trial” in a slide is not this program.

The greedy by utility over cost is a bound, not the integer optimum. The LP relaxation is recorded so a later desk can see the gap. We do not hide the gap in an attention plot.

Reinforcement learning for sites stays a research sleeve. It cannot move a site live. The integer program can, when a human still owns the budget.

49. Stochastic incidence when counts are small

When counts are small, demographic noise is the Itô term on the infectious class, with a diffusion that scales as the square root of I. That is a Feller-type noise, not a language-model residual. It is used when the count is small enough that a deterministic ODE is a lie. It is not used to fatten a band on a large national series.

The noise is not a reason to quote a tighter path. It is a reason to quote a wider one. A notebook that adds the noise and then reports a tighter band has used the noise as decoration.

The deterministic SEIRD remains the skeleton for R0 and for the next-generation inverse. The noise is a nowcast sleeve, not a new radius.

50. What a fitted β is not

A fitted β is a transmission parameter on a dated window under a named mixing assumption and a named reporting regime. It is not a variant name. It is not a policy evaluation. It is not a reason to close a ward. It is not a reason to size an equity clip.

A later desk that wants a policy contrast will need an identification strategy from the forecast paper: a back-door, a front-door, or an instrument. This paper does not supply one. Incidence as a dated regressor is the only object that paper is allowed to take from here.

A later desk that wants an equity column will need a public stop or go date. A fitted β is not that date.

51. Quiet flu weeks and the first week of a variant

The note will look precise on a quiet flu week. Precision on a quiet week is not skill. It is a quiet week. The first week of a new variant is late, because the broken-β gate prefers lateness to a fake percentage. A competitor that is early and wrong is a competitor. We do not average with it.

A holiday inside the window is a break. A case-definition change is a break. A silent hospital is a break. The gate does not negotiate. If the window is broken and reporting cannot be refit from a clean β, the nowcast is refused.

Refusal is a result. A dashboard that cannot be empty is a different product.

52. The product formula, multilinearity, and stress

The map from stage keep-rates to the product is multilinear. Halve one stage, halve the product. There is no hidden compensation between stages in the independent-Bernoulli model. Stress is a lower keep-rate vector, not a hope that a later stage will make up an early miss.

The default and stress products in Section 32 are literature-band arithmetic. They are not a portfolio forecast. A fund that treats 0.59 expected NMEs from a hundred discovery assets as a pipeline of a hundred late-stage assets is reading a different table.

Dependence across assets in the same indication is the intra-class correlation. It fattened variance. It does not fatten the mean.

53. Horizons, restated as a speech rule

Seven days: incidence and occupancy, yes. Fourteen days: the same, with a band. Thirty days: occupancy as a nowcast, not a season. Ninety days: pipeline events as a list, not a path. Five years: a CGE scenario, not a forecast.

A single percent that claims to cover those five rows is the register language the September revision deleted. It is not a result of this paper. It is not restated here as a skill.

The speech rule is the table. If a number is not in the table, it is not spoken.

55. The two infected compartments, in words a later desk can defend

Exposed and infectious are the infected block. New infections enter exposed, not infectious, under the latent rate σ. Putting new infections into I is a different graph and a different next-generation matrix. Recovery and death leave I. They do not enter T. Putting recovery into T produces a radius that is not R0.

The inverse of Σ is written out in Section 19 so a later desk can change the latent graph without guessing. Mean sojourn in E is 1/σ. Mean sojourn in I is 1/(γ+μ). The next-generation entries are β times those sojourns. For the two-by-two block used here, R0 equals β over γ plus μ. That is a spectral radius, not a slogan.

If a hospitalized compartment is added between I and D, both T and Σ grow. R0 is recomputed. The two-compartment formula is not kept by habit. A notebook that keeps β/(γ+μ) after adding H has not computed R0.

56. Reporting, seroprevalence, and the pair that is not identified

Observed counts equal a reporting probability times an incidence map. Early growth of the linearization depends on β, not on reporting. The observed level scales with reporting. Without a dated level anchor the pair is not identified. A seroprevalence point, when one is public and dated, anchors reporting. We do not invent one.

A testing-policy change that is not marked will be eaten by β. That is the broken-β rule. Holidays, case-definition changes, and silent hospitals are the same class of break. The gate is: mark the window, freeze β from the last clean window, refit only reporting, or refuse the nowcast.

A later desk that wants a policy contrast will not get it from a fitted β. They will need an identification strategy from the forecast paper. This paper supplies a dated incidence path, not a do-operator.

57. Stage tables, public registries, and the date that is only a day

ClinicalTrials.gov is a public door. A stop or a completion is a date. The date is a day, not a continuous clock. Efron’s approximation for ties is how Cox treats two assets that exit on the same day. Inventing a hour is refused.

Sponsor size, indication class, and stage are public covariates. A tweet is not a covariate. A language-model polarity on a press release is not a covariate. If the Cox residual is noise, the stage table is the report.

The default keep-rates are literature bands for the modality, then stressed. They are not a house π that fills a fund. The product is multilinear. Halve one stage, halve the product. A board that makes up a later stage after an early miss is inventing a dependence the independent-Bernoulli model does not have.

58. Staffed beds, class, and the later winter

Occupied staffed beds are a stock. Admissions raise the stock. Discharge lowers it. The discharge rate is the reciprocal of mean length of stay for that hospital class when a public LOS series exists. A children’s hospital is not a VA ward. If class is unknown, we do not nowcast that name.

An LSTM on occupancy is a competitor. It must win or lose on a later winter, not on the winter it was trained on. A nowcast that is beautiful on last February and late on the first freeze is a nowcast that liked the training winter.

Thirty days is a nowcast for occupancy, not a season. A single occupancy percent that claims to be a season is the speech the horizon table forbids. Seven and fourteen days are the horizons we will quote for incidence and occupancy, with a band at fourteen.

59. What the September revision deleted, and why it stays deleted

A CGE sleeve, when used at all, has one epidemiological door: labor supply that falls when infectious prevalence is high. There is no polarity inside a utility function. There is no dollar impact without a named region, a named window, and a named counterfactual.

Molecular generation stays in a chemistry notebook. Person-level rows stay out. Maps that point at a building stay out. County is the finest grain. A finer grain is a different legal object.

60. Quiet weeks, variant weeks, and refusal as a result

The note will look precise on a quiet flu week. Precision on a quiet week is not skill. The first week of a new variant is late, because the broken-β gate prefers lateness to a fake percentage. A competitor that is early and wrong is a competitor. We do not average with it.

A dashboard that cannot be empty is a different product. Refusal of a nowcast is a result. The daily board still reads overnight trial stops. It still marks β broken. It does not invent a compartment to fill the empty cell.

Lateness is preferred to a fake percentage. That sentence is the speech rule for the first week of a variant and for a holiday inside the window and for a silent hospital.

62. Closed population

The SEIRD states are normalized by a closed population when that approximation is tolerable.

A city that is not closed will leak people. The leak is not a new pathogen.

County is the finest public grain. A building map is refused.

63. Different sinks

Recovered and dead are different sinks. Mixing them into removed is a different model.

If a death series does not exist, μ is not invented. The model becomes SEIR.

A SEIRD that pretends to know μ without deaths is a costume.

64. Mass action

Mass-action incidence assumes homogeneous mixing in the normalized population.

A mixture β from bad mixing is not a new variant.

Age structure needs a named contact matrix. A mobility tweet is not that matrix.

65. Broken beta gate

Testing-policy moves mark β broken. Holidays, case-definition changes, and silent hospitals do the same.

Freeze β from the last clean window and refit reporting, or refuse the nowcast.

Absorbing a break into β and quoting R0 is refused.

66. Identifiability pair

Early growth identifies a function of β given a generation interval. The level scales with reporting.

Without a dated seroprevalence or other level anchor the pair is not identified.

We do not invent a seroprevalence point.

67. Next-generation split

New infections go in T. Progressions and removals go in Σ.

Putting recovery in T produces a meaningless radius.

If H is added, both matrices grow and R0 is recomputed as a spectral radius.

68. Two-by-two inverse

Mean sojourn in E is 1/σ. Mean sojourn in I is 1/(γ+μ).

R0 equals β over γ plus μ for this block only.

The inverse is written so a later desk can change the latent graph without guessing.

69. Early growth

The positive eigenvalue of T plus Σ is the early growth rate when R0 exceeds one.

It is identified from a log-linear segment only if reporting is stable on that segment.

A testing break on that segment is eaten by the growth rate if it is not marked.

70. Final-size check

In an SIR reduction the attack size solves z equals 1 minus e to the minus R0 z.

At R0 equals 2 the unique root in the unit interval is about 0.80.

A fitted R0 of 2 and a 15 percent observed attack cannot both be true under that reduction.

71. Negative binomial weights

Weights φ over φ plus μ shrink when the mean is large.

Setting φ to infinity restores Poisson and a path that looks tighter than the counts.

φ is estimated. It is not set to one to make a figure pretty.

72. Library log-gamma

The log-likelihood uses a library log-gamma form.

A handwritten log that overflows at large C is refused.

When Poisson and NB disagree on a one-week sign, both are reported. They are not averaged.

73. Pipeline product

Expected terminal count is N0 times the product of stage keep-rates under independence.

Default product is about 0.0059. Stress product is about 0.00049.

Those numbers are why a hundred discovery assets are not a hundred late-stage assets.

74. Multilinearity

Halve one stage, halve the product. There is no hidden compensation in the independent model.

A board that makes up a later stage after an early miss invents a dependence the model does not have.

Intra-class correlation fattens variance, not the mean.

75. Literature bands

π_s comes from published industry bands for the modality, then stressed.

We do not invent a house π that fills a fund.

If the Cox residual is noise, the stage table is the report.

76. Cox public dates

Covariates are stage, indication, sponsor size. A tweet is not a covariate.

Ties use Efron. A date is a day, not an invented hour.

A parametric baseline drawn to look smooth is a different model and must say so.

77. Kaplan-Meier steps

The survival figure is steps. A smooth curve through three names is a marketing curve.

Public stop and go dates are the events.

An asset listed active and dead is a limitation, not a reason to invent a completion.

78. Bed stock

H rises with admissions and falls with discharge. Class-specific coefficients.

A children’s hospital is not a VA ward. Unknown class is not nowcast.

d is the reciprocal of mean LOS when a public series exists.

79. Optimistic stay

A variant that lengthens stay and last year’s d produce an optimistic path.

If d halves and the drive is unchanged, the steady state doubles.

A trip on the occupancy residual is a reason to refit d, not a new disease.

80. Horizons speech

7 days: I and H, yes. 14 days: the same with a band. 30 days: H as a nowcast, not a season.

90 days: pipeline events as a list. 5 years: CGE scenario, not a forecast.

A single percent covering those rows is the language the September revision deleted.

81. LSTM competitor

An occupancy LSTM must win or lose on a later winter.

Beauty on the training winter is not skill.

Incidence may enter as a dated drive. A refit SEIRD inside the LSTM is refused.

82. One CGE door

Labor supply that falls when infectious prevalence is high is the only epidemiological door.

No polarity inside a utility function. No dollar impact without region, window, and counterfactual.

The older billions are not restated as a scenario.

83. Daily board

Pull dated public series. Mark β broken if testing moved. Nowcast I and H at 7 and 14. H at 30 as a nowcast.

Read overnight trial stops from public registries.

Hand one residualized column to equity only if the edge is dated.

84. Refusals

No diagnose, dose, IND, binding headline, national dollar, person-level row, building map, SMILES as intelligence.

County is the finest grain.

RL for sites stays a sleeve. Live allocation is an integer program if it happens at all.

85. Site knapsack

Binary sites, utilities from inclusion and enrollment, costs, budget, minimum site count.

Greedy by utility over cost is a bound. The LP gap is recorded.

An agent that optimizes a trial on a slide is not this program.

86. Renewal competitor

Weekly counts may use a renewal with a generation-interval mass from the pathogen class.

The mass is frozen for the window. Refitting it on the same counts that produce Rt is leaking.

If compartment and renewal disagree on the sign of Rt minus one, both are reported.

87. Stochastic sleeve

Small counts may use a square-root diffusion on I. That is Feller-type noise, not a language residual.

It widens a band. A notebook that adds noise and tightens a band used decoration.

R0 still comes from the deterministic skeleton.

88. Quiet versus variant

Precision on a quiet flu week is a quiet week. The first variant week is late on purpose.

A competitor that is early and wrong is not averaged in.

Refusal of a nowcast is a result. A dashboard that cannot be empty is a different product.

89. Equity door

Equity may receive one residualized column after a public dated trial event.

A fitted β is not that event.

The forecast paper may receive dated I as a regressor, not a refit SEIRD.

90. Drift alarm

Drift in a pathogen residual is an alarm, not a new compartment.

The receiving action is already written: mark β broken.

This paper does not invent the trip.

91. WHO CDC doors

WHO situational counts, CDC influenza-like series, ClinicalTrials, CMS cost reports, company stop filings.

Outputs that require badges or unnamed hospital feeds are refused.

Licensed dated series are allowed when the license and the date exist.

92. What fitted beta is not

A fitted β is a transmission parameter on a dated window under named mixing and named reporting.

It is not a variant name, a policy evaluation, a ward closure, or an equity clip.

A policy contrast needs identification from the forecast paper.

93. Speech rule for clocks

Incidence is a days clock. Pipeline is a years clock. Payroll is a payroll clock.

Join them only on a dated edge. A single dashboard number is the failure mode.

A cash nowcast that replaces a count with a polarity is the other failure mode.

94. September deletions

Measured 30-day skill language, national dollar impact, binding-affinity headline: deleted.

They are not scenarios here.

Lateness is preferred to a fake percentage.

95. Worked R0 equals 2

β=0.40, σ=0.20, γ=0.18, μ=0.02 gives R0=2 and a doubling time of about 8.4 days.

A reporting break on day 10 that is not marked will pull a refit R0 toward 1.4.

That artefact is why the gate exists.

96. Worked product

Default keep-rates multiply to 0.00589. One hundred starts expect 0.59 terminals under independence.

Stress expects 0.05. Those are literature-band arithmetic, not a fund.

Dependence fattens variance only.

97. Worked bed path

d=0.20, a=0.15, I=80, H0=100 tends to 60. If stay doubles and d is kept, the nowcast still aims at 60 while truth aims at 120.

That is an optimistic stale discharge rate.

It is not a new pathogen.

98. NB toy weights

Weekly counts and means in Section 31 give Poisson residuals that look comparable and NB weights that are not.

Week 5 is down-weighted because the mean is larger.

That is overdispersion in a desk’s language.

99. Integer program speech

If site allocation is done, it is the knapsack, not an RL live move.

The human still owns the budget.

The gap to the LP is shown, not hidden in attention.

102. What a SEIRD count is allowed to name

A compartment is a named count of people in a state the model is willing to keep. The five names used here are susceptible, exposed, infectious, recovered, and dead. The count is normalized by a closed population when that approximation is tolerable. The closed population is a county or a named collection of counties, never a building and never a person-level row. County is the finest geographic grain in any public figure. A map that points at a ward, a campus, or a household is a different legal object and is refused.

The arrows are the only dynamics the skeleton is allowed. Susceptible people become exposed at a rate that depends on the infectious count and on a transmission parameter under a named mixing assumption. Exposed people become infectious at a latent rate. Infectious people recover or die. Recovered and dead are different sinks. Mixing them into a single removed class is a modeling choice this paper does not make when a death series exists. When a death series does not exist, the fatality rate is not invented to fill the sink. The sink is closed and the model is an SEIR, not a SEIRD that pretends to know a death rate.

The skeleton is not a hospital. Occupied staffed beds are a later stock driven by a lagged incidence map. The skeleton is not a trial. Trial assets live on a product of stage survivals or on a Cox clock of public stop and go dates. The skeleton is not a wage bill. A computable general-equilibrium sleeve, when used at all, has one epidemiological door and is not allowed to rewrite the compartment names.

A parameter without a series identifier is deleted in review. A compartment without a dated public series is not drawn. A national series may be used as a competitor or as a context line. It is not the operating grain. The operating grain is the county, or a named set of counties whose series can be dated and whose testing regime can be marked.

When the closed-population approximation is not tolerable, the paper does not invent a migration term from a mobility tweet. It either names a published inter-county flow table and treats that table as a different model, or it refuses the nowcast for that window. Open-population language that is only a slogan is refused.

The daily object is a dated incidence series and, when it exists, a dated death series. The weekly object is the same series aggregated after the daily fit, or a renewal competitor that lives on weekly counts from the start. The monthly object is occupancy and cost-report context, not a new compartment. Those three clocks are not one number.

A later desk that wants a person-level hazard, a household contact diary, or a badge file is asking for a different paper. This paper will not supply one by renaming a compartment.

103. Mixing assumptions, written as permissions and refusals

Mass-action incidence assumes that every susceptible person in the closed county meets infectious people in proportion to the infectious share, with a single transmission parameter. That is a homogeneous mixing assumption. It is not a claim that the county is homogeneous. It is a claim that the model is willing to summarize the meeting process by one number on a dated window. A city that is not mixing that way will produce a transmission parameter that is a mixture of meeting processes. The mixture is not a new pathogen. It is a broken mixing assumption.

Frequency-dependent incidence, in which the contact rate does not scale with population size, is a different assumption. It is used when the closed population is a school, a ward class, or another setting where the meeting rate is closer to a capacity than to a density. This paper does not switch from mass-action to frequency dependence because a figure looks nicer. The switch is a named assumption with a named population. If the population used for normalization is the county census and the meetings are a school, the parameter is a mixture again.

Age structure, if used at all, requires a named contact matrix with a public source and a date. The matrix is frozen for the window. It is not estimated from the same counts that produce the transmission parameter. Estimating contacts and transmission on the same short series is how a notebook manufactures an age story. This paper does not invent a contact matrix from a mobility tweet, from a language-model polarity, or from a weekend traffic count.

Household, workplace, and community layers are allowed only as a published multi-layer construction with named rates. A three-layer slogan without a table is not a layer model. Preferential mixing by age, by occupation, or by prior infection is the same rule: name the matrix, date it, freeze it, or refuse the structured model and keep the homogeneous county.

A weekend effect, a holiday, and a school-calendar break are mixing breaks. They are not automatically transmission breaks. The gate is the same gate used for testing: mark the window, decide whether the meeting process changed or the reporting process changed, and do not absorb both into one parameter. A school holiday that empties classrooms and a testing holiday that empties the case file can move the same series. Treating them as one number is how a nowcast becomes a story about a variant that did not happen.

Spatial mixing below the county is refused. Gravity models that point at a tract, a campus, or a building are a different legal object. A commuter flow between named counties may enter only if the flow table is public, dated, and not inferred from a phone-panel that cannot be cited. If the table does not exist, the counties stay separate or they are pooled under a stated administrative rule, not under a learned embedding.

The transmission parameter is therefore a parameter of a named mixing assumption on a named window. It is not a variant name. It is not a policy evaluation. It is not a reason to close a ward. When the mixing assumption is false, the honest sentence is that the assumption is false, not that the pathogen changed.

A later desk that wants a policy contrast will need an identification strategy that this paper does not supply. Mixing is an assumption for the incidence path. It is not a treatment effect.

104. The two-compartment infected block and the next-generation radius

The infected block in this skeleton is the pair of exposed and infectious counts. New infections enter the exposed class. The infectious class produces those new infections. Recovery and death leave the infectious class. That split is the whole next-generation method. Putting recovery into the transmission matrix produces a meaningless radius. Putting new infections into the transition matrix does the same.

At the disease-free equilibrium the county is fully susceptible and the infected counts are zero. Linearization of the infected block yields a transmission matrix whose only nonzero entry is the transmission parameter in the exposed-from-infectious position, and a transition matrix that holds latent progression, recovery, and death. The next-generation matrix is the product of the transmission matrix and the inverse of the negative transition matrix. The basic reproduction number is the spectral radius of that product.

For this two-by-two infected block the spectral radius collapses to a ratio: the transmission parameter divided by the sum of recovery and death rates. In symbols the paper keeps, that is the radius of the next-generation matrix and it equals the transmission parameter over recovery plus fatality. The latent rate cancels in the radius because every exposed person who does not die in latency, and this skeleton does not put a death rate on the exposed class, becomes infectious. The radius counts secondary infectious people, not secondary exposures that never progress.

The two-by-two formula is not a relic to keep after a hospitalized compartment is added. If a hospital class is inserted between the infectious class and the sinks, both matrices grow. The radius is recomputed as a spectral radius. Quoting the two-compartment ratio by habit after the block has changed is a different model pretending to be the same number.

The disease-free equilibrium is locally stable when the radius is less than one and unstable when the radius is greater than one. That is the van den Driessche and Watmough theorem for this compartmental class, under the usual conditions: non-negative new infections, a non-singular M-matrix of transitions, and no new infection in the uninfected compartments. This mass-action SEIRD satisfies those conditions near the disease-free state. The uninfected block is already Hurwitz because it only conserves into recovered and dead.

A time-varying transmission parameter produces an instantaneous reproduction number with the same ratio. That instantaneous number is a local diagnostic on a named series and a named window. It is not a national score. It is not a reason to write a headline about a country. County is still the finest grain.

If the death sink is closed because no death series exists, the radius becomes the transmission parameter over recovery alone. That is an SEIR radius, not a SEIRD radius with a hidden fatality. The paper will not keep the death rate in the denominator after the sink has been closed.

105. Identifiability of reporting and transmission as a pair

Observed counts are a reporting fraction times a flow out of the exposed class, or, in some reductions, a reporting fraction times an incidence term that already contains the transmission parameter. The pair of reporting and transmission is not identified from a count series alone. Early exponential growth identifies a function of the transmission parameter given a generation interval. The observed level scales with reporting. Without a dated level anchor the split cannot be made.

A seroprevalence point, when it is public and dated, is the honest level anchor. We do not invent one. A death-based argument that backs out an infection-fatality rate can be a second anchor only if the fatality rate is taken from a named study and not tuned to make the pair look pretty. If neither anchor exists, the paper reports a product that maps to observed incidence and refuses to split the product into a reporting story and a transmission story.

That is why the broken-transmission rule exists. A testing change is a change in reporting. Treating it as a change in transmission is a different model. The early growth rate, which is the positive eigenvalue of the infected Jacobian when the radius exceeds one, will eat a testing break if the break is not marked. A refit on a window that contains the break will then report a new radius that is a testing artefact.

Identifiability is not a reason to add more compartments. Adding a hospitalized class without a hospital series adds parameters. Adding an asymptomatic class without a testing or seroprevalence split adds parameters. The pair becomes a triple or a quadruple. The paper does not add a class to look complete. It adds a class when a dated series exists for that class or when a literature sojourn is being used as a frozen number, not as a free fit.

A weakly informative prior on latent, recovery, and fatality rates from the pathogen class is not a solution to the reporting problem. Those priors pin sojourn times. They do not pin reporting. A notebook that treats a literature recovery rate as if it identified reporting has confused a sojourn with a level.

The operating sentence is therefore simple. Fit transmission on a window where reporting is believed stable. If reporting moves, mark transmission broken. Refit reporting if a clean transmission from the last clean window can be frozen. If it cannot, refuse the nowcast. Do not publish a split of the pair that the series cannot support.

A later desk that wants a policy contrast on testing will need an instrument or a discontinuity that this paper does not invent. The identifiability note is a refusal, not a research program on testing policy.

106. Broken transmission when testing and reporting change

The transmission parameter is fit to a dated incidence series and frozen for the forecast window. If the series is revised, the fit is revised. If testing policy moves, the parameter is marked broken. A testing change is a change in reporting. Treating it as a change in transmission is how a nowcast becomes a story about a variant that did not happen.

Holidays, case-definition changes, and silent hospitals are the same class of break. A holiday that closes testing sites empties the case file without emptying the county. A case-definition change that adds or drops a class of positives moves the level overnight. A hospital that stops sending a feed looks like a collapse in occupancy or in severe incidence. The gate is the same in each case: mark the window, freeze transmission from the last clean window, refit only reporting, or refuse the nowcast.

Absorbing the break into the transmission parameter and then quoting a basic reproduction number is refused. The radius is a function of transmission. A testing artefact in transmission is a testing artefact in the radius. The daily board will prefer an empty cell to that artefact.

The windowed fit exists so a global pandemic fit cannot hide the week the testing policy changed. A global fit is almost always a mixture of reporting regimes. The mixture can look smooth. Smoothness is not identification. The algorithm is: for each date, form a window, maximize the count likelihood in transmission, and if a documented testing-policy change sits in the window, mark the parameter broken and shrink the window.

Documented means public. A WHO situational note, a CDC influenza-like update that changes a case definition, a state health notice, or a hospital-system outage that is in a public status page can mark a break. A rumor cannot. A language-model polarity cannot. A weekend tweet cannot.

When two breaks sit close together, the window may have no clean stretch long enough to identify transmission. That is a refusal, not a reason to borrow a parameter from a neighboring county. Neighboring counties can have different testing regimes. Pooling them to save a nowcast is how a break becomes a region.

A revision to a past count is not automatically a testing break. It is a data revision. The fit on the revised series is the fit. The old fit is archived. Rewriting the old fit in place so a walk-forward looks cleaner is refused. The as-of stamp on the series is part of the object.

The first week of a new variant is the week this gate is for. The note will look late. Lateness is preferred to a fake percentage. A competitor that is early and wrong is a competitor. We do not average with it.

107. Negative-binomial overdispersion and a dispersion that is estimated

If counts are over-dispersed relative to Poisson, the likelihood uses a negative binomial with mean equal to the SEIRD incidence and a dispersion parameter that is estimated. The score for the transmission parameter is a weighted residual. The weights shrink when the mean is large. That is the point of overdispersion. Setting the dispersion to infinity recovers the Poisson score and a path that looks tighter than the counts. Setting the dispersion to one to make a figure pretty is refused.

We use a library implementation of the log-gamma form, not a handwritten logarithm that overflows at large counts. The mean is the model incidence on that date, after reporting if reporting is being used as a frozen fraction. The dispersion is a free parameter on the window. It is not borrowed from a national figure. It is not set to one because a notebook defaulted to one.

A short weekly series can make the Poisson and negative-binomial nowcasts disagree on the sign of a one-week change. When they disagree, both are reported. They are not averaged into a slogan. Averaging a Poisson path and a negative-binomial path is how a dashboard number is born.

The weight at a date is the dispersion over the dispersion plus the mean. A week whose mean is large is down-weighted relative to a week whose mean is small. That is why a late winter peak does not dominate the score the way it would under Poisson. A notebook that then reports a tighter band than the Poisson band has used the negative binomial as decoration. The honest band is wider when the dispersion is finite.

A weekly series of a few dozen counts can leave the dispersion weakly identified. Weak identification is reported. It is not a reason to pin the dispersion at one. A profile likelihood that is flat in the dispersion is a reason to quote both the Poisson path and a conservative negative-binomial path, not a reason to pretend the Poisson path is the model.

The mean must stay the SEIRD incidence. Replacing the mean by a spline or by a language-model leftover and then calling the leftover a negative-binomial compartment model is a different paper. The dispersion is not a residual from another desk. It is a parameter of this likelihood.

When counts are small enough that a deterministic mean is a lie, demographic noise on the infectious class is the other sleeve. That noise is a Feller-type square-root diffusion. It is not a reason to drop the negative binomial on the observation equation. The two objects are different. One is process noise. One is observation overdispersion. A notebook that uses one as a substitute for the other is confused.

The operating report names the estimated dispersion, the Poisson competitor, and whether they agree on the sign of the next seven-day change. That is the whole speech for overdispersion. A single pretty path is refused.

108. Final size as a consistency check and not as a forecast

In an SIR reduction without a latent class and without a death sink, the final attack size solves a transcendental equation: one minus the eventual susceptible share equals one minus the exponential of minus the basic reproduction number times that attack size. The same equation is a first approximation when the latent period is short relative to infectiousness. We do not use it as a national forecast. We use it to check that a fitted radius is not claiming an attack size the series cannot support.

For a radius greater than one the equation has a unique root in the open unit interval. For a radius at most one the only root in the closed unit interval is zero. That uniqueness is why the check is usable. A fitted radius of two in the closed SIR reduction claims an attack size of about four fifths. A city or county series that never exceeds a fifteen percent attack and a fitted radius of two cannot both be true under that reduction. The fit is broken. The equation is not a scenario.

The check is local. County is the finest grain. A national attack size is not computed. A national dollar impact is not computed from an attack size. The older register’s billions are not a result of this paper and are not restated here as a scenario.

When the skeleton is SEIRD rather than SIR, the death sink and the latent class move the attack size. The SIR equation is then a bound or a first check, not an exact identity. If the fitted radius still implies an attack the dated series cannot support, the fit is still broken. The extra compartments are not a license to ignore the check.

A later winter, a vaccination campaign, or an open-population flow can make a closed SIR reduction the wrong object. Those are reasons to refuse the check as a forecast. They are not reasons to skip the check as a consistency test on a window that was fit under the closed assumption. If the window was fit as closed, the implied attack is a property of the fit.

The final-size check is also a check on reporting. A radius identified from early growth, combined with a reporting fraction that is too small, will imply a large true attack and a small observed attack. That pair can be consistent. It can also be a way to hide a broken radius behind an invented reporting fraction. Without a seroprevalence point the paper will not congratulate itself for that consistency.

The speech rule is: quote the implied attack under the reduction that was actually fit, quote the observed attack on the dated series, and say whether they can coexist. Do not quote a national attack. Do not quote a season. Do not turn the transcendental equation into a headline.

109. The generation-interval renewal as a competitor

When counts are weekly, a renewal representation with a generation-interval mass taken from the pathogen class is sometimes more stable than integrating the ordinary differential equation at a daily step and aggregating. The renewal says that today’s counts equal a reproduction number times a weighted sum of past counts, with weights equal to the generation-interval mass. We allow that representation as a competitor to the windowed SEIRD fit. It is not a replacement.

The generation interval is taken from the literature of the pathogen class and frozen for the window. It is not estimated from a price series. It is not estimated from the same counts that produce the reproduction number. Refitting the mass on the same counts that produce the reproduction number is leaking. It is refused.

If the compartment path and the renewal path disagree on the sign of the reproduction number minus one in a quiet week, we report both. We do not average them into a slogan. Averaging a compartment and a renewal is how a dashboard number is born. Disagreement on a quiet week is information. It often means the generation-interval mass and the latent-plus-infectious sojourn are not saying the same thing.

The mean generation interval should be consistent with the mean latent sojourn plus a fraction of the mean infectious sojourn. If the literature mass used for the renewal has a mean that the SEIRD sojourns cannot support, one of the two objects is using the wrong pathogen class. That is a reason to stop and name the sources, not a reason to split the difference.

Weekly counts hide intra-week reporting. A Monday bulge and a weekend hole can look like a generation interval if a daily model is aggregated carelessly. The renewal on weekly totals is less tempted by that bulge, which is why it can be more stable. It is also less able to see a mid-week testing break. The broken-transmission gate still applies. A case-definition change on a Wednesday is still a break in the weekly total.

The renewal competitor is not allowed to become a national score. County remains the grain. A national weekly influenza-like series from the Centers for Disease Control and Prevention may be used as a context line or as a competitor at the national grain. It does not replace the county nowcast and it does not license a national dollar sentence.

A later desk that wants a nowcast will be handed both paths when they disagree, and the SEIRD path when they agree and the window is clean. The desk will not be handed an average.

110. Pipeline attrition as a product of stage survivals

Expected late-stage count is the product of stage survivals times the starting set, under independent stage risks. The default keep-rates used in the table multiply to about 0.0059. One hundred discovery assets have expected terminal count of about 0.59. The stress product is about 0.00049, an expected count of about 0.05. Those two numbers are why a starting set is not a late-stage set. They are not a fund.

The default keep-rates are literature-band arithmetic for a modality, not a house invention that fills a portfolio. Target to hit, hit to lead, lead to investigational application, then the three clinical phases and a final keep into a named molecular entity: those are the stages. Each keep-rate is taken from published industry bands and then stressed. We do not invent a house keep-rate to make the product look like a pipeline of late-stage assets.

The map from keep-rates to the product is multilinear. Halve one stage, halve the product. There is no hidden compensation between stages in the independent-Bernoulli model. A board that makes up a later stage after an early-stage miss is inventing a dependence the model does not have. Stress is a lower keep-rate vector, not a hope.

Intra-class correlation in an indication increases the variance of the terminal count. It does not restore a killed stage. Dependence across assets in the same indication is the reason a cluster of failures is not a surprise. It is not a reason to raise the mean. The mean stays the product.

If the Cox residual on public stop and go dates is noise, the desk reports the stage table and stops. The product is then the whole pipeline object. A smooth curve drawn through three names is a marketing curve. The figure this paper allows is steps.

A ninety-day speech about the pipeline is an event list: public stop dates, public go dates, public completions, public withdrawals. It is not a path of expected counts. Five years is a scenario for a later economic sleeve, not a forecast of the product. The product does not become a season.

The starting set must be named and dated. A set that grows by rumor is not a set. A set that includes assets without a public identifier is not a set. Person-level investigator rows are not assets. A generated molecule presented as an asset is refused.

The operating sentence is the product, the stress product, and whether a Cox sleeve on public dates beat noise. That is the pipeline speech. A single percentage that claims to be a portfolio skill is refused.

111. Cox models on public stop and go dates

When the question is time-to-exit of a named trial asset, a Cox model on public stop and go dates is the alternative to the stage product. The hazard is a baseline hazard times an exponential of covariates. The partial likelihood cancels the baseline. That is the point of the Cox construction. A notebook that estimates a parametric baseline to make a survival curve look smooth is a different model and must say so.

Covariates are stage, indication class, and sponsor size, all public. A tweet is not a covariate. A language-model polarity is not a covariate. A rumor of a hold is not a covariate until it becomes a dated public stop or a dated public notice. Ties are handled by Efron’s approximation. We do not invent a continuous clock for a date that is only a day.

ClinicalTrials registrations and company trial-stop filings are the date sources. A completion date, a withdrawal date, a termination date, and a primary-completion change that is stamped on the registry are usable. A silent page that never updates is a censoring problem, not a go date. We do not impute a stop from a missing update.

If the Cox residual is noise, the stage table is the report. A smooth Kaplan–Meier that has been drawn through three names is a marketing curve. The figure in this paper is steps. Steps tell the reader how few names were at risk. A smooth curve hides that.

Sponsor size is a public bucket, not a prestige score. Indication class is a public label, not a learned embedding. Stage is the registry stage, not a house stage that has been moved forward to look kinder. If the registry stage and a press stage disagree, the registry stage wins or the name is dropped.

The Cox sleeve does not diagnose. It does not dose. It does not file an investigational application. It does not quote a binding-affinity headline as a covariate. Molecular measurements stay in a chemistry notebook until a wet lab and a public filing exist.

Time-to-exit is not a nowcast of incidence and not a nowcast of occupancy. A public go date may later become a dated edge for another desk, after residualization. The Cox model itself is not that edge. The edge is the date.

A later desk that wants an equity column will need that public date. A later desk that wants a capacity story will need a dated change in load, not a hazard ratio. The hazard ratio is a description of historical public exits. It is not a recommendation to enroll, to stop, or to buy.

112. Hospital beds as a stock with a lagged incidence drive

Occupied staffed beds are a stock. They rise with admissions and fall with discharge. The linear nowcast is a contraction plus a lagged incidence drive. In the discrete form used here, tomorrow’s occupied stock equals one minus the discharge rate times today’s stock, plus an admission coefficient times a lagged infectious or incidence count. That is a linear filter. It is not a new compartment inside the SEIRD skeleton, and it is not a reason to recompute the next-generation radius unless a hospital class has been inserted into the infected block.

Coefficients are hospital-class specific. A children’s hospital is not a veterans ward. A large urban teaching hospital is not a critical-access hospital. If class is unknown, we do not nowcast that name. Unknown class is a refusal, not a reason to borrow a discharge rate from a national average.

The discharge rate is the reciprocal of mean length of stay for that class when a public length-of-stay series exists. If stay lengthens in a variant wave and we keep last year’s discharge rate, the nowcast is optimistic. The arithmetic is simple. On a constant drive the stock tends to the admission drive over the discharge rate. Halve the discharge rate and the true steady state doubles while a stale nowcast still aims at last year’s stock. That is a broken coefficient, not a new disease.

The lag on incidence is a class object as well. Admissions do not arrive on the same day as a positive test. A lag that is estimated on the same short winter that is being nowcast is leaking. The lag is taken from a public length-of-stay and admission-delay literature for the class, or it is estimated on an earlier winter and frozen. It is not tuned to make this week’s occupancy look skillful.

Seven and fourteen days are horizons we will quote for incidence and occupancy. Thirty days is a nowcast for occupancy, not a season. Ninety days is an event list for the pipeline, not a path of beds. Five years is a computable general-equilibrium scenario, not a forecast of the stock. An occupancy path that claims to be a season is refused.

A trip on the occupancy residual is a reason to refit the discharge rate or to mark the incidence drive broken, not a reason to invent a new pathogen. The trip is a boolean from the pattern note. This paper does not invent the trip. A notebook that treats every occupancy residual as a variant is the notebook the broken-transmission rule was written to stop.

Staffed beds are not licensed beds. A licensed count that is not staffed is not a stock this nowcast can fill. CMS cost reports can inform a class-level staffed-bed context. They do not give a daily stock. Unnamed hospital feeds and badge files are refused. The nowcast lives on public or licensed series that can be dated.

The operating sentence names the class, the discharge rate, the lag, and the horizon. It does not name a building. County is still the finest grain on any map. A hospital name may appear in a class table. A pin on a ward is refused.

113. Length of stay by hospital class and the optimistic nowcast

Length of stay is a class object. A children’s hospital, a veterans ward, a large urban teaching hospital, and a critical-access hospital do not share a discharge rate. A public length-of-stay series, when it exists, sets the discharge rate as the reciprocal of the mean stay for that class. When it does not exist, we do not nowcast that class. Inventing a stay from a neighboring class is how an optimistic nowcast is born.

A variant wave that lengthens stay and a discharge rate from last year produce an optimistic occupancy path. On a constant lagged drive the true steady stock is the drive over the discharge rate. If last year’s rate was one fifth and this year’s true rate is one tenth, the nowcast still aims at a stock that is half the true steady state. The path looks calm. The ward is not calm. That gap is a stale coefficient.

Mean stay is not median stay. A class with a long tail of stays will have a mean that the median understates. Using the median as if it were a mean shortens the implied stay and raises the discharge rate. The nowcast then empties the stock too fast. The paper uses the mean when a mean is published, and refuses the class when only a marketing median exists.

Length of stay is also not a clinical recommendation. A sentence that tells a ward how long to keep a patient is clinical advice and is refused. The number used here is a historical class mean from a public series. It is an input to a stock recursion. It is not a protocol.

When stay shortens because a class discharges earlier, the nowcast that keeps last year’s longer stay will overstate occupancy. That error is the mirror of the optimistic variant-wave error. Both errors are stale coefficients. Neither is a new compartment. The residual is a reason to refit the discharge rate on a later clean window, not a reason to add an intensive-care class without an intensive-care series.

CMS cost reports can supply annual or periodic class context. They do not supply a daily stay. Using an annual cost-report stay as a daily discharge rate is a coarse approximation and must be named as such. It is better than inventing a stay. It is not better than a public stay series at the right frequency, when one exists.

A later desk that wants a payroll implication will need a staffed-bed stock and a wage table, not a stay slogan. This paper will hand over a dated occupancy nowcast at seven, fourteen, and thirty days. It will not hand over a wage bill.

The speech rule is the class, the mean stay, the implied discharge rate, and whether the rate is current or stale. A single stay number that claims to cover every hospital in a county is refused unless the county has only one class and that class is named.

114. Speech horizons for incidence, occupancy, and everything else

Seven days: incidence and occupancy may be quoted, with a band. Fourteen days: the same two objects, with a wider band. Thirty days: occupancy as a nowcast, not a season, and not an incidence forecast dressed as occupancy. Ninety days: pipeline events as a list of public stop dates, go dates, completions, and withdrawals, not a path of expected counts. Five years: a computable general-equilibrium scenario with one epidemiological door, not a forecast.

A single percent that claims to cover those five rows is the register language the September revision deleted. It is not a result of this paper. It is not restated here as a skill. The speech rule is the table. If a number is not in the table, it is not spoken.

Incidence at thirty days is not in the table. A thirty-day incidence path is a temptation the broken-transmission gate exists to resist. Reporting can move twice in thirty days. A holiday can sit in the window. A variant can arrive. The paper will nowcast occupancy at thirty days because the stock has inertia. It will not pretend the incidence flow has the same inertia.

Occupancy at ninety days is not in the table. A ward that wants a season should not be handed a ninety-day stock path from this recursion. The recursion is a short-memory filter. It is not a seasonal model. A seasonal competitor must win or lose on a later winter, not on the winter it was trained on.

Pipeline counts at seven days are not in the table. Stage survivals do not move in a week. A public stop date can arrive in a week. That date is an event, not a change in the product. Quoting a new expected terminal count because one name stopped is a misunderstanding of the product. The product is a mean over a starting set. One stop is a realization.

Five-year incidence is not in the table. A computable general-equilibrium sleeve may use a labor path that depends on infectious prevalence, but that path is a scenario door, not a forecast of cases. Anyone who quotes a five-year case path from this paper is quoting a paper this one did not write.

The daily board is allowed to be empty. An empty cell at fourteen days is a result. A dashboard that cannot be empty is a different product. Lateness is preferred to a fake percentage. The first week of a new variant will look late. A quiet influenza week will look precise. Precision on a quiet week is not skill.

Horizons are also a refusal of interpolation. A notebook that draws a smooth line from seven days to five years and calls every point a nowcast has destroyed the table. The objects on those rows are different clocks. They do not interpolate.

115. One epidemiological door into a general-equilibrium sleeve

A computable general-equilibrium sleeve, when used at all, has one epidemiological door: labor supply that falls when infectious prevalence is high. Write that door as a labor index that is a decreasing function of the infectious share. The function is dated. It moves with the incidence path, or it does not move. A labor-supply path that moves before the incidence path is a different model, and it is probably a polarity.

There is no language-model polarity inside a utility function. There is no dollar impact without a named region, a named window, and a named counterfactual. The older register’s billions are not a result of this paper. They are not restated here as a scenario. A sectoral question can use the sleeve. A headline cannot.

The region is not a nation unless a national series is the only series and the paper has already refused a national dollar sentence. In practice the sleeve, if used, is a named region with a named labor table. County remains the finest grain of the epidemiological input. Aggregating counties into a labor region is a stated administrative choice, not a learned region.

The counterfactual is a dated incidence path that was not observed, built from a frozen transmission parameter on a clean window, not from a wish. A counterfactual that sets prevalence to zero everywhere is a different object and must be named as such. A counterfactual that sets prevalence to last year’s quiet week is closer to a usable door.

National dollar impact is refused even when the sleeve is run. The refusal is a speech rule, not a computational accident. A sectoral output change in a named region may be recorded as a scenario table for a later desk that already owns the economic model. It is not a headline. It is not a skill. It is not a thirty-day nowcast.

The door is not occupancy. Occupancy is a hospital stock. Labor supply is a population object. Using occupancy as if it were labor is a category error. Using deaths as if they were labor without a dated fatality and a dated working-age share is another. This paper will not invent that share.

The door is not a trial. A public go date does not move labor supply unless a later note names a dated load change. The pipeline product does not enter the utility function. Mixing a stage keep-rate into a wage equation is how a dashboard number is born.

Five years is the horizon at which the sleeve may be spoken, and only as a scenario. Seven, fourteen, and thirty days are incidence and occupancy horizons. They are not general-equilibrium horizons. Anyone who quotes a seven-day wage bill from this door has left the table.

116. Trial sites as a constrained integer program

Live trial-site allocation, when we do it at all, is a binary program: choose a subset of candidate sites to maximize the sum of site utilities, subject to a budget on costs and a floor on the number of sites. Utilities come from inclusion criteria and historical enrollment. Costs are public or licensed site costs. The budget is a human number. The floor stops a one-site fantasy. An agent that optimizes a trial on a slide is not this program.

The greedy ranking by utility over cost is a bound, not the integer optimum. The linear relaxation is recorded so a later desk can see the gap. We do not hide the gap in an attention plot. A gap that is large is information. It means the greedy story and the integer story are different, and the integer story is the one that can move a site, when a human still owns the budget.

Reinforcement learning for sites stays a research sleeve. It cannot move a site live. The integer program can, when the utilities and costs are dated and the human still signs. A learned utility that cannot be written as inclusion plus enrollment is refused. A utility that includes a binding-affinity headline is refused. A utility that includes a tweet is refused.

Candidate sites are not buildings on a map that the public figure is allowed to pin. County is the finest grain of any public map. A site table may name a center in a methods note that is not a public figure. A public figure that points at a ward is refused.

Inclusion criteria are public protocol text. Historical enrollment is a public or licensed series. If either is missing, the site is not a candidate. Inventing enrollment from a neighboring county is refused. Inventing inclusion from a language model is refused.

The program does not diagnose. It does not enroll a patient. It does not file an investigational application. It does not recommend a dose. It allocates a binary vector under a budget. That is the whole claim.

A later desk that wants a live allocation will be handed the integer vector, the greedy vector, the relaxation, and the gap. It will not be handed an agent. It will not be handed a heat map that pretends the gap is zero.

If the candidate set is smaller than the floor, the program is infeasible and the result is a refusal. Infeasibility is a result. A dashboard that cannot show infeasibility is a different product.

117. Refusals that keep the stack public and non-clinical

Outputs that are refused: badge files, unnamed hospital feeds, a generated molecule presented as intelligence, any sentence that reads as a prescription, a person-level row, a map that points at a building, a binding-affinity headline, a national dollar impact, a thirty-day incidence forecast, a ninety-day occupancy season, and a five-year case path.

The paper does not diagnose. It does not dose. It does not file an investigational application. It does not tell a ward how long to keep a patient. Length of stay is a historical class mean. It is not a protocol. A sentence that could be read as clinical advice is deleted in review, even if it was meant as a description of a stock.

Molecular generation, variational autoencoders for molecular strings, and reinforcement learning for compounds stay in a chemistry notebook until a wet lab and a public filing exist. A single binding-accuracy number is not a result of this paper. It is not restated here as a skill. It is not a covariate in the Cox sleeve. It is not a site utility.

Private hospital feeds that cannot be cited are refused even when they would make the occupancy nowcast look tighter. CMS cost reports and other public or licensed series are the doors. A tighter nowcast that cannot be audited is not a nowcast this paper will sign.

Person-level rows are refused even when they are de-identified in a slide. County is the finest grain. A household, a campus, and a ward are finer. They are different legal objects. A later paper that owns those objects will have to own the law as well.

A national dollar impact is refused even when a general-equilibrium sleeve has been run. The older register’s billions are not results. They are not scenarios in this revision. A sectoral table in a named region may exist for a later desk. A headline may not.

Routes on a marketing site are not sources. This note does not point at a house hallway. The World Health Organization, the Centers for Disease Control and Prevention, the public trial registry, and CMS are the public doors. The next-generation papers and the Cox paper are the method doors. Early pandemic compartment notes are competitors, not a house radius.

Refusal is a result. A dashboard that cannot be empty is a different product. Lateness is preferred to a fake percentage. The first week of a new variant will look late because the broken-transmission gate prefers lateness. That lateness is a refusal of a story.

118. Quiet influenza weeks and the first week of a variant

The note will look precise on a quiet influenza week. Precision on a quiet week is not skill. It is a quiet week. Counts are small or stable, reporting is usually calm, and the windowed fit has little to do. A competitor that advertises that precision as a thirty-day skill is advertising the week, not the model.

The first week of a new variant is late. The broken-transmission gate prefers lateness to a fake percentage. Early growth may have changed. Testing may have changed. Length of stay may be about to change. The gate will not absorb all three into one parameter. A competitor that is early and wrong is a competitor. We do not average with it.

Quiet weeks are also where Poisson and negative-binomial paths can disagree on the sign of a one-week change, and where the renewal competitor can disagree with the compartment path on the sign of the reproduction number minus one. Those disagreements are reported. They are not averaged. They are more informative on a quiet week than a single pretty path.

The first variant week is where a stale discharge rate becomes an optimistic occupancy nowcast. Stay lengthens, last year’s rate stays, and the stock path aims too low. The occupancy residual will trip if the pattern note is working. The trip is a boolean. It is not a new disease name. It is a reason to mark a coefficient broken.

CDC influenza-like series are useful on a quiet influenza week as a context line. They are not a county nowcast. WHO situational counts are useful when a variant is being named in a public note. They are not a reason to jump the grain from county to nation. ClinicalTrials will not move in that first week unless a sponsor stops or starts a public study. CMS cost reports will not move at all. Each door stays on its clock.

A later winter is the test of whether the quiet-week precision was luck. The winter the model was trained on is not a test. An occupancy competitor, including a long short-term memory on the stock, must win or lose on a later winter. Winning on a quiet week inside the training winter is not a win.

The speech on those two weeks is therefore different by design. Quiet week: both competitors, estimated dispersion, no slogan. First variant week: broken marks, empty cells if needed, occupancy residual as a trip, no early radius. Anyone who wants the same sentence on both weeks wants a dashboard this paper will not build.

119. County as the finest geographic grain

County is the finest geographic grain in any public figure. That sentence is a refusal of tracts, campuses, wards, buildings, and households. It is also a refusal of a learned embedding that clusters streets into a new region the reader cannot name. If a later desk wants a finer grain, it wants a different legal object and a different paper.

A named collection of counties is allowed when the collection is an administrative unit that already exists: a state, a hospital referral region that can be built from counties, a public health district. The collection is listed. It is not inferred. Pooling counties to save a nowcast after a testing break is still a break. The pool inherits the messiest testing regime in the pool unless reporting is refit county by county.

A map that points at a building is refused even when the building is a hospital whose class is used in the occupancy nowcast. The class table may name hospital types. The public figure may shade counties. The pin is the problem, not the name in a methods table.

Inter-county flows require a published flow table. A commuter file, a hospital-referral file, or another dated public table can support an open-population variant of the skeleton. A mobility tweet cannot. A phone-panel that cannot be cited cannot. Without the table the counties stay separate, or they are pooled under the administrative rule above.

National series are context. WHO situational counts and national CDC influenza-like lines can sit on a figure as competitors or as context. They do not license a national dollar impact. They do not replace the county nowcast. They do not move the finest grain.

A county with missing days is not filled from its neighbor. Missing days are missing. A silent hospital inside a county is a break for that county’s occupancy, not a reason to borrow a stock from the next county. Borrowing is how a break becomes a region.

The closed-population normalization uses the county population, or the listed collection, as the denominator. Using a national population to normalize a county series is a different model and will break the radius. Using a daytime population without a dated table is a mobility slogan.

A later desk that wants a site-level integer program still does not get a public pin. The site table is not the public figure. The public figure remains a county shade. That split is the grain rule, not a formatting preference.

120. Public doors: situational counts, influenza-like series, trial registries, and cost reports

The public doors are dated. The World Health Organization supplies situational counts and named notices that can mark a testing or case-definition break. The Centers for Disease Control and Prevention supply influenza-like series and related public surveillance lines. The public trial registry supplies registrations, completions, withdrawals, and primary-completion changes. CMS supplies cost reports and related public hospital-class context. Company trial-stop filings supply dated stops and goes when the registry is late.

Each door has a clock. Situational counts and influenza-like series move in days or weeks. Registry dates move when a sponsor updates a page. Cost reports move on a fiscal clock. Joining them is allowed only when a dated edge exists. A registry go date that changes a regional load is an edge. A cost-report stay that updates a discharge rate is an edge. A language-model polarity that claims to join them is not an edge.

Inputs that are not public or licensed and dated do not enter. Badge files do not enter. Unnamed hospital feeds do not enter. A generated molecule does not enter. A binding-affinity number does not enter. A tweet does not enter. A rumor of a hold does not enter until it is a dated public stop.

Revisions are first-class. A situational count that is revised later is a new as-of series. The fit is revised. The old fit is archived. Rewriting history so a walk-forward looks cleaner is refused. The same rule applies to registry dates that move and to cost reports that restate.

The doors are not clinical advice. A situational notice is not a prescription. An influenza-like line is not a diagnosis. A registry page is not an enrollment recommendation. A cost report is not a staffing order. The paper will not translate those doors into a sentence that reads as a protocol.

The doors are not a national dollar impact. A WHO line is not a wage bill. A CDC line is not a sectoral output. A registry date is not a market recommendation. A cost report is not a national hospital bill. The general-equilibrium sleeve, if used, still has one door and still cannot become a headline.

A later desk that wants the equity residual will receive one residualized column after a public, dated trial event from the registry or a filing. A later desk that wants a forecast-stack regressor will receive a dated incidence path, not a refit of the compartment model inside an ensemble. Those two hand-offs are the only joins this paper allows.

No route on a marketing site is a source. This note names the doors as doors. It does not point the reader down a hallway of products.

121. Exposed sojourn, infectious sojourn, and consistency with a generation interval

The latent rate is the reciprocal of the mean exposed sojourn when that sojourn is taken as exponential. The recovery-plus-fatality rate is the reciprocal of the mean infectious sojourn. Those two sojourns, together with the mixing assumption, are the timing content of the SEIRD skeleton. They are given weakly informative priors from the literature of the pathogen class. They are not estimated freely on a two-week county series.

The early growth rate is a function of the transmission parameter and of those sojourns. It is the positive eigenvalue of the infected Jacobian when the radius exceeds one. Identifying transmission from a log-linear segment therefore requires that the sojourns be treated as known, or nearly known, on that segment. Estimating sojourns and transmission together on a short segment is how a notebook manufactures a timing story.

The generation-interval mass used by the renewal competitor must be consistent with those sojourns. A literature mass whose mean is far from the mean latent sojourn plus a fraction of the mean infectious sojourn is a sign that the two competitors are not talking about the same pathogen class. The paper will name both sources and stop, not split the difference.

Exponential sojourns are a convenience. A gamma sojourn with the same mean and a shape greater than one is a different skeleton and a different next-generation calculation. The two-by-two inverse that gives the simple radius assumes the exponential block written in the proofs. Changing the sojourn family without recomputing the radius is a habit the paper refuses.

Death during latency is not in this skeleton. If a later note adds it, the transition matrix changes and the radius changes. Keeping the simple ratio after that addition is refused. The same warning applies if a hospitalized class is inserted with its own sojourn.

A weekend reporting hole is not a sojourn. Treating a Monday bulge as a short generation interval is a reporting artefact. The broken-transmission gate and the weekly renewal competitor exist in part to keep that artefact out of the timing parameters.

The operating report names the literature sojourns, the implied radius formula, and the mean of the frozen generation-interval mass. If those three numbers cannot sit in the same sentence, the window is not fit. That refusal is a timing refusal, not a data outage.

122. Death sink, recovered sink, and when the model is only SEIR

Recovered and dead are different sinks. The recovered class is a living exit. The dead class is not. Mixing them into a single removed class is a modeling choice this paper does not make when a death series exists. The fatality rate then has a series to answer to. The recovery rate has a series or a literature sojourn. They are not one number.

When a death series does not exist, the fatality rate is not invented to fill the sink. The sink is closed and the model is an SEIR. The next-generation radius becomes the transmission parameter over recovery, not over recovery plus a hidden fatality. Keeping the death rate in the denominator after the sink has been closed is a way to make a radius look smaller. It is refused.

A death series that arrives late is a revision. The fit on the revised pair of incidence and deaths is the fit. A fatality rate that was frozen because deaths were missing is unfrozen only on a clean window after deaths exist. It is not unfrozen on the first noisy week of a new death feed.

Infection-fatality is not case-fatality. Case-fatality mixes reporting into the denominator. Using a case-fatality as if it were a fatality rate in the skeleton will eat reporting into the sink. That is another way the pair of reporting and transmission gets lost. If a fatality argument is used as a level anchor, the fatality number must come from a named study and must be treated as infection-fatality, not as a dashboard case-fatality.

The dead class is not a labor door. A general-equilibrium sleeve that wants working-age mortality needs a dated age share and a dated fatality, and even then this paper will not turn it into a national dollar impact. The dead class is not occupancy. Deaths are not beds. Using deaths as a drive for the stock without an admission series is a different filter and must be named.

A later desk that wants a mortality nowcast will not receive one at thirty days from this paper. Thirty days is an occupancy nowcast. Deaths, when they are used, are a sink in the skeleton and a check on the fatality rate. They are not a headline and they are not clinical advice.

The speech rule is: name which sinks are open, name the series that opened them, and name the radius formula that follows. If the death sink is closed, say SEIR. If it is open, say SEIRD. Do not say SEIRD when μ is a decoration.

123. Log-linear early growth and the infected Jacobian

Near the disease-free state the susceptible share is near one and the infected block is a linear system. The Jacobian of that block is the sum of the transmission matrix and the transition matrix. Its characteristic polynomial has a constant term that is negative precisely when the next-generation radius exceeds one. There is then a positive real eigenvalue. That eigenvalue is the early exponential growth rate of incidence.

The growth rate is identified from a log-linear segment of a dated count series only if reporting is stable on that segment. A testing-policy break that is not marked will be eaten by the growth rate and then by the transmission parameter. The worked numbers in the paper show the pattern: a clean radius of two and a doubling time of about eight days can become a testing artefact near one and a half if reporting halves and the window is not split.

The growth rate depends on transmission and on the sojourns. It does not depend on reporting. The observed level does. That is the identifiability split again. A notebook that fits a log-linear slope and a level and then reports both a radius and a reporting fraction without an external level anchor has reported a pair the segment cannot identify.

A short segment that includes a weekend hole is not a log-linear segment. A short segment that includes a holiday is not a log-linear segment. The paper would rather have no growth rate than a growth rate that is a calendar. Weekly totals can be used when daily reporting is a calendar. The renewal competitor is often the better object on those weeks.

The doubling time is the log of two over the growth rate. It is a translation of the eigenvalue, not a new parameter. Quoting a doubling time after the window has left the disease-free neighborhood is a misuse. Once a material share of the county is no longer susceptible, the linearization is a poor description and the growth rate is no longer the eigenvalue of the disease-free Jacobian.

A later desk that wants a national doubling time will not receive one. County is the grain. A national line may be a competitor. It is not an operating eigenvalue. A national dollar sentence built from a doubling time is refused twice: once as a grain error and once as a dollar error.

The operating report names the segment dates, the slope, the implied growth rate, the implied radius given frozen sojourns, and whether reporting was marked stable. If reporting was not stable, the slope is not reported as a radius.

124. Series revisions, as-of stamps, and the windowed fit

Parameters are estimated on a sliding window, not on a global pandemic. A global fit hides the week the testing policy changed. The objective is a Poisson or negative-binomial likelihood on new counts, with latent, recovery, and fatality rates given weakly informative priors from the literature and transmission free. A parameter without a series identifier is deleted in review.

If the series is revised, the fit is revised. The as-of stamp is part of the object. A nowcast produced on Tuesday from Monday’s file is not rewritten on Thursday when Monday’s file is revised, except as a new as-of object. Archiving both is how a later desk audits the board. Rewriting Tuesday in place is how a walk-forward becomes a memoir.

Window length is not a free art. A documented testing-policy change inside the window marks transmission broken and shrinks the window. If the remaining clean stretch is too short to identify transmission given the sojourn priors, the nowcast is refused. Borrowing a window length from a neighboring county is refused. Neighboring counties can have different calendars.

A silent hospital is a revision of the occupancy series and, when the hospital was contributing severe incidence, a revision of the incidence series. It is marked. It is not interpolated. Interpolation across a silence is how a break becomes a smooth path.

WHO situational counts are revised. CDC influenza-like series are revised. Registry dates move. CMS cost reports restate. Each revision is a new as-of. The daily board pulls the latest public file and keeps the previous file. It does not pretend the latest file was known last week.

The forecast stack may receive a dated incidence path as a regressor. It may not receive a path that was silently revised without an as-of stamp. The equity stack may receive one residualized column after a public dated trial event. It may not receive a column that was back-filled when a registry date moved. Those two sentences are the same revision rule in two receiving desks.

A later desk that wants a prettier expanding-window score will have to live with the archived as-of objects. Dropping revision weeks to raise a score is a different product.

125. What the daily board is allowed to say, and what it must not say

The daily board pulls dated public series only. It marks transmission broken if testing policy moved. It nowcasts incidence and occupancy at seven and fourteen days, and occupancy at thirty days as a nowcast. It reads overnight trial stops and completions from public registries. It hands one residualized column to an equity stack only if the edge is dated. It hands a dated incidence path to a forecast stack as a regressor, not as a license to refit the skeleton inside an ensemble.

It does not diagnose. It does not dose. It does not file an investigational application. It does not quote a binding-affinity headline. It does not quote a national dollar impact. It does not nowcast incidence at thirty days. It does not nowcast occupancy at ninety days. It does not shade a tract. It does not pin a building. Lateness is preferred to a fake percentage.

The board is allowed to be empty. An empty cell is a result. A dashboard that cannot be empty is a different product. The first week of a new variant will have empty cells if the gate cannot freeze a clean transmission parameter. A quiet influenza week will have numbers that look precise. The board will not use the quiet week as an advertisement for the variant week.

Overnight registry reads are event lists. They are not a change in the stage product. They are not a Cox refit unless a pre-declared sleeve says so. A single stop is a realization. A cluster of stops in one indication is a reason to remember intra-class correlation. It is not a reason to raise the mean product.

The board does not speak for a ward. Occupancy is a class stock. A hospital name in a class table is not a staffing order. A discharge rate is not a length-of-stay protocol. Anyone who copies a board sentence into a clinical memo has left this paper.

The board does not speak for a portfolio. A public go date may become a dated residual column after another desk residualizes it. The board does not size a clip. The board does not quote a binding number as if it were a trial result.

The speech rule is the horizon table plus the refusal list. If a sentence is not in those two objects, the board does not say it.

126. Stochastic incidence when county counts are small

When counts are small, demographic noise is an Itô term on the infectious class, with a diffusion that scales as the square root of the infectious count. That is a Feller-type noise, not a language-model residual. It is used when the count is small enough that a deterministic ordinary differential equation is a lie. It is not used to fatten a band on a large national series. County is the grain on which smallness is judged.

The noise scale is estimated from the residual of the deterministic fit. If it wants to be huge, the horizon is shortened. We do not decorate the stochastic equation to keep a long path. Seven days may survive. Fourteen days may not. Thirty-day incidence is not in the table anyway. Thirty-day occupancy still uses the stock recursion, which has its own residual, not this diffusion.

The noise is not a reason to quote a tighter path. It is a reason to quote a wider one. A notebook that adds the noise and then reports a tighter band has used the noise as decoration. The honest use is a wider nowcast on a small county series, or a refusal when the scale is huge.

The deterministic SEIRD remains the skeleton for the basic reproduction number and for the next-generation inverse. The noise is a nowcast sleeve, not a new radius. Linearizing a stochastic system and calling the Lyapunov exponent a radius is a different paper. This paper will not do that on a twelve-count county week.

Observation overdispersion and process noise are different. The negative-binomial dispersion sits on the count given the mean. The Feller term sits on the infectious state. Using one as a substitute for the other is a confusion. Both can be present on a small series. Both must be named.

A later desk that wants a national band will not receive this sleeve. A national series is not small in the sense that makes demographic noise the right story. Overdispersion may still be right. The Feller term is not automatically right because a figure looks scientific.

The operating sentence names the estimated noise scale, the horizon that survived, and whether the deterministic path is still the mean. If the scale is huge, the sentence is a refusal.

127. Mass-action as a mixture, and when a contact matrix may enter

A fitted transmission parameter under mass-action is a mixture whenever the county is not mixing that way. Age groups, occupations, households, and schools can all sit inside one number. The number can still be useful as a windowed reduced form. It cannot be useful as a variant name. When the mixture changes because a school calendar changed, the honest mark is a mixing break, not a new pathogen.

A contact matrix may enter only when it is named, public, and dated. The matrix is frozen for the window. Transmission may then be a scalar on top of the matrix, or a small set of group transmissions if the series can support them. Estimating a full matrix from a short county series is refused. Estimating the matrix and the scalar on the same counts is refused.

Preferential mixing, assortative mixing by age, and layered household-workplace-community constructions are special cases of a named matrix or a named set of matrices. A slogan that says the county is assortative, without a table, is not a matrix. The paper will keep mass-action rather than pretend.

Frequency-dependent incidence is the other reduced form. It is used when meetings are closer to a capacity than to a density. A school is often closer to frequency dependence. A county census denominator with school meetings is a mixture of the two reduced forms. The paper will not switch reduced forms mid-window to chase a residual.

Open-population flows are not a contact matrix. They are a flow table between named counties. A gravity slogan is not a table. If the table exists, the skeleton changes and the next-generation operator changes. If it does not, the county stays closed or the nowcast is refused.

The mixture view is also why neighboring counties should not donate a transmission parameter. Their mixtures can differ even when the pathogen does not. A donated parameter is a donated mixture. The broken-transmission gate already forbids absorbing a testing break. This note forbids absorbing a mixing difference the same way.

A later desk that wants a school-closure contrast will need an identification strategy and a named matrix. This paper will not supply a treatment effect from a mixture residual.

128. Intra-class correlation, multilinearity, and what stress is allowed to mean

Under independent stage risks the variance of the terminal count is the usual Bernoulli variance of the product. Dependence across assets in the same indication increases that variance. It does not raise the mean. Intra-class correlation is therefore a reason to expect clusters of failures or clusters of survivals. It is not a reason to rewrite a keep-rate.

The product map is multilinear. Halve any one keep-rate, halve the product. There is no hidden compensation. Stress is a lower keep-rate vector taken from published stress bands, not a hope that a later stage will make up an early miss. A board that makes up a later stage is inventing a dependence the independent-Bernoulli model does not have.

The default product near 0.0059 and the stress product near 0.00049 are literature-band arithmetic. They are not a portfolio forecast. A fund that treats 0.59 expected named entities from a hundred discovery assets as a pipeline of a hundred late-stage assets is reading a different table. The starting set is a discovery set. The terminal set is a mean.

Correlation estimated from a handful of names in one indication is weak. Weak correlation is reported. It is not pinned at zero to make the variance look kind, and it is not pinned at one to make a cluster look inevitable. If the Cox residual on public dates is noise, the desk reports the stage table, the product, the stress product, and stops. It does not estimate a rich correlation matrix from three names.

Indication class is a public label. It is the same class used as a Cox covariate. A learned embedding that rewrites indication class to fatten a correlation is refused. Sponsor size is not an indication. Mixing sponsor and indication into one intra-class factor is how a notebook manufactures a cluster.

A ninety-day event list can show a cluster. The cluster is a realization. It is compatible with a positive intra-class correlation. It is not a change in the default product unless a later note revises the literature bands. Revising the bands because this quarter was painful is a different product.

The operating sentence is the default product, the stress product, the intra-class note, and whether Cox beat noise. Anyone who wants a single pipeline percentage has left the table.

129. Occupancy residuals, later winters, and competitors that must lose honestly

A trip on the occupancy residual is a reason to refit a discharge rate or to mark the incidence drive broken. It is not a reason to invent a new disease. The trip is a boolean from the pattern note. This paper does not invent the trip. A notebook that names a variant from an occupancy residual has skipped the incidence series and the testing gate.

The linear stock recursion is a short-memory filter. It will look skilled on a quiet week and late when stay changes. That is the same speech pattern as the incidence note, on a different clock. Thirty days is the longest occupancy nowcast. A competitor that quotes a ninety-day stock path is a different product.

A long short-term memory on occupancy is a competitor. It must win or lose on a later winter, not on the winter it was trained on. Winning on Tuesdays inside the training winter is not a win. Mean-week error is not the test. Error on peaks is the test. A model that is good on Tuesdays and late on the first freeze is a model that will look skilled on a mean-week score and useless to a class stock.

Incidence as a dated regressor may enter an occupancy competitor. A refit of the SEIRD skeleton inside that competitor is refused. The compartment paper owns the compartment. The occupancy paper owns the path. This note is the compartment paper’s description of the stock it is willing to drive. It is not a license for another desk to refit transmission inside a neural stock.

CMS cost reports can tell a later winter whether a class’s staffed-bed context moved. They cannot tell a daily residual why it tripped last Tuesday. Using a cost report to overwrite a daily residual is a clock error.

County remains the grain of the public figure. A class stock may be named. A building may not be pinned. A later winter test that is only available at a building feed we cannot cite is not a test this paper will use.

The operating report names the residual, whether it tripped, whether the discharge rate was stale, and how a competitor did on a later winter. It does not name a variant from the residual alone.

130. What a fitted transmission parameter is not

A fitted transmission parameter is a number on a dated window under a named mixing assumption and a named reporting regime. It is not a variant name. It is not a policy evaluation. It is not a reason to close a ward. It is not a reason to size an equity clip. It is not a binding-affinity headline. It is not a national dollar impact.

A later desk that wants a policy contrast will need an identification strategy from a forecast paper: a back-door, a front-door, or an instrument. This paper does not supply one. Incidence as a dated regressor is the only object that paper is allowed to take from here. A refit of the skeleton inside an ensemble is refused.

A later desk that wants an equity column will need a public stop or go date. A fitted transmission parameter is not that date. Residualizing a radius into a return is a category error. Residualizing a dated event is a different desk’s job.

A later desk that wants a clinical sentence will not receive one. The parameter does not diagnose, dose, or enroll. A sentence that says a county should do something because transmission moved is clinical or policy advice this paper will not write.

A later desk that wants a five-year labor path may use infectious prevalence through the one general-equilibrium door. It may not use the transmission parameter as a productivity shock. The door is prevalence, dated. The parameter is a windowed reduced form.

When the window is marked broken, the parameter is not a number the board speaks. An old clean parameter may be frozen to refit reporting. A broken parameter is not frozen as a story. Quoting a broken parameter with a wide band is still quoting a testing artefact if the break was reporting.

The speech rule is the window, the mixing assumption, the reporting mark, the implied radius, and the final-size check. Anything louder is a different paper.

131. Efron ties, censoring, and what a registry date can bear

Registry dates are days. Many names share a day. Efron’s approximation is how the Cox partial likelihood treats those ties. A continuous-time fantasy that spreads the ties uniformly through the day is a different model. We do not invent a continuous clock for a date that is only a day. Breslow’s approximation is a competitor and may be recorded. It is not a reason to hide the tie count.

Censoring is the usual right-censoring of names that have not stopped or completed by the as-of stamp. A silent registry page is censoring, not a hidden stop. Imputing a stop from silence is refused. A primary-completion date that moves forward is a revision of a covariate clock, not automatically an event. The event is a stop, a completion, a withdrawal, or a termination that the registry or a filing stamps.

Left truncation appears when a name enters the risk set only after a public start. Using a discovery date as if it were a trial start is a clock error. The Cox sleeve on public trial dates starts when the registry says the trial started, or it does not start.

Sponsor size and indication class are fixed covariates on the window. Time-varying covariates are allowed only when they are public and dated: a stage change on the registry, a sponsor-size bucket change after a public filing. A tweet that claims a stage change is not a covariate. A binding-affinity number is not a covariate.

If the partial-likelihood residual is noise, the stage table is the report. Noise means the covariates do not move the hazard beyond what a later note’s residual check allows. The desk does not then draw a smooth survival curve through three names. The figure is steps. The steps show the risk set.

A later desk that wants a live site allocation will not read a hazard ratio as a site utility. Site utility is inclusion plus historical enrollment. A hazard ratio is a description of historical exits. Mixing them is how an agent slide is born. The integer program and the Cox sleeve stay separate.

The operating sentence names the event definition, the tie rule, the risk-set rule, and whether the residual was noise. It does not name a molecule. It does not name a dose.

132. Labor paths, dated prevalence, and why polarity cannot open the door

The only epidemiological door into the economic sleeve is labor as a function of infectious prevalence. The function is decreasing, dated, and frozen for the scenario. It is not estimated from a wage series on the same window that produced the incidence path. Estimating both on the same window is leaking. The incidence path is an input. The labor function is a scenario assumption.

A language-model polarity cannot open the door. Polarity is not prevalence. A news leftover that moves before cases move is the usual failure. If the labor path moves first, the sleeve is marked as a polarity sleeve and is not this door. The September revision exists in part to keep that leftover from being rewritten as a dollar impact.

Prevalence is a county object first. Aggregating to a labor region is a listed administrative choice. Using a national prevalence to move a county wage is a grain error. Using occupancy to move labor is a category error. Using a trial go date to move labor is a category error unless a later note names a dated load change.

The counterfactual must be a dated incidence path. Setting prevalence to zero is a named extreme. Setting it to a quiet influenza week is often the more honest extreme. Setting it to a first variant week that the gate refused to nowcast is not allowed. A refused nowcast cannot become a scenario door by renaming.

National dollar impact remains refused. A sectoral table in a named region may be written for a later desk that already owns the economic model. The table is a scenario. It is not a thirty-day nowcast. It is not a skill. It is not a headline.

Five years is the only horizon at which this door is spoken. Seven, fourteen, and thirty days stay with incidence and occupancy. Anyone who wants a seven-day labor index from prevalence has left the speech table.

The operating sentence names the labor function, the prevalence path, the region list, and the counterfactual. If any of those four is missing, the sleeve is not run.

133. Site utilities, greedy gaps, and why an agent cannot own the budget

Site utility is inclusion criteria plus historical enrollment. Inclusion is public protocol text. Enrollment is a public or licensed history. A utility that cannot be written that way is refused. A utility that includes a tweet, a polarity, a binding number, or a generated molecule is refused. A utility that includes a hazard ratio from the Cox sleeve is refused. Those objects live on other clocks.

Cost is a public or licensed site cost. Budget is a human number. The floor on the number of sites is a human number. If the human will not name the budget and the floor, the program is not run. An agent that invents a budget on a slide is not a human.

The greedy ranking by utility over cost is computed and stored. The integer optimum is computed and stored. The linear relaxation is computed and stored. The gap between greedy and integer, and the gap between integer and relaxation, are the audit. Hiding the gaps in an attention plot is refused. A large gap means the slide story and the live story differ.

Infeasibility is a result. If every affordable set is smaller than the floor, the program says so. If every set that meets the floor blows the budget, the program says so. A dashboard that converts infeasibility into a heat map of almost-sites is a different product.

Reinforcement learning for sites stays a research sleeve. It cannot move a site live. It cannot own the budget. It cannot replace inclusion text. A later paper may compare a learned policy to the integer program on historical enrollments. That comparison is a competitor, and it must be run on later trials, not on the trials that trained the policy.

County remains the public grain. The site table is not a public pin. A methods table may name centers. A public figure may not point at them.

The program does not enroll, dose, diagnose, or file. It returns a binary vector, or it returns a refusal. That is the whole live claim.

134. Writing the next-generation inverse and the two-compartment radius in symbols

The infected compartments are the exposed class and the infectious class. New infections and transitions at the disease-free equilibrium are the matrices T and Σ written in the proofs. Only the exposed-from-infectious entry of T is a new infection: one infectious person produces β latent infections per unit time in a fully susceptible county. Σ holds progressions and removals. The next-generation matrix is minus T times the inverse of Σ. The basic reproduction number is the spectral radius of that matrix: R0 equals ρ of minus T Σ inverse, and for this two-compartment infected block that radius equals β over γ plus μ.

The inverse of Σ exists because the latent rate is positive and the sum of recovery and fatality is positive. Direct inversion produces a non-positive matrix whose negative is a non-negative sojourn matrix, as required of an M-matrix inverse. The product minus T Σ inverse then has a single nonzero eigenvalue, β over γ plus μ. That is the whole two-by-two calculation. Putting recovery into T, or putting new infections into Σ, produces a different matrix and a meaningless radius.

If a hospital class H is inserted between the infectious class and the sinks, T and Σ grow. R0 is recomputed as a spectral radius. The two-compartment ratio is not kept by habit. If the death sink is closed, μ leaves the denominator and the model is SEIR. If an asymptomatic class is added without a series, the pair of reporting and transmission becomes a worse identification problem, not a better radius.

The disease-free point is locally stable when R0 is less than one and unstable when R0 is greater than one. The early growth rate is the positive eigenvalue of T plus Σ when R0 exceeds one. That eigenvalue is not R0. It is a function of β, σ, γ, and μ. Quoting the eigenvalue as if it were the radius is a translation error.

A time-varying β produces an instantaneous Rt equal to βt over γ plus μ under the same two-compartment block. Rt is a local diagnostic on a named county series. It is not a national score. It is not a dollar. It is not a reason to close a ward.

The operating sentence names T, Σ, the inverse, the radius, and whether the infected block is still two-by-two. If the block has grown, the sentence names the new spectral radius and stops using the simple ratio.

135. Broken-β, reporting ρ, and why the pair is the gate

β is fit to a dated incidence series and frozen for the forecast window. If the series is revised, the fit is revised. If testing policy moves, β is marked broken. Observed counts scale with a reporting fraction ρ. Early growth identifies a function of β given a generation interval. The level scales with ρ. The pair (ρ, β) is not identified from counts alone. A testing change is a change in ρ. Treating it as a change in β is the failure the gate exists to stop.

Holidays, case-definition changes, and silent hospitals are the same class of break. The gate is: mark the window, freeze β from the last clean window, refit only ρ, or refuse the nowcast. Absorbing the break into β and then quoting R0 equals β over γ plus μ is refused, because that R0 is then a testing artefact.

A seroprevalence point, when public and dated, can anchor ρ. We do not invent one. A death-based infection-fatality argument can be a second anchor only if the fatality is a named study number, not a tuned dashboard case-fatality. Without an anchor the paper reports a product that maps to observed incidence and refuses the split.

The windowed algorithm is the operational form of the gate. For each date, form a window, maximize the count likelihood in β, and if a documented testing-policy change sits in the window, mark β broken and shrink the window. Documented means a WHO notice, a CDC case-definition update, a public hospital outage, or another dated public note. A rumor is not documentation.

The first week of a new variant is why the gate prefers lateness. A quiet influenza week is why a competitor can look skilled without being skilled. The board will not use the quiet week as an advertisement for the variant week, and it will not quote a broken β with a wide band as if the band repaired the pair.

A later desk that wants a policy contrast on testing needs an instrument this paper does not invent. The identifiability note is a refusal. The broken-β mark is a boolean. The receiving action is already written: do not treat the mark as a new pathogen.

136. Estimating φ, refusing φ equals one, and the negative-binomial score

If counts are over-dispersed relative to Poisson, the likelihood is negative binomial with mean equal to the SEIRD incidence and dispersion φ that is estimated. The score for β is a weighted residual. The weight at a date is φ over φ plus the mean. Weights shrink when the mean is large. That is the overdispersion. Setting φ to infinity recovers the Poisson score. Setting φ to one to make a figure pretty is refused. φ is not set to one.

We use a library log-gamma implementation. A handwritten logarithm that overflows at large county-week totals is not a likelihood. The mean stays the SEIRD incidence after any frozen ρ. Replacing the mean by a spline and calling the leftover a compartment model is a different paper.

A short weekly series can leave φ weakly identified. Weak identification is reported. It is not a reason to pin φ at one. A profile that is flat in φ is a reason to quote both the Poisson path and a conservative negative-binomial path. When those paths disagree on the sign of a one-week change, both are reported. They are not averaged.

The worked weekly example in the paper is the score in numbers: a late week with a larger mean is down-weighted relative to an early week with a smaller mean. A notebook that then draws a tighter band than Poisson has used the negative binomial as decoration. The honest band is wider when φ is finite.

Process noise on the infectious class is not φ. φ is observation overdispersion. The Feller square-root term is process noise. Both may sit on a small county series. Neither is a language-model leftover. Neither is a reason to quote a national band.

The operating report names the estimated φ, the Poisson competitor, and the sign agreement on the next seven-day change. Fourteen-day incidence uses the same likelihood on a longer horizon and a wider band. Thirty-day incidence is not spoken. Anyone who sets φ to one in a thirty-day figure has left both the likelihood and the speech table.

137. The bed recursion in symbols, class lags, and the thirty-day occupancy nowcast

The occupancy stock is written H sub t plus one equals one minus d times H sub t plus a times a lagged infectious or incidence drive. In compact form, H_{t+1} = (1-d) H_t + a I_lag. d is the discharge rate for a named hospital class. a is the admission coefficient for that class. I_lag is a lagged incidence or infectious count, not a polarity and not a trial date. The recursion is a linear filter. It is not a new infected compartment unless H has been inserted into the next-generation block, in which case R0 is recomputed.

d is the reciprocal of mean length of stay for the class when a public stay series exists. A children’s hospital, a veterans ward, a teaching hospital, and a critical-access hospital do not share d. If class is unknown, the name is not nowcast. If stay lengthens and last year’s d is kept, the nowcast aims at a/d while the true steady stock is higher. That is an optimistic nowcast from a stale d, not a new pathogen.

The lag is frozen from an earlier winter or from a public admission-delay note. Estimating the lag on the same short winter that is being nowcast is leaking. CMS cost reports may inform class context. They do not set a daily d. Unnamed feeds are refused.

Seven and fourteen days are quoted for incidence and occupancy. Thirty days is a nowcast for occupancy only. The stock has inertia. The incidence flow does not get the same thirty-day courtesy. Ninety days is a pipeline event list. Five years is a labor-door scenario. An occupancy season is refused.

A residual trip refits d or marks the drive broken. It does not name a variant by itself. A later-winter competitor, including a long-memory stock model, must lose or win on a later winter and on peaks, not on mean Tuesdays.

The operating sentence names the class, d, a, the lag, the horizon, and whether thirty days is being spoken as occupancy. If someone quotes H_{t+1} as a thirty-day incidence path, they have left the table.

139. Closed population and what leaks

The SEIRD states are shares of a closed population when that approximation is tolerable. A city that is not closed leaks people across the boundary. The leak is a mixing and a reporting problem, not a new pathogen. Treating a leak as a variant is how a nowcast becomes a story.

Spoken, the object is a normalized count in a named region on a dated window. The region is a county or coarser. A building is not a region. A person is not a compartment.

Misread, the closed-population share is treated as a headcount that can be summed into a national dollar. It cannot. The September revision deleted that dollar language. It stays deleted.

Refused: a map that points at a building, a person-level row, and a compartment invented to absorb a boundary leak.

A later desk that wants an open population will name births, deaths from other causes, and migration. This paper does not invent those flows to keep a path pretty.

140. Recovered and dead as different sinks

Recovered and dead leave the infectious class by different doors. Mixing them into a single removed class is a modeling choice this paper does not make when a death series exists. When a death series does not exist, the fatality rate is not invented. The model becomes SEIR.

Spoken, R and D are different columns. μ is a rate on a dated death series, not a hope. γ is a recovery rate on a dated recovery or removal-minus-death construction that is named.

Misread, a single removed class is quoted as if it were this paper’s SEIRD. It is not. A SEIRD that pretends to know μ without deaths is a costume.

Refused: inventing μ to fill a sink, and quoting a death path as a recovery path.

If a later desk adds a hospitalized class between I and D, both the new-infection matrix and the transition matrix grow. The two-compartment R0 formula is not kept by habit.

141. Mass-action mixing

Mass-action incidence assumes homogeneous mixing in the normalized population. A city that is not mixing that way produces a β that is a mixture. The mixture is not a new variant. It is a broken mixing assumption.

Spoken, β is a transmission parameter on a dated window under a named mixing assumption and a named reporting regime. It is not a variant name and not a policy evaluation.

Misread, a change in mixing is a change in pathogen. That sentence is how a testing break becomes a variant headline.

Refused: a contact matrix invented from a mobility tweet, and an age structure without a named matrix.

Age structure, if used at all, is a later note with a named contact matrix. This paper does not invent one.

142. The broken-β gate as a speech rule

Testing-policy moves, holidays, case-definition changes, and silent hospitals mark β broken. The gate is: freeze β from the last clean window and refit only reporting, or refuse the nowcast. Absorbing the break into β and then quoting R0 is refused.

Spoken, a broken window is a marked window. Refusal of a nowcast is a result. A dashboard that cannot be empty is a different product.

Misread, a testing change is a transmission change. That is the identifiability failure: early growth depends on β, the observed level depends on reporting, and the pair is not split without a dated level anchor.

Refused: inventing a seroprevalence point, and quoting R0 from a window that contains an unmarked testing break.

Lateness is preferred to a fake percentage. The first week of a new variant is late on purpose. A competitor that is early and wrong is not averaged in.

143. Identifiability of reporting and transmission

Early growth of the infected linearization depends on β and the progression rates, not on reporting. The observed level scales with reporting. Without a public dated seroprevalence point or another level anchor, the pair is not identified.

Spoken, the identified object is a product that maps to observed incidence. The split is not identified. That is why the broken-β rule exists.

Misread, a level change is a β change. A notebook that fits both freely on counts alone has not identified either.

Refused: a house reporting number that fills a path, and a seroprevalence point that was not public and dated.

A later desk that wants a policy contrast will not get it from a fitted β. They need an identification strategy from the forecast paper. This paper supplies a dated incidence path, not a do-operator.

144. Next-generation split of T and Σ

New infections belong in T. Progressions and removals belong in Σ. Putting recovery into T produces a radius that is not R0. For the two-compartment infected block, R0 equals β over γ plus μ, which is the spectral radius of minus T Σ inverse.

Spoken, R0 is a spectral radius at the disease-free equilibrium. Local stability when R0 is less than one and instability when it is greater than one is the van den Driessche–Watmough theorem, not a slogan.

Misread, R0 is a national forecast or an attack-size claim without the final-size check. The final-size equation is a consistency check, not a forecast.

Refused: keeping β/(γ+μ) after adding a hospitalized class, and quoting R0 from a window with an unmarked reporting break.

The inverse of Σ is written so a later desk can change the latent graph without guessing. Mean sojourn in E is 1/σ. Mean sojourn in I is 1/(γ+μ).

145. Final-size consistency

In an SIR reduction the attack size z solves z equals 1 minus e to the minus R0 z. For R0 greater than one there is a unique root in the open unit interval. At R0 equals 2 that root is about 0.80.

Spoken, the equation is a check on a fitted R0 against an observed attack. A fitted R0 of 2 and a 15 percent observed attack cannot both be true under the closed SIR reduction.

Misread, the transcendental equation is a national forecast. It is not. This paper does not use it as one.

Refused: a scenario that quotes the 0.80 as a city’s future, and a fitted R0 that the series cannot support.

Strict concavity of 1 minus e to the minus R0 z minus z forbids a second root. The algebra is in Section 21. The speech is here.

146. Negative-binomial overdispersion

If counts are over-dispersed relative to Poisson, the likelihood uses a negative binomial with mean equal to the SEIRD incidence and a dispersion φ that is estimated. Weights φ over φ plus μ shrink when the mean is large.

Spoken, φ is estimated. The path is not tightened by setting φ to one or to infinity. A library log-gamma form is used so large counts do not overflow a handwritten log.

Misread, a Poisson path that looks tight is more honest. It is less honest when the counts are over-dispersed. Tightness is not honesty.

Refused: setting φ to make a figure pretty, and averaging Poisson and NB nowcasts that disagree on a one-week sign.

When they disagree, both are reported. Averaging them is how a dashboard number is born.

147. Pipeline product and stress

Expected terminal count is the starting set times the product of stage keep-rates under independent stage risks. Default literature-band keep-rates multiply to about 0.0059. Stress multiplies to about 0.00049. One hundred discovery assets expect 0.59 terminals under the default and 0.05 under stress.

Spoken, those products are literature-band arithmetic, not a fund. The map is multilinear: halve one stage, halve the product. There is no hidden compensation.

Misread, a hundred discovery assets are a hundred late-stage assets. They are not. A board that makes up a later stage after an early miss invents a dependence the independent model does not have.

Refused: a house π that fills a fund, and a mean quoted without the variance that intra-class correlation fattens.

Intra-class correlation in an indication fattens variance. It does not restore a killed stage. If the Cox residual on public dates is noise, the stage table is the report.

148. Cox on public stop and go dates

Spoken, the partial likelihood cancels the baseline. Ties use Efron’s approximation because a date is a day, not an invented hour. The events are public registry dates.

Misread, a smooth survival curve through three names is this paper’s Kaplan–Meier. The figure here is steps. A parametric baseline drawn to look smooth is a different model and must say so.

Refused: a tweet covariate, a language-model polarity on a press release as a covariate, and an invented continuous clock.

An asset listed active and dead is a limitation. It is not a reason to invent a completion.

149. Staffed beds as a stock

Occupied staffed beds rise with admissions and fall with discharge. Coefficients are hospital-class specific. A children’s hospital is not a VA ward. If class is unknown, that name is not nowcast.

Spoken, d is the reciprocal of mean length of stay for that class when a public LOS series exists. Seven and fourteen days are quoted for occupancy. Thirty days is a nowcast, not a season.

Misread, a single occupancy percent is a thirty-day season. That speech is forbidden by the horizon table. An LSTM that is beautiful on the training winter is not skilled until it wins on a later winter.

Refused: nowcasting an unknown class, using last year’s d in a variant wave that lengthened stay without marking the coefficient broken, and a refit SEIRD inside the occupancy member.

If d halves and the drive is unchanged, the steady state doubles. A nowcast that still aims at last year’s stock is optimistic. That is a stale discharge rate, not a new disease.

150. Horizons as a speech table

Spoken, only the row that matches the object and the horizon. A single percent that claims to cover those five rows is the register language the September revision deleted.

Misread, a 30-day occupancy nowcast is a seasonal forecast. It is not. A 90-day event list is a path. It is not.

Refused: blending those rows into one skill number, and restating the deleted 30-day skill language as a scenario.

The daily board pulls dated public series, marks β broken if testing moved, nowcasts at the allowed horizons, and reads overnight trial stops. It does not diagnose, dose, or file an IND.

151. One door into an economy

A CGE sleeve, when used at all, has one epidemiological door: labor supply that falls when infectious prevalence is high. There is no language-model polarity inside a utility function.

Spoken, a sectoral question can use the sleeve. A headline cannot. There is no dollar impact without a named region, a named window, and a named counterfactual.

Misread, the older register’s billions are a result of this paper. They are not. They are not restated as a scenario.

Refused: a polarity in a utility function, a national dollar without the three names, and a labor path that moves before the incidence path.

The door is dated. A labor-supply path that leads incidence is probably a polarity.

152. Public doors and refusals

Inputs are WHO situational counts, CDC influenza-like series, ClinicalTrials registrations, CMS cost reports, and company trial-stop filings, public or licensed and dated.

Spoken, the doors are those series and those papers: van den Driessche and Watmough, Diekmann, Heesterbeek and Metz, Cox. No product route is a source.

Misread, a badge file or an unnamed hospital feed is an input. It is not. County is the finest grain.

Refused: person-level rows, building maps, SMILES as intelligence, binding-affinity headlines, RL that moves a site live.

Live site allocation, if it happens, is a binary knapsack with a recorded LP gap. The human still owns the budget.

153. Quiet week and first variant week

Precision on a quiet flu week is a quiet week. It is not skill. The first week of a new variant is late because the gate prefers lateness to a fake percentage.

Spoken, refusal is a result. The board still reads stops. It still marks β broken. It does not invent a compartment to fill an empty cell.

Misread, lateness is a failed product. Lateness is the product of the gate. A competitor that is early and wrong is a competitor.

Refused: averaging with the early-wrong competitor, and a dashboard that cannot be empty.

A holiday, a case-definition change, and a silent hospital are the same class of break as a testing-policy move. The gate does not negotiate.

154. What a fitted β is not

A fitted β is a transmission parameter on a dated window under named mixing and named reporting. It is not a variant name, a policy evaluation, a ward closure, or an equity clip.

Spoken, equity may receive one residualized column after a public dated trial event. A fitted β is not that event. The forecast paper may receive dated incidence as a regressor, not a refit SEIRD.

Misread, a fitted β is a do-operator. It is not. A policy contrast needs identification from the forecast paper.

Refused: sizing an equity clip from β, closing a ward from β, and filling a contrast cell with β.

Drift in a pathogen residual is an alarm. The written action is mark β broken. This paper does not invent the trip.

155. Renewal competitor on weekly counts

When counts are weekly, a renewal with a generation-interval mass from the pathogen class is sometimes more stable than integrating the ODE daily and aggregating. The mass is frozen for the window.

Spoken, the renewal is a competitor, not a replacement. If the two disagree on the sign of Rt minus one in a quiet week, both are reported.

Misread, averaging a compartment and a renewal is honesty. It is how a dashboard number is born.

Refused: estimating the generation interval from a price series, and refitting the mass on the same counts that produce Rt.

The generation-interval literature of the pathogen class is the door. A price series is not.

156. Stochastic incidence when counts are small

When counts are small, a square-root diffusion on the infectious class is a Feller-type noise, not a language-model residual. It is used when a deterministic ODE is a lie.

Spoken, the noise widens a band. R0 still comes from the deterministic skeleton. The noise is a nowcast sleeve, not a new radius.

Misread, adding noise and then reporting a tighter band is more honest. It is decoration.

Refused: using the noise to fatten a band on a large national series that does not need it, and quoting the noise as a new R0.

The deterministic SEIRD remains the skeleton for the next-generation inverse.

157. Site knapsack if allocation is done

Live trial-site allocation, when done at all, is a binary program: utilities from inclusion criteria and historical enrollment, costs, a budget, and a minimum site count.

Spoken, greedy by utility over cost is a bound. The LP relaxation gap is recorded. An agent that optimizes a trial on a slide is not this program.

Misread, reinforcement learning that proposes sites is live allocation. It is a research sleeve. It cannot move a site live.

Refused: hiding the LP gap in an attention plot, and a live move without a human-owned budget.

The integer program can move a site when a human still owns the budget. The sleeve cannot.