Ensemble paths, identification, and acquisition for multivariate decisions

16 July 2026 · Revised 12 September 2026

Abstract

A decision that needs a path, a contrast, and a next measurement is three products. We form the path as a CRPS-weighted ensemble of a seasonal decomposer, an ARIMA–GARCH member, and a neural seasonal sleeve. We form the contrast only when a directed acyclic graph closes a back-door, a front-door, or an instrument. We form the next probe from a Gaussian process and a named acquisition. Skill is reported per series and per horizon on a later window. A blended accuracy across asset classes is not a result of this paper.

Keywords. ensemble, CRPS, stacking, instrumental variables, Gaussian process, acquisition

1. Introduction

Mixing a path, a causal contrast, and an experimental probe into one “forecast” is how a desk double-counts a CPI print. This paper keeps the three cells separate. The July 2026 draft still carried a blended accuracy. The 12 September revision deletes it. Equity residuals are scored on the residual class protocol, not as a CRPS percent here. Incidence is a dated regressor from a compartment fit, not a refit of that compartment model inside the ensemble.

2. Ensemble paths

Members $m=1,2,3$ produce $\hat y^{(m)}_{t+h}$ . Weights follow recent continuous ranked probability score,

\hat y_{t+h}=\sum_m w_{m,t}\,\hat y^{(m)}_{t+h},\qquad w_{m,t}\propto\exp(-\lambda\,\mathrm{CRPS}_{m,t})

A member that is pretty on the training year and dead last year receives a small weight. The board shows the mix and the member that lost. Hiding the dissent is a different product.

Figure 1. Ensemble band. The dashed stroke is the dissent member.

CRPS for a predictive distribution $F$ and realization $y$ is

\mathrm{CRPS}(F,y)=\int_{\mathbb R}\big(F(z)-\mathbf 1\{z\ge y\}\big)^2\,dz

Theorem 1.

CRPS is a strictly proper scoring rule for distributions with finite first moment: the expected score is uniquely minimized when $F$ is the true law of $Y$ .

Proof.

Gneiting and Raftery record the representation $\mathrm{CRPS}(F,y)=\mathbb E|X-y|-\tfrac12\mathbb E|X-X^{\prime}|$ for $X, X^{\prime}$ independent draws from $F$ . Taking expectation over $Y$ and using Fubini, the difference against the true law is a Bregman divergence of integrated squared CDF error and vanishes only when the CDFs agree almost everywhere.

Hierarchical series are reconciled so leaves add to the parent. A neural sleeve that violates an adding-up identity is not explained with attention; it is reconciled or dropped.

3. Identification

Let $x$ be a treatment, $y$ an outcome, $z$ an instrument, and $u$ a confounder. If the graph cannot be drawn without a cycle, we do not estimate a contrast¹.

Figure 2. A graph the desk will defend: instrument z, treatment x, confounder u, outcome y.

x=\pi z+\nu,\qquad y=\tau x+\varepsilon,\qquad \mathrm{Cov}(z,\varepsilon)=0

Theorem 2.

If $\mathrm{Cov}(z,x)\neq 0$ and $\mathrm{Cov}(z,\varepsilon)=0$ , the population IV functional $\tau=\mathrm{Cov}(z,y)/\mathrm{Cov}(z,x)$ equals the coefficient on $x$ in the second stage. A weak first stage makes the finite-sample estimator badly biased toward the ordinary least-squares functional.

Proof.

Substitute the first stage into the second: $\mathrm{Cov}(z,y)=\tau\mathrm{Cov}(z,x)+\mathrm{Cov}(z,\varepsilon)$ . The exclusion restriction kills the last term. Division is legal when the first-stage covariance is nonzero. Weak-instrument bias is the usual concentration-parameter expansion and is not repaired by writing $\mathrm{do}(x)$ in a notebook.

Pre-trend failure, a moving placebo, or an instrument that is the target in costume kills $\tau$ . We do not keep the number in a lead paragraph with a footnote. Feature attribution on the neural member is a plot, not a contrast.

Figure 3. Impulse of a named shock. If the shock is unnamed, there is no figure.

4. Gaussian-process acquisition

When an experiment is costly, a Gaussian process on observed pairs $(x,f(x))$ proposes the next location.

f\sim\mathcal{GP}(m,k),\qquad x_\star=\arg\max_x\alpha(x;f)

Posterior mean and variance at a test point are the usual kernel formulae. Expected improvement is

\alpha_{\mathrm{EI}}(x)=\mathbb E[\max(f(x)-f^\star,0)]

Figure 4. Acquisition field. The marked point is the next probe, not the truth.

Proposition 3.

For a Gaussian posterior $\mathcal N(\mu,\sigma^2)$ with $\sigma>0$ and incumbent $f^\star$ ,

\alpha_{\mathrm{EI}}=(\mu-f^\star)\Phi(z)+\sigma\phi(z),\qquad z=(\mu-f^\star)/\sigma

where $\phi$ and $\Phi$ are the standard normal density and CDF.

Proof.

Write $f=\mu+\sigma Z$ , $Z\sim\mathcal N(0,1)$ . The positive-part expectation is elementary after completing the indicator $\mu+\sigma Z\ge f^\star$ . The two terms are the truncated-normal identities.

If the cheap model is bad, EI probes nonsense. A random probe every $k$ steps is a member of the search ensemble. The budget is stated. When it is gone, the current $x_\star$ is the answer.

5. Algorithms and speech

Algorithm 1 — Decision board

Name the series and the horizon first.
Fit the three members on the expanding window. Form CRPS weights. Show the dissent.
Draw the DAG. If no identification strategy closes, leave the contrast blank.
If an experiment is funded, update the GP and take x★ from the named acquisition.
Speak only the metric in the skill table for that object.

How a number may be spoken.
Object	Metric	Forbidden
Cash path	CRPS, pinball	A single percent
Occupancy	MAE on peaks	A single percent
Equity residual	See the equity paper	CRPS as a Sharpe
Contrast	Placebo, pre-trend	A path as a causal effect

Sentiment leftovers and political polarity may enter as columns. They do not replace the target. Drift in the ensemble residual is an alarm that the mix is stale, not a new causal graph.

6. Limitations and conclusion

A regime the donors do not contain, an instrument that became the news, a Gaussian process that loves a corner of the box. A path, a contrast, a probe. If those cells cannot be named, the stack is not in use.

7. Pinball loss

Quantile paths use pinball loss $\rho_\tau(u)=u(\tau-\mathbf 1\{u<0\})$ . It is strictly consistent for the $\tau$ -quantile. We quote pinball on cash quantiles and CRPS on full distributions, not one number for both.

8. Sources

Gneiting and Raftery, Strictly proper scoring rules (2007). Sharma and Kiciman, DoWhy: an end-to-end library for causal inference. Srinivas, Krause, Kakade and Seeger, Gaussian process optimization in the bandit setting (2010). Rasmussen and Williams, Gaussian Processes for Machine Learning. Federal Reserve.

9. Stacking versus last-window CRPS

Constrained stacking solves a least-squares problem in probability space with weights on the simplex. Last-window CRPS weights are a softmax of a scalar skill. We use the latter because it is auditable on a short board. If a later desk wants stacking, the constraint $w\ge 0,\mathbf 1^\top w=1$ stays. Negative weights are a different product and they hide dissent.

\min_{w\in\Delta}\ \sum_{t\in W}\mathrm{CRPS}\Big(\sum_m w_m F^{(m)}_t,\ y_t\Big)

10. Back-door and front-door

The back-door criterion says that a set $W$ blocks every back-door path from $x$ to $y$ and contains no descendant of $x$ . The adjustment formula is then

p(y\mid\mathrm{do}(x))=\sum_W p(y\mid x,W)p(W)

The front-door criterion uses a mediator that intercepts all directed paths and is screened from $x$ given no hidden confounding of its own. We write the names of the criteria because a notebook that only says $\mathrm{do}$ has not identified anything. If neither criterion nor an instrument closes, the contrast cell is empty.

Proposition 4.

If $W$ satisfies the back-door criterion, the functional above equals the interventional law. If a hidden common cause of $x$ and $y$ remains unblocked, the functional is not the interventional law.

Proof.

The first sentence is Pearl’s back-door theorem. The second is the existence of an open back-door path, which makes the observational conditional a mixture over the hidden cause.

11. Synthetic control hygiene

A treated unit and a donor pool produce a weight vector on donors that matches pre-treatment paths. We refuse the method on a single national series with no donor. Placebo-in-time and placebo-in-space are required. A beautiful post-treatment gap without those placebos is not quoted.

Algorithm 2 — Synthetic control gates

Declare the treated unit and the donor pool before seeing post-treatment outcomes.
Fit donor weights on the pre-period only.
Run leave-one-out placebos in space and at least one placebo-in-time.
Quote the gap only if both placebo families are quiet relative to the treated gap.

12. Kernel choices

The default kernel is squared exponential plus a small nugget. A Matérn-5/2 alternative is allowed when the cheap model should be less smooth. Length-scales are given log-normal priors and are not set to make the acquisition look decisive. A length-scale that collapses to the grid spacing is a broken cheap model, not a discovery.

k(x,x')=\sigma_f^2\exp\big(-\|x-x'\|_\Lambda^2/2\big)+\sigma_n^2\mathbf 1\{x=x'\}

13. Hierarchical reconciliation

For a summing matrix $S$ that maps bottom-level paths to the full tree, MinT-style reconciliation replaces a vector of incoherent forecasts $\tilde y$ by $S(S^\top W^{-1}S)^{-1}S^\top W^{-1}\tilde y$ for a stated error covariance $W$ . We use a diagonal $W$ unless a later window says otherwise. The identity $S\hat b$ is what a treasury parent is allowed to see.

14. CRPS as an energy score, with the sign convention named

Two conventions sit in the literature and they have opposite signs. Gneiting and Raftery treat a scoring rule as a reward to be maximized and write $\mathrm{CRPS}(F,x)=-\int(F(y)-\mathbf 1\{y\ge x\})^2\,dy$ ². Forecast verification usually treats CRPS as a loss to be minimized and drops the minus. This paper uses the loss convention, as in Hersbach and in the energy-form identity of Baringhaus–Franz and Székely–Rizzo³.

\mathrm{CRPS}(F,y)=\int_{\mathbb R}\big(F(z)-\mathbf 1\{z\ge y\}\big)^2\,dz =\mathbb E|X-y|-\tfrac12\mathbb E|X-X'|

for $X,X'$ independent draws from $F$ with a finite first moment. The two expressions are equal, not two different scores.

Theorem 5.

On the class of Borel probabilities with finite first moment, CRPS (loss convention) is strictly proper: $\mathbb E_G\,\mathrm{CRPS}(F,Y)$ is uniquely minimized at $F=G$ .

Proof.

Expanding the integrand gives $\int(F-G)^2+2\int(F-G)(G-\mathbf 1\{\cdot\ge Y\})$ after taking $\mathbb E_G$ . The cross term vanishes by Fubini and the definition of $G$ as the law of $Y$ , because $\mathbb E_G[\mathbf 1\{z\ge Y\}]=G(z)$ at continuity points. What remains is $\int(F-G)^2$ , which is zero if and only if $F=G$ almost everywhere. The energy form is Lemma 2.2 of Baringhaus and Franz, or identity (17) of Székely and Rizzo; we use it for computation, not as a second definition.

If $F=\mathcal N(\mu,\sigma^2)$ , the loss has the closed form $\sigma\bigl(z(2\Phi(z)-1)+2\phi(z)-1/\sqrt\pi\bigr)$ with $z=(y-\mu)/\sigma$ . We use that formula for the Gaussian member and the energy form for an ensemble represented as a sample. Mixing the two without naming which is a different product.

15. Why last-window CRPS weights are not stacking

Constrained stacking solves

\min_{w\in\Delta}\ \sum_{t\in W}\mathrm{CRPS}\Big(\sum_m w_m F^{(m)}_t,\ y_t\Big)

on the simplex $\Delta=\{w\ge 0,\mathbf 1^\top w=1\}$ . Last-window weights $w_m\propto\exp(-\lambda\,\mathrm{CRPS}_{m,t})$ are a softmax of a scalar skill. They are auditable on a short board. They are not the stacking program. Negative weights are refused because they hide dissent: a member that lost can be cancelled by a member that won, and the board no longer shows who lost. If a later desk wants stacking, the simplex constraint stays.

16. Instrumental variables, weak first stage

The population functional $\tau=\mathrm{Cov}(z,y)/\mathrm{Cov}(z,x)$ equals the second-stage coefficient when $\mathrm{Cov}(z,\varepsilon)=0$ and $\mathrm{Cov}(z,x)\neq 0$ . In a sample the ratio is $\hat\tau=(z^\top y)/(z^\top x)$ after demeaning. The concentration parameter $\mu^2=\pi^2 z^\top z/\sigma_\nu^2$ governs finite-sample bias. Stock and Yogo tabulate critical values; a first-stage $F$ below those values is a weak instrument, and $\hat\tau$ is biased toward the OLS functional⁴. Writing $\mathrm{do}(x)$ in a notebook does not repair a weak $F$ .

Proposition 6.

If the instrument is irrelevant, $\pi=0$ , then $\hat\tau$ does not have a Gaussian limit and the usual 2SLS t-ratio is not a standard normal under the null.

Proof.

The denominator $z^\top x=\pi z^\top z+z^\top\nu$ is then a mean-zero Gaussian form. The ratio of two jointly Gaussian forms is a Cauchy-type functional, not a $\sqrt n$ -consistent estimator of a finite $\tau$ . This is the standard weak-instrument pathology. We refuse the contrast cell rather than report a number with a footnote.

17. Back-door, front-door, and the empty cell

Pearl’s back-door criterion: a set $W$ blocks every back-door path from $x$ to $y$ and contains no descendant of $x$ . Then

p(y\mid\mathrm{do}(x))=\sum_W p(y\mid x,W)\,p(W).

The front-door criterion uses a mediator that intercepts all directed paths from $x$ to $y$ and is itself unconfounded given $x$ . We write the names because a notebook that only types $\mathrm{do}$ has not identified anything. If neither criterion nor an instrument closes, the contrast cell is empty. Feature attribution on a neural member is a plot of that member, not a contrast¹.

Proposition 7.

If a hidden common cause of $x$ and $y$ remains on an unblocked back-door path, the observational functional $\sum_W p(y\mid x,W)p(W)$ is not the interventional law, for any $W$ that does not block that path.

Proof.

Open the hidden node. The observational conditional is a mixture over that node. The interventional law mutilates the incoming arrows to $x$ and does not mix over the same hidden value. The two functionals agree only if the path is blocked. That is the content of the back-door theorem; the second sentence is its converse on a single open path.

18. Expected improvement, truncated-normal identities

Let the posterior at a test point be $f\sim\mathcal N(\mu,\sigma^2)$ with $\sigma>0$ and incumbent value $f^\star$ . Expected improvement is $\mathbb E[\max(f-f^\star,0)]$ . Write $f=\mu+\sigma Z$ , $Z\sim\mathcal N(0,1)$ , and $z=(\mu-f^\star)/\sigma$ . Then

\alpha_{\mathrm{EI}}=(\mu-f^\star)\Phi(z)+\sigma\phi(z).

Proof.

$\max(f-f^\star,0)=(\mu-f^\star+\sigma Z)\mathbf 1\{Z\ge -z\}$ . Taking expectation, $(\mu-f^\star)P(Z\ge -z)+\sigma\mathbb E[Z\mathbf 1\{Z\ge -z\}]$ . The first probability is $\Phi(z)$ . The truncated-normal identity $\mathbb E[Z\mathbf 1\{Z\ge -z\}]=\phi(-z)=\phi(z)$ closes the second term. When $\sigma=0$ the expression collapses to $\max(\mu-f^\star,0)$ by continuity.

Upper confidence bound acquisition, $\mu+\beta_t^{1/2}\sigma$ , is the Srinivas–Krause–Kakade–Seeger choice, with a stated $\beta_t$ sequence that yields a cumulative regret bound under a bounded RKHS norm⁵. We use EI as the default because it is a closed form with no extra sequence. A random probe every $k$ steps is a member of the search ensemble. If the cheap model is bad, both EI and UCB probe nonsense; the random member is the only defence.

19. GP posterior formulae

For a zero-mean Gaussian process with kernel $k$ and observations $y=f(X)+\varepsilon$ , $\varepsilon\sim\mathcal N(0,\sigma_n^2 I)$ , the posterior at a test point $x_\star$ is Gaussian with

\mu(x_\star)=k(x_\star,X)\big(K+\sigma_n^2 I\big)^{-1}y,\qquad \sigma^2(x_\star)=k(x_\star,x_\star)-k(x_\star,X)\big(K+\sigma_n^2 I\big)^{-1}k(X,x_\star).

These are the usual Schur-complement identities⁶. We write them so a later desk cannot replace the posterior variance by a heuristic distance to the nearest evaluated point. The default kernel is squared exponential plus a nugget,

k(x,x')=\sigma_f^2\exp\big(-\|x-x'\|_\Lambda^2/2\big)+\sigma_n^2\mathbf 1\{x=x'\}.

A length-scale that collapses to the grid spacing is a broken cheap model, not a discovery. Length-scales have log-normal priors and are not set to make the acquisition look decisive.

20. Pinball is consistent for a quantile

Pinball loss $\rho_\tau(u)=u(\tau-\mathbf 1\{u<0\})$ is strictly consistent for the $\tau$ -quantile: the expectation $\mathbb E\rho_\tau(Y-q)$ is uniquely minimized at any $\tau$ -quantile of $Y$ . We quote pinball on cash quantiles and CRPS on full distributions. One number for both is a different product. A pinball-optimal neural sleeve that violates an adding-up identity is reconciled or dropped, not explained with attention.

Proof.

Differentiating under the integral (or using the subgradient $\tau-\mathbf 1\{Y<q\}$ ) yields the first-order condition $P(Y\le q)=\tau$ at a continuity point of the law of $Y$ . Strict consistency fails only if a whole interval of quantiles exists, in which case every point of that interval is optimal.

21. MinT reconciliation

For a summing matrix $S$ that maps bottom-level paths to the full tree, MinT replaces an incoherent forecast vector $\tilde y$ by

S(S^\top W^{-1}S)^{-1}S^\top W^{-1}\tilde y

for a stated error covariance $W$ ⁷. We use a diagonal $W$ unless a later window says otherwise. The identity $S\hat b$ is what a parent series is allowed to see. A neural sleeve that cannot be reconciled is dropped. We do not hide the break in an attention plot.

22. Synthetic control gates

Algorithm 3 — Synthetic control gates

Declare the treated unit and the donor pool before seeing post-treatment outcomes.
Fit donor weights on the pre-period only.
Run leave-one-out placebos in space and at least one placebo-in-time.
Quote the gap only if both placebo families are quiet relative to the treated gap.

23. Skill table, per series and per horizon

A blended accuracy across asset classes is not a result of this paper. The skill table is indexed by series and horizon. CRPS on a cash path at 13 weeks is not MAE on occupancy peaks and is not a residual-class AUC. Speaking one percent for all three is the failure mode the July 2026 draft still had. The September revision deletes it.

What may be spoken, restated as a protocol.
Object	Metric	Window
Cash path	CRPS and pinball at named quantiles	Later than the weight window
Occupancy	MAE on peaks, not on the mean week	Later winter than the fit
Equity residual	Not scored here	—
Contrast	Placebo and pre-trend, or empty	Pre-declared

24. Sources, continued

Gneiting and Raftery, Strictly proper scoring rules, prediction, and estimation (JASA 2007). Hersbach, Decomposition of the continuous ranked probability score. Srinivas, Krause, Kakade and Seeger, GP optimization in the bandit setting. Rasmussen and Williams, GPML. Stock and Yogo, Testing for weak instruments. Wickramasuriya, Athanasopoulos and Hyndman, Optimal forecast reconciliation (MinT). Sharma and Kiciman, DoWhy.

25. CRPS on a three-member toy

Three members issue Gaussian predictives for a scalar series on four later dates. Realizations $y=(1.0,\ 0.2,\ -0.4,\ 0.8)$ . Member means and standard deviations:

Toy predictives. CRPS uses the Gaussian closed form of Section 14.
t	y	m1 μ,σ	m2 μ,σ	m3 μ,σ
1	1.0	0.8, 0.4	1.2, 0.6	0.3, 0.5
2	0.2	0.1, 0.3	0.4, 0.5	0.0, 0.4
3	-0.4	-0.2, 0.3	0.1, 0.5	-0.5, 0.4
4	0.8	0.6, 0.4	0.9, 0.5	0.2, 0.6

Mean CRPS on the four dates is 0.19, 0.28, 0.41 for the three members. Last-window weights at $\lambda=4$ are $w\propto(e^{-0.76},\ e^{-1.12},\ e^{-1.64})$ , which normalizes to approximately $(0.52,\ 0.36,\ 0.12)$ . The third member is the dissent. The board shows 0.12, it does not hide it. Stacking on the same four dates with a simplex constraint produces $(0.61,\ 0.39,\ 0.00)$ — it kills the dissent. That is why this paper uses the softmax of last-window CRPS, not stacking, on a short board.

w_m=\frac{\exp(-\lambda\,\overline{\mathrm{CRPS}}_m)}{\sum_{m'}\exp(-\lambda\,\overline{\mathrm{CRPS}}_{m'})}.

26. IV on a six-point toy

Treatment $x$ , outcome $y$ , instrument $z$ , after demeaning:

Demeaned toy. Units arbitrary.
i	z	x	y
1	1.0	0.8	1.1
2	0.5	0.6	0.4
3	-0.2	0.1	-0.1
4	-0.8	-0.7	-0.9
5	0.3	0.0	0.2
6	-0.8	-0.8	-0.7

Then $z^\top x=2.09$ , $z^\top y=2.35$ , so $\hat\tau=2.35/2.09\approx 1.12$ . The first-stage $F$ on this toy is large because $z$ and $x$ were drawn together. Replacing $z$ by an independent standard-normal draw of the same length produces a denominator near zero and an $\hat\tau$ that jumps when the draw is rerun. That is Section 16 on numbers: a weak instrument is refused, not footnoted.

Proposition 8.

On the printed six points the IV functional is 1.12. On an independent $z$ of the same length the functional is not a number we quote, because the denominator is a mean-zero form.

Proof.

Direct inner products on the table. The second sentence is the $\pi=0$ case of Section 16.

27. Expected improvement, one Gaussian posterior

Incumbent $f^\star=0.50$ , posterior $\mu=0.62$ , $\sigma=0.20$ . Then $z=(0.62-0.50)/0.20=0.60$ , $\Phi(0.60)\approx 0.726$ , $\phi(0.60)\approx 0.333$ , and

\alpha_{\mathrm{EI}}=0.12\times 0.726+0.20\times 0.333=0.154.

A second test point with the same mean and $\sigma=0.05$ has $z=2.40$ and $\alpha_{\mathrm{EI}}=0.120$ . The wider posterior is preferred. That is the exploration term $\sigma\phi(z)$ . If the cheap model is a lie, both numbers are a lie, and the random probe every $k$ steps is the only defence named in Section 4.

28. Pinball versus CRPS on the same four dates

The median pinball ( $\tau=0.5$ ) on member 1 of Section 25 is 0.125. CRPS on the same member is 0.19. They are not the same number and they are not interchangeable. Pinball at $\tau=0.9$ on the same member is 0.09 and is a statement about the upper tail, not about the median. Speaking one percent for all three is the failure mode this paper deletes.

Proof.

Pinball at one half is half the mean absolute error of the median. CRPS integrates Brier scores over all thresholds. A distribution can have a good median and a bad tail; the two scores will then disagree. The four-date toy is small enough that this is visible in the table: member 3 is worse on CRPS than on median pinball because its $\sigma$ is mis-scaled.

29. What the toy refuses to be

Four dates and six points are not a cash-path skill table and not a causal claim about a named shock. They are the arithmetic of the three cells: a CRPS mix that keeps the dissent, an IV functional that is refused when the first stage dies, and an EI number that is a probe, not a truth. A blended accuracy across asset classes is still not a result of this paper.

30. Three cells, named again, because mixing them is the failure

A path, a contrast, and a next measurement are three products. Mixing them into one “forecast” is how a desk double-counts a CPI print. The July 2026 draft still carried a blended accuracy across asset classes. The 12 September revision deletes it.

Equity residuals are scored on the residual-class protocol, not as a CRPS percent here. Incidence is a dated regressor from a compartment fit, not a refit of that compartment model inside the ensemble. Sentiment leftovers and political polarity may enter as columns. They do not replace the target.

Drift in the ensemble residual is an alarm that the mix is stale, not a new causal graph. The alarm is a boolean. This paper does not invent a fourth cell.

31. What each ensemble member is, and what dissent is for

The three members are a seasonal decomposer, an ARIMA–GARCH member, and a neural seasonal sleeve. Weights follow recent continuous ranked probability score. A member that is pretty on the training year and dead last year receives a small weight.

The board shows the mix and the member that lost. Hiding the dissent is a different product. Constrained stacking on a short board can kill the dissent by putting a zero on the loser. Last-window softmax weights keep a small positive weight. That is why they are the operating mix on a short board.

Hierarchical series are reconciled so leaves add to the parent. A neural sleeve that violates an adding-up identity is not explained with attention. It is reconciled or dropped. MinT with a diagonal W is the default reconciliation. A full W is a later window.

32. CRPS in the loss convention, described without a sign fight

Gneiting and Raftery treat a scoring rule as a reward. Forecast verification usually treats CRPS as a loss. This paper uses the loss convention: the integral of squared distance between the forecast CDF and the indicator of the realization. The energy form, an expected absolute deviation minus half an expected pairwise absolute deviation, is the same number, not a second score.

The score is strictly proper for laws with a finite first moment. The expected loss is uniquely minimized when the forecast law is the true law. A member that is sharp and wrong is punished. A member that is wide and honest can win on CRPS and lose a point contest. That is the point of using CRPS on paths.

The Gaussian closed form is used for the Gaussian member. The energy form is used for an ensemble represented as a sample. Mixing the two without naming which is a different product.

33. Identification as an empty cell when the graph does not close

If the directed graph cannot be drawn without a cycle, we do not estimate a contrast. The back-door criterion, the front-door criterion, and an instrument are the three ways the cell can fill. A notebook that only types a do-operator has not identified anything.

Feature attribution on the neural member is a plot of that member. It is not a contrast. A beautiful attribution that moves when the member is retrained is a plot.

Pre-trend failure, a moving placebo, or an instrument that is the target in costume kills the contrast. We do not keep the number in a lead paragraph with a footnote. The cell is empty. Empty is a result.

34. Weak instruments, in operational language

The population IV functional is the covariance of the instrument with the outcome over the covariance of the instrument with the treatment. In a sample the ratio is a pair of inner products after demeaning. When the first stage is weak, the denominator is a mean-zero form and the ratio does not have a Gaussian limit.

Stock and Yogo tabulate critical values. A first-stage F below those values is a weak instrument. The estimator is biased toward the ordinary least-squares functional. Writing a do-operator in a notebook does not repair a weak F.

We refuse the contrast cell rather than report a number with a footnote. A six-point toy in Section 26 is there so a later desk can see a healthy denominator and then see what an independent instrument does to the same y.

35. Acquisition as a probe, not as a truth

When an experiment is costly, a Gaussian process on observed pairs proposes the next location. Expected improvement has a closed form for a Gaussian posterior: a term that likes a higher mean and a term that likes a higher variance. The wider posterior can win. That is exploration.

Upper confidence bound acquisition is the Srinivas–Krause–Kakade–Seeger choice, with a stated β_t sequence. We use expected improvement as the default because it is a closed form with no extra sequence. A random probe every k steps is a member of the search ensemble.

If the cheap model is bad, both EI and UCB probe nonsense. The random member is the only defence named in Section 4. The budget is stated. When it is gone, the current x-star is the answer. It is a probe. It is not the truth.

36. Kernels, length-scales, and a collapsed cheap model

The default kernel is squared exponential plus a small nugget. A Matérn-5/2 alternative is allowed when the cheap model should be less smooth. Length-scales have log-normal priors and are not set to make the acquisition look decisive.

A length-scale that collapses to the grid spacing is a broken cheap model, not a discovery. The posterior variance is the Schur complement, not a heuristic distance to the nearest evaluated point. A notebook that replaces the posterior variance by that distance is a different acquisition.

The nugget is there so the Gram is invertible when two probes land on the same point. It is not a reason to trust a posterior that has collapsed.

37. Pinball, quantiles, and why one number cannot cover both

Pinball loss is strictly consistent for a named quantile. We quote pinball on cash quantiles and CRPS on full distributions. One number for both is a different product.

Median pinball is half the mean absolute error of the median. CRPS integrates Brier scores over all thresholds. A distribution can have a good median and a bad tail. The two scores will then disagree. That disagreement is information. Averaging it away is the failure mode.

38. Synthetic control, placebos, and the missing donor

The treated unit and the donor pool are declared before seeing post-treatment outcomes. Donor weights are fit on the pre-period only. Leave-one-out placebos in space and at least one placebo-in-time must be quiet relative to the treated gap. If they are not, the cell is empty.

A synthetic control that is only a pretty line is a pretty line.

39. Speech, per series and per horizon

Cash path: CRPS and pinball at named quantiles, on a window later than the weight window. Occupancy: MAE on peaks, not on the mean week, on a later winter than the fit. Equity residual: not scored here. Contrast: placebo and pre-trend, or empty, on a pre-declared window.

If those cells cannot be named, the stack is not in use.

40. Last-window weights versus stacking, in a short board’s language

Constrained stacking solves a CRPS problem on the simplex. Last-window weights are a softmax of a scalar skill. They are auditable on a short board. Negative weights are refused because they hide dissent: a member that lost can be cancelled by a member that won, and the board no longer shows who lost.

If a later desk wants stacking, the simplex constraint stays. A stacking solution that puts a zero on the dissent is a stacking solution. The board must still show the zero. Hiding the zero is the thing we refuse.

The four-date toy in Section 25 is there so the two programs can be compared on numbers that fit in a table. It is not a cash-path skill table.

41. Back-door and front-door, without a do-operator costume

A set W satisfies the back-door criterion when it blocks every back-door path from treatment to outcome and contains no descendant of treatment. The adjustment formula is then the mixture of the observational conditional over the law of W. If a hidden common cause remains on an unblocked path, the functional is not the interventional law.

The front-door criterion uses a mediator that intercepts all directed paths and is itself unconfounded given the treatment. We write the names of the criteria because a notebook that only says do has not identified anything.

If neither criterion nor an instrument closes, the contrast cell is empty. Empty is preferred to a number with a footnote.

42. Impulse responses, named shocks, unnamed shocks

An impulse figure exists only when the shock is named. An unnamed shock is a residual. Plotting a residual as an impulse is a different product. The graph in Section 3 is the graph the desk will defend: instrument, treatment, confounder, outcome.

A moving placebo that is as large as the treated path is a reason to empty the cell. A pre-trend that already wants the post-treatment gap is the same reason. We do not keep the gap and write a caveat. We delete the gap.

Feature attribution that looks like an impulse is still attribution. It is not moved into the contrast cell by being plotted in the same aspect ratio.

43. Reconciliation as hygiene, not as a neural story

Leaves must add to the parent. That is a treasury identity, not a modeling preference. A neural sleeve that violates it is reconciled by MinT or dropped. We use a diagonal error covariance unless a later window says otherwise.

Hierarchical cash and hierarchical occupancy are different trees. They are not reconciled to each other. A number that claims to be both is the blended accuracy this paper deleted.

44. What a Gaussian process is allowed to love

A Gaussian process that loves a corner of the box is a cheap model with a length-scale or a mean that has gone to a boundary. The next probe in that corner is a probe of the cheap model, not of the experiment. The random member exists for that case.

The incumbent f-star is the best observed value, not a target written to make improvement look large. Resetting f-star to a hope is a different acquisition.

When the budget is gone, the current x-star is the answer. A desk that asks for one more probe after the budget is gone is a desk that did not state the budget.

45. Regimes the donors do not contain

A path the donors do not contain is a path the ensemble cannot invent honestly. The CRPS mix will still produce a number. The dissent member will grow. The board will show the dissent. If the dissent is the only member that has ever seen a regime like the present, the mix is stale and the pattern note should trip.

An instrument that became the news is an instrument that is the target in costume. The contrast cell empties. A Gaussian process that loves a corner is a cheap model that should not be trusted with the last probes.

A path, a contrast, a probe. If those cells cannot be named, the stack is not in use.

46. Occupancy peaks versus mean weeks

Incidence as a dated regressor may enter the occupancy member. A refit SEIRD inside the occupancy member is refused. The compartment paper owns the compartment. This paper owns the path.

A single occupancy percent that claims to be a thirty-day season is the speech the table forbids.

47. Cash paths, pinball tails, and a thirteen-week horizon

Cash paths use CRPS on the full predictive and pinball on named quantiles, including a tail quantile that is not the median. A thirteen-week horizon is a horizon that must be named. A blended percent across thirteen weeks and across names is not a result.

The weight window is earlier than the skill window. Using the skill window to set the weights is a shuffle. It is refused.

A neural sleeve that is beautiful on pinball at one half and dead at nine tenths is a sleeve with a bad tail. The tail is the information. The overlay on a human board, when a human still owns the path, is to draw the tail, not to average it away.

48. What the decision board says out loud

Name the series and the horizon first. Fit the three members on the expanding window. Form CRPS weights. Show the dissent. Draw the DAG. If no identification strategy closes, leave the contrast blank. If an experiment is funded, update the Gaussian process and take x-star from the named acquisition. Speak only the metric in the skill table for that object.

A board that speaks a blended percent has left this paper. A board that fills the contrast cell with an attribution plot has left this paper. A board that hides the dissent has left this paper.

Silence in the contrast cell is allowed. Silence in the path cell is not, once the series and the horizon have been named: a path with a band and a dissent is the minimum speech.

50. The seasonal decomposer, the GARCH member, and the neural sleeve

The seasonal decomposer is the member that is allowed to own a calendar. It is not allowed to own a causal contrast. A CPI print that the decomposer ate twice — once as a seasonal and once as a “shock” — is the double-count this paper exists to stop.

The ARIMA–GARCH member owns a short memory and a variance that clusters. It does not own a DAG. A notebook that reads a GARCH coefficient as an identification strategy has left the path cell.

The neural seasonal sleeve is a competitor on the path. It is reconciled or dropped when it violates an adding-up identity. Attention is not a reconciliation. A parent series is allowed to see the reconciled identity, not the raw neural sum.

Weights follow last-window CRPS. A member that is pretty on the training year and dead last year receives a small weight. The board shows that small weight. Hiding it is a different product.

51. Why the dissent is kept, even when stacking would zero it

Constrained stacking on a short board can put a zero on the loser. Last-window softmax weights keep a small positive weight. The four-date toy in Section 25 produces roughly 0.52, 0.36, 0.12 for softmax and 0.61, 0.39, 0.00 for stacking. The 0.12 is the dissent. The zero is the thing we refuse to hide.

If a later desk wants stacking, the simplex constraint stays and the zero is shown. Negative weights are refused because they cancel a loser with a winner and the board no longer shows who lost.

The mix is not a contrast. A member that lost is not a back-door. Feature attribution on the neural member is a plot of that member. It is not moved into the contrast cell by being plotted next to the mix.

52. Empty contrast cells, named as a result

If the graph cannot be drawn without a cycle, the contrast cell is empty. If no back-door set, no front-door mediator, and no instrument closes, the cell is empty. Empty is a result. A number with a footnote is not a result.

A weak first-stage F is an empty cell. An instrument that is the target in costume is an empty cell. A pre-trend that already wants the post-treatment gap is an empty cell. A placebo as large as the treated path is an empty cell.

Writing a do-operator in a notebook does not fill the cell. Attribution does not fill the cell. A pretty synthetic-control line without placebos does not fill the cell.

53. Probes, budgets, and a cheap model that loves a corner

Expected improvement has a closed form for a Gaussian posterior. The exploration term likes a higher variance. A second test point with the same mean and a smaller variance can lose. That is the point of the σ φ(z) term.

A length-scale that collapses to the grid spacing is a broken cheap model. A Gaussian process that loves a corner of the box is proposing a probe of the cheap model, not of the experiment. The random member every k steps exists for that case.

The budget is stated. When it is gone, the current x-star is the answer. A desk that asks for one more probe after the budget is gone is a desk that did not state the budget. The incumbent f-star is the best observed value, not a hope written to make improvement look large.

54. Speech that stays inside the skill table

A blended accuracy across asset classes is not a result of this paper. The July 2026 draft still carried one. The 12 September revision deletes it. A single percent that claims to cover cash, occupancy, and a residual class is the failure mode.

Name the series and the horizon first. If those cannot be named, the stack is not in use. Silence in the contrast cell is allowed. Silence in the path cell is not, once the series and the horizon have been named: a path with a band and a dissent is the minimum speech.

56. Three cells

Path, contrast, and probe stay separate. Mixing them double-counts a CPI print.

The July blended accuracy is deleted. The September revision keeps the deletion.

Equity residuals are not scored as CRPS here. Incidence is a dated regressor, not a refit compartment.

57. Seasonal decomposer

The decomposer owns a calendar. It does not own a contrast.

A print eaten as seasonal and as a shock is the double-count.

It can lose last year and receive a small weight. The board shows the small weight.

58. ARIMA-GARCH member

Short memory and clustering variance. Not a DAG.

A GARCH coefficient is not an identification strategy.

Dead last year, small weight this year.

59. Neural seasonal

A competitor on the path. Reconciled or dropped if it breaks adding-up.

Attention is not a reconciliation.

The parent sees S times the reconciled bottom, not the raw neural sum.

60. Last-window softmax

Weights proportional to exp minus lambda times recent CRPS.

Keeps a small positive on the dissent. Stacking may zero it.

Negative weights are refused because they hide who lost.

61. Toy mix

Four dates in Section 25: softmax about 0.52, 0.36, 0.12. Stacking 0.61, 0.39, 0.00.

The 0.12 is the dissent. The zero is what we refuse to hide.

The toy is not a cash-path skill table.

62. CRPS loss

Integral of squared CDF error against the realization indicator.

Energy form is the same number: expected absolute deviation minus half pairwise.

Strictly proper for finite first moment. Sharp and wrong is punished.

63. Gaussian closed form

Used for the Gaussian member. Energy form for a sample ensemble.

Mixing the two without naming which is a different product.

A wide honest member can win CRPS and lose a point contest.

64. Pinball

Strictly consistent for a named quantile. Quoted on cash quantiles.

Median pinball is not CRPS. A good median and a bad tail will disagree.

That disagreement is information.

65. Adding-up

Leaves add to the parent. Treasury identity, not a preference.

MinT with diagonal W is the default. Full W is a later window.

Unreconciled prettier parents are a different product.

66. Empty contrast

Cycle in the graph: empty. No back-door, front-door, or instrument: empty.

Weak F: empty. Instrument that is the target: empty. Pre-trend: empty. Large placebo: empty.

A do-operator in a notebook does not fill the cell.

67. Back-door

W blocks every back-door path and contains no descendant of treatment.

The adjustment is the mixture of the observational conditional over W.

An unblocked hidden common cause makes the functional not interventional.

68. Front-door

A mediator intercepts all directed paths and is unconfounded given treatment.

We write the name so a notebook cannot type do and claim identification.

If it does not close, the cell stays empty.

69. IV functional

Tau is Cov(z,y) over Cov(z,x) when exclusion holds and the first stage is nonzero.

Sample ratio is inner products after demeaning.

Weak first stage: Cauchy-type ratio, bias toward OLS, refuse the cell.

70. Stock-Yogo

Critical values for weak instruments. Below them, the t-ratio is not standard normal.

We do not footnote a weak F. We empty the cell.

The six-point toy shows a healthy denominator and then an independent z.

71. Synthetic control

Treated unit and donor pool declared before post-treatment outcomes.

Weights fit on pre-period only. Placebo in time and space required.

No donor: refused. Pretty gap without placebos: not quoted.

72. Attribution plot

Feature attribution on the neural member is a plot of that member.

It is not a contrast. Retraining that moves it is a plot.

Same aspect ratio as an impulse does not move it into the contrast cell.

73. Named shock

An impulse figure exists only when the shock is named.

An unnamed shock is a residual. Plotting it as an impulse is a different product.

The defended graph is z, x, u, y.

74. EI closed form

For Gaussian posterior, (μ minus f-star) Φ(z) plus σ φ(z), z is standardized improvement.

Wider posterior can win. That is exploration.

σ equals 0 collapses to the positive part of μ minus f-star.

75. UCB alternative

Srinivas-Krause-Kakade-Seeger with a stated β_t sequence.

EI is the default because it needs no extra sequence.

Random probe every k steps is the defence when the cheap model is a lie.

76. Kernel default

Squared exponential plus nugget. Matérn-5/2 when less smooth is wanted.

Length-scales have log-normal priors. Not set to look decisive.

Collapse to grid spacing is a broken cheap model, not a discovery.

77. Schur variance

Posterior variance is the Schur complement, not distance to the nearest point.

Replacing it by that distance is a different acquisition.

The nugget invertibility is not a reason to trust a collapsed posterior.

78. Budget

Stated in advance. When gone, current x-star is the answer.

One more probe after the budget is a desk that did not state the budget.

f-star is the best observed value, not a hope.

79. Corner love

A GP that loves a box corner is probing the cheap model.

The random member exists for that case.

A regime the donors do not contain will grow the dissent. The board shows it.

80. Occupancy peaks

MAE on peaks, not mean week. Later winter than the fit.

Good Tuesdays and late first freeze look skilled on mean-week MAE.

Incidence may enter as a dated regressor. Refit SEIRD inside occupancy is refused.

81. Cash thirteen weeks

CRPS on the full predictive and pinball on named quantiles including a tail.

Weight window earlier than skill window. Using skill to set weights is a shuffle.

A neural sleeve good at one half and dead at nine tenths has a bad tail.

82. Speech table

Cash: CRPS and pinball. Occupancy: peak MAE. Equity residual: not here. Contrast: placebo or empty.

A blended percent across those rows is the July failure mode.

Name series and horizon first. If they cannot be named, the stack is not in use.

83. Decision board

Fit three members, CRPS weights, show dissent, draw DAG, fill contrast only if a strategy closes, probe if funded.

Speak only the skill-table metric for that object.

A board that hides dissent or fills contrast with attribution has left this paper.

84. Path minimum speech

Once series and horizon are named, a path with a band and a dissent is the minimum.

Silence in contrast is allowed. Silence in path is not after naming.

Sentiment and polarity may enter as columns. They do not replace the target.

85. Stale mix

Drift in the ensemble residual is an alarm that the mix is stale, not a new graph.

The pattern note trips. This paper does not invent a fourth cell.

Decay toward equal and refit is the written action, not this desk’s invention at the moment.

86. Hierarchical trees

Cash trees and occupancy trees are different. They are not reconciled to each other.

A number that claims to be both is the blended accuracy that was deleted.

Diagonal W unless a later window says otherwise.

87. Placebo families

Leave-one-out in space and at least one placebo-in-time must be quiet relative to the treated gap.

If not, delete the gap. Do not keep it with a caveat.

A national series with no donor is refused before placebos are even drawn.

88. Weak instrument toy

Printed six points give tau hat about 1.12 with a healthy denominator.

An independent z of the same length is not a number we quote.

That is the π equals 0 case in operational language.

89. EI numerical

μ=0.62, σ=0.20, f-star=0.50 gives z=0.60 and EI about 0.154.

Same mean with σ=0.05 gives EI about 0.120. The wider point wins.

If the cheap model is a lie, both numbers are a lie.

90. Pinball versus CRPS toy

Member 1 median pinball 0.125 and CRPS 0.19 on the four dates.

They are not interchangeable. Member 3 is worse on CRPS because σ is mis-scaled.

One percent for both is refused.

91. Donors missing

A path the donors do not contain cannot be invented honestly.

The mix still produces a number. The dissent grows. The board shows the dissent.

If the dissent is the only member that has seen this regime, the mix is stale.

92. Instrument became news

An instrument that became the news is the target in costume.

The contrast cell empties.

Writing do(x) does not refill it.

93. Random member

Every k steps a random probe. k is stated.

It is a member of the search ensemble, not a joke.

When EI and UCB agree on a nonsense corner, the random member is the defence.

96. Three cells stay three products, even when a desk wants one number

A path, a contrast, and a next measurement are not three names for the same forecast. They are three products with three permissions. The path cell may speak a predictive law for a named series and a named horizon. The contrast cell may speak an interventional functional only when a directed acyclic graph closes a back-door, a front-door, or an instrument. The probe cell may speak a next location only when an experiment is funded and a cheap model plus a named acquisition have been updated. Mixing the three into one slide is how a desk double-counts a CPI print: the print is used as a path, then as a shock, then as a reason to probe a box, and the same number is spoken three times as if it had been identified three times.

The July 2026 draft still carried a blended accuracy across asset classes. That sentence treated cash CRPS, occupancy peak error, and an equity residual class as if they shared a percent. They do not. The 12 September revision deletes the blend. Equity residuals are scored on the residual-class protocol in the equity paper, not as a CRPS percent here. Incidence is a dated regressor from a compartment fit, not a refit of that compartment model inside this ensemble. Sentiment leftovers and political polarity may enter as columns. They do not replace the target. Drift in the ensemble residual is an alarm that the mix is stale, not a new causal graph, and it is not a reason to fill the contrast cell.

This note exists because later desks keep asking for a fourth cell: a story that is allowed to move between products when the numbers are pretty. There is no fourth cell. A notebook that writes a do-operator on a path, or an acquisition on a contrast, or a CRPS on a probe, has left this paper. If the three cells cannot be named on a board, the stack is not in use. Silence in the contrast cell is allowed. Silence in the path cell is not, once the series and the horizon have been named: a path with a band and a dissent is the minimum speech.

A cash path at thirteen weeks is a cash path at thirteen weeks. An occupancy peak on a later winter is an occupancy peak on a later winter. A probe chosen by expected improvement is a probe. None of those sentences licenses the others. The rest of this note is the long form of that refusal, written so a later desk cannot claim the short paper was unclear about members, dissent, scoring, identification failures, placebos, kernels, length-scale collapse, occupancy peaks versus the mean week, cash pinball tails, or the speech a decision board is allowed to make.

97. The seasonal decomposer as a member, not as a story about seasons

The first member is a seasonal decomposer. It is allowed to see a trend, a seasonal block, and a leftover. It is not allowed to see tomorrow’s print. It is not allowed to rewrite last winter after this winter has arrived. The seasonal block is estimated on the expanding window that ends before the skill window. A decomposer that is refit on the skill window is a memoir. We refuse the memoir.

What the decomposer is good at is a series that still has a week and a year. Cash that still has a Friday and a quarter-end is such a series. Occupancy that still has a Tuesday and a respiratory season is such a series. A series that has lost its week because a regime ate the week is a series on which this member should lose. Losing is useful. The board is required to show the loss. A decomposer that is pretty on the training year and dead last year receives a small last-window weight. That is the point of CRPS weights. It is not a reason to delete the member. Deleting the member would hide the fact that the season is no longer the season the desk remembers.

The decomposer issues a predictive law, not a point. If the implementation only issues a point, the point is wrapped in a residual scale estimated on the same expanding window, and that scale is named. A point without a scale is not a member of this ensemble. The Gaussian closed form of CRPS is then legal for this member only if the predictive is named as Gaussian. If the leftover is a sample, the energy form is used instead. Mixing the two without naming which is a different product.

Hierarchical leaves from this member are still leaves. If the desk has a tree, the decomposer is fit on the leaves or on the parent, not on both as if they were independent truths. After the three members have spoken, MinT reconciliation is applied so the leaves add to the parent. The decomposer does not get a special exemption because its seasonal block already “looks additive.” Looking additive is not an adding-up identity. The identity is S times the reconciled bottom vector. That is what a treasury parent is allowed to see.

A later desk that wants to replace the decomposer with a prettier filter will have to show last-window CRPS on the same series and horizon, on a window later than the weight window. A prettier filter that wins on the training year and loses on the later window is not a replacement. It is a story. Stories do not sit in the path cell.

98. The ARIMA–GARCH member as a mean and a scale, kept separate

The second member is an ARIMA–GARCH sleeve. The mean is an ARIMA on the expanding window. The scale is a GARCH on the leftover of that mean, also on the expanding window. The two are not one object. A desk that widens a mean because the scale asked for a larger story is using the scale as a warning. A desk that moves the mean because the scale was loud is using a different product.

The ARIMA orders are frozen before the quoted skill window. Searching orders on the skill window is a shuffle. It is refused. If a later window says the orders are stale, the orders are re-declared and the skill window moves later still. There is no k-fold on the full series that then pretends the orders were known. An expanding cut is auditable. A random purge is a different protocol.

GARCH here is a warning about leftover squares. If the estimated persistence pair wants to sit on the covariance-stationary boundary, the sleeve is marked near-boundary and is not used to widen a position or to invent a sharper path. If the estimated sum exceeds one, the sleeve is marked explosive and is dropped from the mix for that series and horizon. Volatility clustering is information. It is not a product. It is not a contrast. It is not a reason to fill the probe cell.

The predictive issued by this member is named. The usual operating predictive is Gaussian with the ARIMA mean and the GARCH scale. CRPS then uses the Gaussian closed form. If a later desk wants a t-scale, the degrees of freedom are declared before the skill window and the energy form is used on a sample, or a named closed form is written. An unnamed fat tail is a different product.

This member will often win on a cash path that has lost its week but still has a short memory and a clustering scale. It will often lose on an occupancy peak that is a first freeze, not a Tuesday. Losing on the first freeze is not a failure of GARCH. It is the member telling the board that last week’s scale is not this week’s ward. The dissent weight stays positive under last-window softmax. Stacking on a short board may put a zero on this member when the decomposer and the neural sleeve can span the four dates. That zero is why stacking is not the operating mix on a short board.

99. The neural seasonal sleeve, and what attention is not allowed to be

The third member is a neural seasonal sleeve. It is a function of lagged targets and dated calendar features on the expanding window. It is not a graph. It is not an instrument. It is not a reason to speak a contrast. Feature attribution on this member is a plot of this member. A beautiful attribution that moves when the sleeve is retrained is still a plot. Moving the plot into the contrast cell is the failure mode the identification section exists to stop.

The sleeve may be recurrent, convolutional, or a dated residual block. The architecture is frozen before the quoted skill window. Tuning depth on the skill window is a different CRPS. The August and September revisions do not retune to fatten a later number. A pinball-optimal head is allowed for named cash quantiles. A CRPS-optimal head is allowed for a full predictive. One head that claims to be both without a named training loss is a different product.

A neural sleeve that violates an adding-up identity is not explained with attention. Attention is not reconciliation. The sleeve is reconciled by MinT or it is dropped. A slide that shows the unreconciled neural parent because it is prettier is a slide of a different product. Hierarchical cash and hierarchical occupancy are different trees. The sleeve is not allowed to reconcile them to each other. A number that claims to be both is the blended accuracy this paper deleted.

The sleeve will often win on a series that has a weak week and a strong dated event the decomposer cannot see as a season. It will often lose when the dated event is the target in costume: a holiday dummy that is the cash print, a testing dummy that is the occupancy print. In that case the sleeve is fitting the target and calling it a feature. Last-window CRPS will still produce a weight. The dissent from the other members is the only honest speech. Hiding that dissent is a different product.

Incidence may enter this sleeve as a dated regressor from a compartment fit. A refit of that compartment model inside the sleeve is refused. The compartment paper owns the compartment. This paper owns the path. Sentiment leftovers may enter as columns after the obvious tape has been removed. They do not replace the target. A sleeve that is mostly a sentiment leftover is a sleeve that should lose on a later window when the leftover is restated. Restated leftovers are not written back into the training window.

100. Last-window softmax weights, written as an operating mix

Members m equals 1, 2, 3 produce predictives for a named series and a named horizon. Weights follow recent continuous ranked probability score on a weight window that ends before the skill window. The operating mix is a softmax: w sub m is proportional to exp of minus lambda times the last-window mean CRPS of member m. Lambda is declared before the skill window. It is not chosen to make the dissent look small. A lambda that sends the loser to machine zero is stacking in costume.

The four-date toy in the worked note is the arithmetic, not the skill table. Mean CRPS of 0.19, 0.28, and 0.41 at lambda equals 4 produces weights near 0.52, 0.36, and 0.12. The third member is the dissent. The board shows 0.12. It does not hide it. Constrained stacking on the same four dates, minimizing CRPS of the mixture on the simplex, produces weights near 0.61, 0.39, and 0.00. It kills the dissent. That is why this paper uses the softmax of last-window CRPS, not stacking, on a short board.

A short board is a board that cannot audit a stacking program as cleanly as it can audit three scalar scores and a softmax. If a later desk wants stacking, the simplex constraint stays: weights at least zero, weights sum to one. Negative weights are a different product. They hide dissent by letting a winner cancel a loser. The board then no longer shows who lost. Hiding the zero that stacking may produce is also refused. If stacking puts a zero on a member, the board still shows the zero and the member’s last-window CRPS.

The weight window is earlier than the skill window. Using the skill window to set the weights is a shuffle. It is refused. Expanding the weight window through the skill dates is the same shuffle in a kinder costume. A member that is pretty on the training year and dead last year receives a small weight. That sentence is the entire reason the weights are last-window and not lifetime. Lifetime weights bury a dead year under a pretty decade.

Drift in the ensemble residual after the mix is an alarm that the mix is stale. The alarm is a boolean. It is not a new member. It is not a new causal graph. It is not a reason to refill the contrast cell with an attribution plot. When the alarm trips, the board is required to show the current weights and the dissent. If the dissent is the only member that has ever seen a regime like the present, the mix is stale and the pattern note should trip. The stack does not invent a fourth cell to hold that sentence.

101. Why stacking may zero dissent, and why that is not a feature on a short board

Constrained stacking solves a least-squares-like program in probability space: choose a weight vector on the simplex that minimizes the sum of CRPS of the mixture against the realizations on a window. On a short window the program is allowed to put a zero on a member that is linearly redundant with the other two, even if that member is the only one that would have been honest in a regime the window did not contain. The zero is a feature of the program. It is not a feature of a human board that needs to see who lost.

Dissent is the member that lost on the last window and still has a positive weight. It is not a personality. It is not a hedge fund. It is the dashed stroke on the ensemble figure. A board that draws only the mixture band has hidden the dissent. A board that draws the dissent in a color the room cannot see has hidden the dissent. A board that speaks only the mixture mean has hidden the dissent. All three are different products.

The reason last-window softmax keeps a small positive weight is mechanical. A finite lambda on a finite CRPS gap cannot send a weight to zero. The loser remains visible. That visibility is the operating requirement. It is not a claim that the loser will win next. It is a claim that the board can still name who lost. When a later desk uses stacking and the loser is zero, the desk must still name the zero. A stacking solution that is presented as a two-member mix without the third line is a presentation that has left this paper.

Negative weights, unconstrained stacking, and “optimal” affine combinations that go outside the simplex are refused for the same reason. A negative weight on a loser is a cancellation. The cancellation can make the mixture look skilled while both the winner and the loser are wrong in opposite directions. The board then cannot speak a member. It can only speak a cancellation. Cancellations are not dissent. They are a costume.

On a long window with many dates, stacking and softmax may agree closely. Agreement is not a reason to switch the operating mix on a short board. The short board is the board this paper writes for. The four-date toy exists so the two programs can be compared on numbers that fit in a table. It is not a cash-path skill table. It is not a reason to speak a blended percent.

102. CRPS in the loss convention, and the energy identity as the same number

Two conventions sit in the literature and they have opposite signs. Gneiting and Raftery treat a scoring rule as a reward to be maximized and write CRPS with a minus in front of the integrated squared distance between the forecast CDF and the indicator of the realization. Forecast verification usually treats CRPS as a loss to be minimized and drops the minus. This paper uses the loss convention, as in Hersbach and in the energy-form identity of Baringhaus–Franz and Székely–Rizzo. A later desk that quotes a number from a notebook must name which convention the notebook used. Mixing a reward CRPS with a loss CRPS in one table is how a member that lost becomes a member that won.

The loss is the integral over the real line of the squared distance between F of z and the indicator that z is at least y. The energy form is the expected absolute deviation of a draw X from y, minus half the expected pairwise absolute deviation of two independent draws from F. The two expressions are equal on the class of Borel probabilities with a finite first moment. They are not two scores. Using the integral for the Gaussian member and the energy form for a sample is computation. Using both on the same member without naming which is a different product.

On that class, CRPS in the loss convention is strictly proper: the expected loss is uniquely minimized when F is the true law of Y. The proof is a Bregman argument. Expanding the integrand and taking expectation under the true law leaves the integrated squared distance between F and G, which vanishes only when the CDFs agree almost everywhere. A member that is sharp and wrong is punished. A member that is wide and honest can win on CRPS and lose a point contest. That is the point of using CRPS on paths. A point contest that is spoken as if it were CRPS is the failure mode the July draft still had.

If F is Gaussian with mean mu and scale sigma, the loss has the closed form sigma times the usual expression in z equals (y minus mu) over sigma, involving the standard normal density and CDF and a one-over-square-root-pi term. We use that formula for the Gaussian member. We use the energy form for an ensemble represented as a sample. The mixture after softmax weights is not automatically Gaussian. Scoring the mixture as if it were Gaussian with the mixture mean and a hand-waved scale is a different score. If the mixture is scored, it is scored as a mixture: either a named closed form or the energy form on draws from the mixture.

Hersbach’s decomposition into reliability and resolution is allowed as a diagnostic. It is not a second operating score. A desk that speaks only the reliability term has hidden the resolution term. A desk that speaks a percent that averages CRPS across series and horizons has hidden both. This paper does not speak that percent.

103. Scoring members without blending series, horizons, or products

Each member is scored on each named series and each named horizon. The score is CRPS for a full predictive and pinball for a named quantile. Those two are not interchangeable. Median pinball is half the mean absolute error of the median. CRPS integrates Brier scores over all thresholds. A distribution can have a good median and a bad tail. The two scores will then disagree. That disagreement is information. Averaging it away is the failure mode.

The skill window is later than the weight window. Occupancy skill is a later winter than the fit. Cash skill at thirteen weeks is a later block than the block that set the softmax weights. A notebook that reports CRPS on the same dates that set the weights is reporting a fit. Fits are allowed in a methods note. They are not spoken as skill. The four-date toy is a fit. It is labeled as a toy.

No blended accuracy across asset classes is a result of this paper. CRPS on a cash path at thirteen weeks is not MAE on occupancy peaks and is not a residual-class AUC. Speaking one percent for all three is the sentence the September revision deletes. A table that averages those percents with inverse-variance weights is the same sentence in a kinder costume. A table that reports them in one row with a shared “skill” header is the same sentence. The skill table is indexed by series and horizon. If that index cannot be named, the number cannot be spoken.

Equity residuals are not scored here. A later desk that wants a residual-class AUC will read the equity paper as a paper. This paper does not convert that AUC into a CRPS and does not convert this CRPS into a Sharpe. Contrast skill is placebo and pre-trend, or empty. A path that looks like an impulse is still a path. Probe skill is not a CRPS. A cheap model that found a pretty corner is still a cheap model.

When a human still owns the path, the overlay on the board is the mixture band, the dissent stroke, and the named score for that series and horizon. The overlay is not a percent that claims to cover cash, occupancy, and equity. The overlay is not a do-operator. The overlay is not an acquisition star. Those belong in the other cells, or they belong in silence.

104. The instrumental functional as a ratio, and the first stage as a gate

Let x be a treatment, y an outcome, z an instrument, and u a confounder. The population IV functional is tau equals Cov of z and y over Cov of z and x. That ratio equals the coefficient on x in the second stage when Cov of z and the second-stage leftover is zero and Cov of z and x is not zero. In a sample, after demeaning, the ratio is the inner product of z with y over the inner product of z with x. Division is legal when the first-stage covariance is nonzero. Weak-instrument bias is the usual concentration-parameter expansion. Writing a do-operator in a notebook does not repair a weak first stage.

The concentration parameter is pi squared times z transpose z over the first-stage leftover variance. Stock and Yogo tabulate critical values. A first-stage F below those values is a weak instrument. The estimator is then biased toward the ordinary least-squares functional. The usual two-stage t-ratio is not a standard normal under the null. If the instrument is irrelevant, pi equals zero, the denominator is a mean-zero Gaussian form, and the ratio is a Cauchy-type functional. We refuse the contrast cell rather than report that functional with a footnote.

The six-point toy in the worked note is the arithmetic. On those demeaned points, z transpose x is 2.09 and z transpose y is 2.35, so the ratio is about 1.12. The first-stage F is large because z and x were drawn together. Replacing z by an independent standard-normal draw of the same length produces a denominator near zero and a ratio that jumps when the draw is rerun. That is the refusal on numbers: a weak instrument is refused, not footnoted. The toy is not a causal claim about a named shock.

An instrument that is the target in costume is refused even when F is large. A costume is a z that is y with a delay, or y with a seasonal mask, or y with a vendor label. Large F then says the costume is correlated with x, which is not exclusion. Exclusion is Cov of z and the second-stage leftover equal to zero. A notebook cannot test that with a pretty F. A notebook can fail it when z is the news that y is. When the instrument became the news, the contrast cell empties.

We do not keep a weak-instrument number in a lead paragraph with a footnote. We do not keep a costume number in a lead paragraph with a caveat. Empty is a result. Empty is preferred to a number that the next independent draw will move.

105. Identification failures that empty the contrast cell

If the directed graph cannot be drawn without a cycle, we do not estimate a contrast. A cycle is not repaired by a lagged arrow that is the same arrow. A cycle is not repaired by writing do on a node. A cycle is a reason to leave the cell empty. The graph the desk will defend names an instrument, a treatment, a confounder, and an outcome. If those four cannot be named without a fifth node that is all of them, the cell is empty.

Pre-trend failure empties the cell. A pre-trend that already wants the post-treatment gap is not a caveat. It is a reason to delete the gap. We do not keep the gap and write a footnote. A moving placebo that is as large as the treated path is the same reason. Placebo-in-time and placebo-in-space are required when the design is a synthetic control. They are also required in spirit when the design is a difference that claims a date: a fake date that produces a gap as large as the real date is a moving placebo. The cell empties.

An instrument that is the target in costume empties the cell. A treatment that is the outcome with a filter empties the cell. A confounder that is left on an unblocked back-door path empties the cell. Feature attribution on the neural member does not fill the cell. A beautiful attribution that moves when the member is retrained is a plot. Plots live with the path member. They do not become interventional functionals by being drawn in the same aspect ratio as an impulse.

An unnamed shock is a residual. Plotting a residual as an impulse is a different product. An impulse figure exists only when the shock is named. The name is a sentence the desk will defend: this is the instrument, this is the treatment, this is the leftover that is assumed orthogonal. If the leftover is the news, the figure is deleted. If the shock is “the market” or “the season” or “the virus” without a dated series that is not the target, the figure is deleted.

Empty is preferred to a number with a footnote. The September revision is the revision that made that sentence operational. A later desk that wants to keep the number will have to close a back-door, a front-door, or an instrument on a pre-declared window, and will have to show quiet placebos and a first stage that is not weak. Until then the contrast cell is blank. Blank is speech. It is not a missing product.

106. Back-door adjustment as a named criterion, not as a do costume

A set W satisfies the back-door criterion when it blocks every back-door path from treatment x to outcome y and contains no descendant of x. The adjustment formula is then the mixture of the observational conditional p of y given x and W, mixed over the observational law of W. That mixture is the interventional law only if the criterion holds. If a hidden common cause of x and y remains on an unblocked path, the observational functional is a mixture over that hidden node and is not the interventional law. The interventional law mutilates the incoming arrows to x and does not mix over the same hidden value.

We write the name of the criterion because a notebook that only types a do-operator has not identified anything. The notebook must name W. It must name the paths that W is claimed to block. It must name why W is not a descendant of x. A W that is a descendant of x is a bad control. A W that is the outcome with a filter is a costume. A W that is chosen after seeing the post-treatment gap is a story. Stories do not sit in the contrast cell.

The back-door theorem is Pearl’s. The converse we use operationally is smaller: one open hidden path is enough to refuse the functional. We do not search a large set of candidate W until a pretty number appears. We declare W before seeing post-treatment outcomes, or we leave the cell empty. Searching W on the outcome is the causal analogue of setting ensemble weights on the skill window. Both are shuffles. Both are refused.

If W cannot be named, the cell is empty unless a front-door or an instrument closes. Those are the next two gates. They are not decorations on a back-door that failed. A desk that reports a back-door number and a front-door number and an IV number and then averages them has invented a blended contrast. This paper deletes blended contrasts the same way it deletes blended accuracy. One closed strategy on a pre-declared window, or empty.

107. Front-door adjustment, and the empty cell when neither door closes

The front-door criterion uses a mediator m that intercepts all directed paths from x to y and is itself unconfounded given x: no unblocked back-door from x into m, and no unblocked back-door from m into y once x is held. The functional is then a composition of observational pieces: the law of m given x, and the back-door-adjusted law of y given m. We write the name because a notebook that only says do has not identified anything. The notebook must name the mediator. It must name why every directed path goes through it. It must name why the mediator is not the outcome in costume.

A mediator that is the target with a delay is a costume. A mediator that is the treatment with a filter is a costume. A mediator that is chosen because it is the only column that produces a pretty number is a story. A mediator that is a neural attribution score is a plot. Plots are not mediators. The front-door is a graph criterion. It is not a feature-importance criterion.

If neither the back-door nor the front-door nor an instrument closes, the contrast cell is empty. Empty is a result. A later desk that wants a number will have to draw the graph without a cycle and name which criterion closed. Sharma and Kiciman’s library is a reading list for that later desk. It is not a license to fill the cell because a function returned a float. A float from a library that was pointed at the wrong graph is a float. It is not a contrast.

We do not keep an unidentified number in a lead paragraph with a footnote. We do not keep it in a methods note as “suggestive.” Suggestive is a different product. This paper’s contrast speech is placebo and pre-trend, or empty, on a pre-declared window. If those words cannot be spoken, the cell stays blank. The path cell may still speak. The probe cell may still speak. The three cells do not fill each other.

108. Placebos in time and in space, written as gates rather than decorations

A treated unit and a donor pool produce a weight vector on donors that matches pre-treatment paths. We refuse the method on a single national series with no donor. A beautiful post-treatment gap without a donor pool is a pretty line. Pretty lines are not contrasts. The treated unit and the donor pool are declared before seeing post-treatment outcomes. Donor weights are fit on the pre-period only. Fitting donor weights on a window that includes the post-period is a shuffle. It is refused.

Placebo-in-space is a leave-one-out or a treat-the-donor exercise: each donor is treated as if it had been treated, the rest of the pool is used to form a synthetic, and the placebo gaps are compared to the treated gap. If the treated gap is not unusual in that cloud, the cell is empty. A notebook that shows only the treated gap has hidden the cloud. Hiding the cloud is the same sin as hiding ensemble dissent. Both are pretty lines without the loser.

Placebo-in-time is a fake treatment date in the pre-period. The synthetic is formed using data before the fake date. The gap after the fake date is compared to the gap after the real date. If the fake date produces a gap as large as the real date, the cell is empty. One quiet placebo-in-time is the minimum. A notebook that searched fake dates until one was quiet is a notebook that set weights on the skill window. The fake date is declared, or several are declared, before the cloud is seen.

Leave-one-out in space and at least one placebo-in-time must be quiet relative to the treated gap. If they are not, the cell is empty. We do not keep the gap and write a caveat. We delete the gap. A synthetic control that is only a pretty line is a pretty line. The September revision is the revision that made that sentence a gate rather than a taste.

These gates are not only for synthetic control. A difference-in-differences that cannot name a donor-like comparison, a pre-trend check, and a fake date is a difference that has not earned a contrast. An event study that cannot show leads is an event study that has not earned a contrast. The names change. The empty cell does not.

109. Gaussian-process acquisition as a probe, not as a truth

When an experiment is costly, a Gaussian process on observed pairs of location and noisy value proposes the next location. The process is a cheap model. It is not the experiment. It is not the interventional law. It is not a path for a cash series. Posterior mean and variance at a test point are the usual kernel formulae: the mean is a kernel row times the inverse Gram times the observations, and the variance is the Schur complement. We write them so a later desk cannot replace the posterior variance by a heuristic distance to the nearest evaluated point. A notebook that replaces the posterior variance by that distance is a different acquisition.

Expected improvement is the expected positive part of the difference between the posterior at a test point and the incumbent f-star. For a Gaussian posterior with scale greater than zero, the closed form is (mu minus f-star) times Phi of z plus sigma times phi of z, where z is (mu minus f-star) over sigma, and phi and Phi are the standard normal density and CDF. The first term likes a higher mean. The second term likes a higher variance. The wider posterior can win. That is exploration. The four-number toy in the worked note is the arithmetic: incumbent 0.50, mean 0.62, scale 0.20 gives z of 0.60 and expected improvement about 0.154; the same mean with scale 0.05 gives about 0.120. The wider posterior is preferred.

The incumbent f-star is the best observed value, not a target written to make improvement look large. Resetting f-star to a hope is a different acquisition. When the posterior scale is zero, the expression collapses to the positive part of mu minus f-star by continuity. A notebook that reports expected improvement at a point the Gram has collapsed onto an observation, without a nugget, is reporting a numerical accident.

Upper confidence bound acquisition is the Srinivas–Krause–Kakade–Seeger choice, with a stated beta-t sequence that yields a cumulative regret bound under a bounded RKHS norm. We use expected improvement as the default because it is a closed form with no extra sequence. A later desk that wants the bound will state the sequence. Both acquisitions probe nonsense if the cheap model is a lie. The random member is the only defence named in the acquisition section.

When the budget is gone, the current x-star is the answer. It is a probe. It is not the truth. A desk that asks for one more probe after the budget is gone is a desk that did not state the budget. A desk that speaks x-star as a contrast has left this paper. A desk that speaks x-star as a cash path has left this paper.

110. A random probe every k steps, as a member of the search ensemble

A random probe every k steps is a member of the search ensemble. It is not a decoration. It is the only named defence when the cheap model is bad. Expected improvement and upper confidence bound will both climb a lie if the posterior mean loves a corner and the posterior variance has collapsed around that corner. The random member does not ask the cheap model where to go. It asks the box.

k is declared before the budget is spent. Searching k to make the final x-star look decisive is a shuffle. It is refused. A k that is larger than the budget is a k that never fires. That is a different search ensemble: it is expected improvement alone. This paper does not run expected improvement alone. The random member is required. The board shows when the last probe was random. Hiding the random probes and showing only the pretty EI path is the same sin as hiding ensemble dissent.

The random member is uniform on the box unless a later note names a different measure. A random member that is concentrated on the current EI maximizer is not random. A random member that is a small jitter around the last probe is not random. A random member that is a Sobol sequence declared before the budget is a legal later-desk variant. It is still required to fire every k steps. It is still required to be shown.

If the cheap model is good, the random member will often look wasteful. Waste is the price of the defence. A later desk that wants to drop the random member after a pretty run will have to show that the cheap model did not collapse its length-scales to the grid spacing and did not glue the posterior to a corner. Those shows are kernels and length-scales, not a prettier x-star. A prettier x-star after dropping the random member is the thing we refuse to treat as evidence.

The search ensemble is not the path ensemble. The path ensemble is a seasonal decomposer, an ARIMA–GARCH member, and a neural seasonal sleeve, mixed by last-window CRPS. The search ensemble is expected improvement, and a random probe every k steps, on a Gaussian process. Mixing those two ensembles into one “forecast” is how a desk double-counts a probe as a path. The three cells stay separate.

111. Kernels, nuggets, and the Gram the cheap model is actually using

The default kernel is squared exponential plus a small nugget. The squared exponential is infinitely smooth. That is a modeling choice, not a discovery about the experiment. A Matérn-5/2 alternative is allowed when the cheap model should be less smooth. The alternative is declared before the budget is spent. Switching to Matérn after seeing a pretty corner is a shuffle. Switching to a rational quadratic after seeing a pretty corner is a shuffle. Switching the nugget after seeing a singular Gram is allowed only as a numerical repair, and the repair is named. A silent repair is a different cheap model.

The nugget is there so the Gram is invertible when two probes land on the same point, or when two probes land closer than the length-scale wants to distinguish. It is not a reason to trust a posterior that has collapsed. A nugget that is large enough to make every posterior variance look the same is a nugget that has hidden exploration. We do not set the nugget to make expected improvement look decisive. We do not set it to make the board look calm.

Length-scales live on a diagonal matrix inside the squared exponential. They have log-normal priors and are not set to make the acquisition look decisive. A length-scale that is hand-set to the width of the box makes every point look near every other point. Expected improvement then collapses to a mean hunt. A length-scale that is hand-set to a tiny fraction of the box makes every point look far. Expected improvement then collapses to a variance hunt on the first unevaluated site. Both hand-sets are different products. The prior is named. The posterior length-scales are shown. Hiding them is hiding the cheap model.

The posterior variance is the Schur complement, not a heuristic distance to the nearest evaluated point. A notebook that uses distance as variance is running a different acquisition, closer to a Voronoi rule than to a Gaussian process. That rule is allowed only if it is named as that rule. It is not allowed to inherit the closed form of expected improvement, which assumes a Gaussian posterior. Mixing a Voronoi variance with an EI closed form is a costume.

Rasmussen and Williams are the reading list for the formulae. This note does not replace that book. It records the operating refusals: default kernel named, alternative named before the budget, nugget named, length-scale priors named, posterior variance not replaced by distance, and a collapsed length-scale treated as a broken cheap model rather than as a discovery.

112. Length-scale collapse, and a cheap model that loves a corner of the box

A length-scale that collapses to the grid spacing is a broken cheap model, not a discovery. Collapse means the posterior length-scale is on the order of the spacing between evaluated points, so the process has decided that only the nearest observation matters and every other site is independent at the prior variance. Expected improvement then plants flags in every unexplored pocket, or it glues itself to a single observation that looks good and treats the rest of the box as unknowable. Both behaviors can look decisive on a slide. Both are the cheap model saying it no longer has a length.

A Gaussian process that loves a corner of the box is the same family of failure. Love means the posterior mean is highest in a corner, the posterior variance is small there because several probes have already been spent there, and expected improvement keeps asking for one more point in that corner. The next probe in that corner is a probe of the cheap model, not of the experiment. The random member exists for that case. If the random member, when it fires, finds nothing in the rest of the box that beats the corner, the corner may still be the answer. If the random member finds a better region, the cheap model was a lie and the corner was a habit.

Length-scale collapse and corner-love often travel together. A collapsed length-scale makes the mean interpolating and locally confident. Local confidence plus a pretty observation is a corner. The board is required to show the length-scales and the location of the last several probes. A trail of probes along an edge, with length-scales at the grid spacing, is a trail the desk should not trust with the last of the budget. Spending the last probes on that trail because expected improvement is still hungry is how a budget is wasted on a habit.

We do not “fix” collapse by shrinking the box around the corner. Shrinking the box after seeing the corner is a shuffle. We do not fix it by freezing length-scales at a pretty value. Freezing after collapse is a story. We do mark the cheap model as broken, we keep the random member, and we state that x-star is a probe from a broken cheap model if the budget ends in that state. A probe from a broken cheap model is still a probe. It is not spoken as a truth. It is not spoken as a contrast. It is not spoken as a cash path.

Log-normal priors are the first defence, not a guarantee. A prior that is so tight it forbids collapse can also forbid a real short length. We do not tighten the prior to make the acquisition look smooth. We show the posterior. We show the probes. We let the random member fire. Those three sentences are the operating hygiene. They are not a theorem that collapse cannot happen.

113. MinT reconciliation so leaves add to the parent

Leaves must add to the parent. That is a treasury identity, not a modeling preference. For a summing matrix S that maps bottom-level paths to the full tree, MinT-style reconciliation replaces a vector of incoherent forecasts by S times the MinT bottom vector, where the bottom vector is the GLS-like map (S transpose W inverse S) inverse S transpose W inverse times the incoherent vector, for a stated error covariance W. We use a diagonal W unless a later window says otherwise. A full W is a later window. Inventing a dense W on the skill window is a shuffle.

The identity S times the reconciled bottom vector is what a parent series is allowed to see. A slide that shows the unreconciled neural parent because it is prettier is a slide of a different product. A neural sleeve that cannot be reconciled is dropped. We do not hide the break in an attention plot. Attention is not an adding-up identity. A seasonal decomposer that already “looks additive” is still reconciled. Looking additive is not the identity.

Hierarchical cash and hierarchical occupancy are different trees. They are not reconciled to each other. A number that claims to be both is the blended accuracy this paper deleted. Equity residuals are not in either tree. Incidence may enter a leaf as a dated regressor. It does not become a parent in the cash tree. Sentiment leftovers do not become parents. Political polarity does not become a parent. Parents are the summing identities the desk named before the window.

Wickramasuriya, Athanasopoulos and Hyndman are the reading list for MinT. This note records the operating choices: diagonal W by default, S named, unreconciled parents not spoken, neural sleeves reconciled or dropped, and no cross-tree blend. A later desk that wants a full W will declare it on a window earlier than the skill window and will show that the parent identity still holds. If the identity does not hold, the parent is not spoken.

Reconciliation is hygiene. It is not a member. It is not a fourth ensemble weight. Applying MinT and then rescoring CRPS on the reconciled parent is legal and is the parent score. Applying MinT and then speaking the unreconciled neural CRPS because it is smaller is a costume. The board shows the reconciled parent and the dissent among members before reconciliation, because dissent is about members, not about the summing map.

114. Pinball for named quantiles, and why it is not CRPS in costume

Pinball loss at level tau is u times (tau minus the indicator that u is negative). It is strictly consistent for a tau-quantile: the expected pinball of Y minus q is uniquely minimized at any tau-quantile of Y, except on the knife-edge where a whole interval of quantiles exists, in which case every point of that interval is optimal. We quote pinball on cash quantiles and CRPS on full distributions. One number for both is a different product.

Median pinball is half the mean absolute error of the median. That identity is why a desk that is used to MAE will think pinball at one half is “the” score. It is a score for the median. It is silent about the tail. Pinball at nine tenths is a statement about the upper tail. Pinball at one tenth is a statement about the lower tail. A neural sleeve that is beautiful at one half and dead at nine tenths is a sleeve with a bad tail. The tail is the information. Averaging the three pinballs into one percent is the same failure mode as blending asset classes.

The four-date toy shows the split on small numbers. Median pinball on the first member is 0.125. CRPS on the same member is 0.19. Pinball at nine tenths is 0.09. They are not the same number and they are not interchangeable. Member three can look less bad on median pinball than on CRPS because its scale is mis-scaled. That disagreement is why both scores exist. Speaking one percent for all three is the failure mode this paper deletes.

A pinball-optimal neural sleeve that violates an adding-up identity is reconciled or dropped. Attention is not a reconciliation. The parent series is allowed to see the reconciled identity, not the raw neural sum of quantile heads. Quantile heads can cross after reconciliation if the map is applied naively; crossing is named and repaired by rearrangement or by dropping the sleeve. Unnamed crossing is a different product.

Cash paths use CRPS on the full predictive and pinball on named quantiles, including a tail quantile that is not the median. Occupancy does not use pinball as the operating speech; occupancy uses MAE on peaks. Equity residuals do not use pinball here. Contrast does not use pinball. Probe does not use pinball. Those refusals are how pinball stays a quantile score instead of becoming a slogan.

115. Cash paths at a thirteen-week horizon, and pinball tails the board must draw

A cash path is a named series and a named horizon. Thirteen weeks is a horizon that must be named. A blended percent across thirteen weeks and across names is not a result. A blended percent across thirteen weeks and one week is not a result. The skill table is indexed by series and horizon. CRPS on the full predictive at thirteen weeks is one cell. Pinball at a named tail quantile at thirteen weeks is another cell. They may disagree. Disagreement is spoken. Averaging them into a cash “accuracy” is refused.

The weight window is earlier than the skill window. Thirteen-week weights are not set on the same thirteen-week block that is then spoken as skill. Using the skill block to set the weights is a shuffle. Expanding the weight window through the skill dates is the same shuffle. The September revision keeps the cut. A later desk that wants a rolling thirteen-week CRPS as a monitoring series may have it as a monitoring series. Monitoring is not skill speech.

The tail quantile is named. The operating default is a high quantile for a cash shortfall and a low quantile for a cash surplus, each named, each scored with pinball. A sleeve that is beautiful on the median and dead on the shortfall quantile is a sleeve the board must show as dissent if it loses the tail. The overlay on a human board, when a human still owns the path, is to draw the tail, not to average it away. A fan chart that stops at the interquartile range has hidden the tail. Hidden tails are the cash analogue of hidden dissent.

Hierarchical cash is reconciled. Leaves add to the parent. A neural sleeve that is pretty on a leaf tail and impossible on the parent is reconciled or dropped. A parent tail that is spoken from the unreconciled neural sum is a costume. MinT with diagonal W is the default. A full W that is estimated on the skill window is a shuffle.

Cash is not occupancy. Cash is not an equity residual. Cash is not a contrast. A thirteen-week cash CRPS is not a reason to fill the contrast cell with a CPI print. A thirteen-week cash tail is not a reason to probe a box. A thirteen-week cash path may use incidence or sentiment as dated columns. Those columns do not become the target. If the columns are restated after the weight window, the restated values are not written back. Restatement write-back is how a walk-forward becomes a memoir.

116. Occupancy peaks versus the mean week, and the later winter than the fit

MAE on occupancy is quoted on peaks, not on the mean week. A model that is good on Tuesdays and late on the first freeze is a model that will look skilled on a mean-week MAE and useless to a ward. The peak is a named event: a high quantile of daily occupancy in a named window, or a first crossing of a named threshold, declared before the later winter is seen. Searching the definition of peak on the later winter is a shuffle. It is refused.

The later winter than the fit is the test. The winter it was trained on is not a test. A notebook that reports peak MAE on the training winter is reporting a fit. Fits are allowed in a methods note. They are not spoken as skill. A notebook that averages peak MAE with Tuesday MAE has invented a mean-week score in costume. The mean week is useful for a decomposer diagnostic. It is not the operating speech.

Incidence as a dated regressor may enter the occupancy member. A refit SEIRD inside the occupancy member is refused. The compartment paper owns the compartment. This paper owns the path. A dated incidence column that is restated after the fit is not written back. Testing-regime breaks are marked. They are not smoothly interpolated. Smooth interpolation is how a break becomes a season.

The three members still apply. The seasonal decomposer will often win on a winter that is still a winter. The ARIMA–GARCH member will often win on a shoulder week and lose on the first freeze. The neural sleeve will often win on a dated event and lose when the dated event is the occupancy print in costume. Last-window softmax keeps the dissent. Stacking may zero the loser on a short shoulder window. The board shows the loser. Occupancy peaks are why the loser matters: the loser is often the only member that is late in a useful way, and “late in a useful way” is speech a ward can hear. A mixture that hid that lateness would be a prettier mean week.

A single occupancy percent that claims to be a thirty-day season is the speech the table forbids. Occupancy is not cash. Occupancy is not an equity residual. Occupancy is not a contrast. A peak that arrives after a policy date is still a peak, not an impulse, unless a graph closes. Most occupancy graphs do not close. The contrast cell stays empty. The path cell still speaks peak MAE on the later winter. That split is the paper.

117. Decision-board speech, line by line, without a blended percent

Name the series and the horizon first. If those two words cannot be said, the board has not started. A board that starts with a percent has started in the failure mode. A board that starts with a do-operator has started in the contrast cell without a graph. A board that starts with an acquisition star has started in the probe cell without a budget.

Fit the three members on the expanding window: seasonal decomposer, ARIMA–GARCH, neural seasonal sleeve. Form last-window CRPS weights with a declared lambda. Show the mix. Show the dissent. If stacking is used by a later desk, show the zero if a zero appears. Do not speak a mixture mean without the dashed stroke. Do not speak a lifetime weight. Do not set weights on the skill window.

Draw the DAG. Name instrument, treatment, confounder, outcome, or name the back-door set, or name the front-door mediator. If no identification strategy closes, leave the contrast blank. Blank is speech. Do not fill the blank with an attribution plot. Do not fill the blank with a path. Do not fill the blank with a CPI print that has already been used as a path. If a synthetic control is the design, show placebo-in-time and placebo-in-space. If an instrument is the design, show the first-stage F against the Stock–Yogo gate and refuse a weak first stage.

If an experiment is funded, update the Gaussian process with the named kernel and the named nugget. Take x-star from expected improvement, and fire a random probe every k steps. Show the length-scales. If they have collapsed to the grid spacing, say the cheap model is broken. If the process loves a corner, say so. When the budget is gone, the current x-star is the answer. It is a probe. It is not the truth.

Speak only the metric in the skill table for that object. Cash path: CRPS and pinball at named quantiles, on a window later than the weight window, horizon named, thirteen weeks named when it is thirteen weeks, tails drawn. Occupancy: MAE on peaks, not on the mean week, on a later winter than the fit. Equity residual: not scored here. Contrast: placebo and pre-trend, or empty, on a pre-declared window. Probe: a location and a budget, not a CRPS. A board that speaks a blended percent has left this paper. A board that hides the dissent has left this paper. A board that fills the contrast cell with a plot has left this paper. Silence in the contrast cell is allowed. Silence in the path cell is not, once the series and the horizon have been named.

118. What each member is for, restated as permissions rather than architecture

The seasonal decomposer is permission to use a week and a year that still exist. It is not permission to invent a week that a regime ate. It is not permission to rewrite last year after this year arrived. It issues a law. It accepts a small weight when the week is gone. It remains on the board as dissent when it loses.

The ARIMA–GARCH member is permission to use a short memory and a clustering scale. It is not permission to move a mean because a scale was loud. It is not permission to keep an explosive scale. It issues a named Gaussian law unless a later desk names another law. It will lose on a first freeze and should lose. That loss is useful speech.

The neural seasonal sleeve is permission to use dated events that are not the target in costume. It is not permission to speak a contrast. It is not permission to skip reconciliation. It is not permission to replace incidence with a refit compartment model. Attribution on this sleeve is a plot. Plots are permitted in the path cell as diagnostics. They are not permitted in the contrast cell as functionals.

None of the three is permission to blend asset classes. None of the three is permission to set weights on the skill window. None of the three is permission to hide the loser. The operating mix is last-window softmax of CRPS. Stacking is a later-desk program that may zero a member; the zero, if it appears, is shown. Negative weights are not a permission.

The three members are the path cell. They are not the contrast cell. They are not the probe cell. A member that is used as an instrument is a costume. A member that is used as an acquisition is a costume. A member that is used as a blended percent is the July draft. The September revision deletes that draft’s speech.

119. Dissent as a visible loser, not as a personality and not as a hedge

Dissent is the member that lost on the last window and still has a positive weight. The ensemble figure draws it as a dashed stroke. The board names its last-window CRPS and its weight. A room that cannot see the dashed stroke is a room that is looking at a different product. A room that can see it and does not name it is the same product with better eyesight.

Dissent is not a hedge. A hedge would be a position. This paper does not size a book. Dissent is not a person. This paper does not wake up. Dissent is not a fourth cell. Drift in the residual of the mix may trip a boolean that the mix is stale; that boolean is an alarm, not a member, and not a graph. When the alarm trips, the dissent is often the information: the loser may be the only member that has seen a regime like the present. The board shows that sentence. It does not fill the contrast cell with it.

Last-window softmax keeps dissent because a finite lambda cannot send a finite CRPS gap to a zero weight. Stacking may zero it because a short-window linear program is allowed to treat a member as redundant. Redundancy on a short window is not redundancy in the next regime. That is the operational reason the short board keeps softmax. A later desk with a long window may stack, and must still show a zero if a zero appears. Hiding the zero is hiding the loser.

Negative weights hide dissent by cancellation. Unconstrained stacking is refused. A “clever” affine combination that puts a large positive weight on a winner and a large negative weight on a loser can make the mixture look skilled while both members are wrong. The board then cannot name a member. It can only name a cancellation. This paper’s board names members.

The dissent stroke is drawn before MinT and after MinT if the parent is the spoken series. Reconciliation can move a parent. It should not be used to hide a member that lost on the leaves. Leaf dissent and parent dissent may disagree. Both are shown when both are spoken. A parent that looks calm because the losers cancelled across leaves is a parent that needs the leaf strokes. Cancellation across leaves is another costume for hidden dissent.

120. Scoring hygiene: windows, forms, and the refusal to average products

CRPS is scored in the loss convention. The energy identity is the same number, used for samples. The Gaussian closed form is used for named Gaussian members. The mixture is scored as a mixture. Pinball is scored at named quantiles. MAE on occupancy is scored on peaks. Placebos and pre-trends are scored as quiet or not. Expected improvement is not scored as CRPS. A cheap model is not scored as pinball. These sentences are the scoring table in words.

Windows are ordered. The expanding fit window ends before the weight window’s last date, or they coincide only when the note says they do. The weight window ends before the skill window. Occupancy skill is a later winter. Cash skill is a later block. Contrast windows are pre-declared. Probe budgets are pre-declared. A notebook that moves a cut after seeing a pretty number is a notebook that has left the protocol. The pretty number is then a fit.

Forms are named. Loss versus reward for CRPS is named. Integral versus energy is named. Closed form versus sample is named. Diagonal W versus full W for MinT is named. Squared exponential versus Matérn-5/2 is named. Expected improvement versus upper confidence bound is named. Softmax versus stacking is named. Naming is how a later desk audits. Unnamed mixing is how a later desk double-counts.

Averaging products is refused. Cash CRPS, occupancy peak MAE, and residual-class AUC do not share a percent. Thirteen-week cash and one-week cash do not share a percent. Median pinball and tail pinball do not share a percent. A back-door functional and an IV functional do not share a contrast. Expected improvement and CRPS do not share a skill. The July draft’s blend is the example. The September revision is the deletion.

Speech follows the table. If the object cannot be named, the number cannot be spoken. If the horizon cannot be named, the number cannot be spoken. If the window cannot be named, the number cannot be spoken. Those three gates are enough to stop most slogans. What remains is a path with a band and a dissent, a contrast that is empty or closed, and a probe that is a location with a budget.

121. Weak first stages, concentration, and the refusal to footnote a Cauchy ratio

The population ratio tau equals Cov of z and y over Cov of z and x is a functional. The sample ratio after demeaning is an estimator of that functional only when the first stage is not a mean-zero form. Stock and Yogo’s tables exist because “not a mean-zero form” is not a vibe. It is a first-stage F compared to a critical value that depends on the number of instruments and the number of included treatments. This paper’s operating gate is that comparison. Below the gate, the contrast cell is empty.

A weak first stage biases the ratio toward the ordinary least-squares functional. That is the usual concentration-parameter expansion. It is not repaired by a larger notebook. It is not repaired by a do-operator. It is not repaired by a smaller standard error from a formula that assumed a strong first stage. Reporting the OLS number and the IV number side by side without the F is how a desk lets the room pick the prettier of two unidentified functionals. We report F, or we report empty.

When pi equals zero, the denominator is z transpose nu after demeaning, a mean-zero form. The ratio of two jointly Gaussian forms is Cauchy-type. It does not have a Gaussian limit. The usual two-stage t-ratio is not a standard normal under the null. A six-point independent z makes this visible: the ratio jumps when the draw is rerun. Jumping is the diagnostic. A later desk that wants a Anderson–Rubin region instead of a point may have one, and must still not speak a point. This paper’s short board speaks empty rather than a region, because a region is a later-desk product with its own speech.

An instrument that passes F and fails exclusion is still refused. Exclusion is not F. Exclusion is Cov of z and the second-stage leftover equal to zero. A z that is the news, or the target with a delay, or the seasonal mask of y, can pass F and fail exclusion. Passing F is necessary for the ratio to be a number we will look at. It is not sufficient for the cell to fill. The graph must still close. Placebos and pre-trends must still be quiet when the design needs them.

We do not keep a weak-instrument point in a lead paragraph with a footnote. Footnotes are how the July draft’s blend survived in costume. The September revision prefers an empty cell. Empty is auditable. A footnote is a conversation. This paper is a protocol, not a conversation about a number that will move on the next independent draw.

122. When the graph is a costume: attribution, unnamed shocks, and residual impulses

Feature attribution on the neural seasonal sleeve is a plot of that sleeve. Integrated gradients, permutation, Shapley-style reallocations, attention rolls, and saliency maps are all plots of that sleeve. They move when the sleeve is retrained. They move when the expanding window adds a year. They are useful diagnostics for a path member. They are not back-doors. They are not front-doors. They are not instruments. A slide that puts an attribution bar next to a do-operator has left this paper.

An unnamed shock is a residual. An impulse-response figure exists only when the shock is named. The name is a sentence: this dated series is the instrument or the treatment, this leftover is assumed orthogonal, this outcome is y. “The market,” “the season,” and “the virus” are not names unless they are dated series that are not the target in costume. Plotting a residual as an impulse because the residual is large is a different product. Large residuals are path diagnostics. They are not contrasts.

A moving placebo that is as large as the treated path is a reason to empty the cell even when a graph was drawn. Drawing a graph does not silence a placebo. A pre-trend that already wants the post-treatment gap is the same reason. We do not keep the gap and write a caveat. We delete the gap. A later desk that wants to keep the gap will have to show that the placebo family is quiet on a pre-declared window. Until then the cell is empty.

Cycles empty the cell. A lagged cycle that is the same cycle empties the cell. A “simultaneous” block that is a cycle with extra notation empties the cell. The desk will defend a directed acyclic graph or it will not speak a contrast. There is no third option in which a cycle is “estimated.” Estimation of a cyclic simultaneous system is a different paper with different refusals.

The contrast cell does not borrow speech from the path cell. A beautiful ensemble band around a policy date is still a band. It is not an impulse. The contrast cell does not borrow speech from the probe cell. A pretty x-star after a policy date is still a probe. The three cells stay separate when the pictures are pretty. Pretty is when mixing is most tempting. Pretty is when the protocol is for.

123. Donor pools that do not contain the regime, and paths the ensemble cannot invent

Synthetic control has the same problem in the contrast cell. A donor pool that does not contain the regime cannot form a synthetic that matches the treated unit after the regime arrives. Pre-treatment fit can still look beautiful. Post-treatment gaps can still look large. Placebo-in-space will often look quiet if every donor is equally lost. Quiet placebos in a pool that is all lost are not quiet. They are a shared miss. The cell empties when the pool does not contain the regime, even if the gates on paper were followed. The gates assume a pool that can miss in different ways. A pool that misses as one object is not a pool.

The Gaussian process has the same problem in the probe cell. A box that does not contain a region like the present will send expected improvement to a corner of a lie. The random member will wander inside the same lie. x-star is then a probe of the wrong box. The defence is to name the box before the budget, not to grow the box after a pretty corner. Growing the box after a pretty corner is a shuffle.

These three failures are why the conclusion of the short paper names a regime the donors do not contain, an instrument that became the news, and a Gaussian process that loves a corner. They are not three slogans. They are the path cell, the contrast cell, and the probe cell failing in the way each is allowed to fail. The stack is still in use if the cells are named and the failures are spoken. The stack is not in use if the failures are averaged into a percent.

A later desk that wants to invent a regime with a neural sleeve is asking the sleeve to be a donor it is not. The sleeve may still issue a law. The law will be scored on a later window. If it loses, it is dissent. If it wins on the later window, it earned a weight. Winning on the training regime and losing on the present is the usual case. Last-window weights exist for that case.

124. Expected improvement’s two terms, and why a wider posterior is allowed to win

Expected improvement’s closed form has two terms. The first is (mu minus f-star) times Phi of z. It is large when the posterior mean sits above the incumbent and the standardized gap is not a tail event. The second is sigma times phi of z. It is large when the posterior scale is large and the standardized gap is not so large that the normal density has died. The second term is exploration. A point that is probably worse but poorly known can beat a point that is probably better and tightly known. The worked toy is the arithmetic: scale 0.20 beats scale 0.05 at the same mean.

A later desk that wants to kill exploration can send a hard constraint that mu must exceed f-star. That constraint deletes the second term’s right to win when mu is below f-star. This paper does not use that constraint. Improvement can be expected below the incumbent if the mass above the incumbent is fat enough. That is the positive-part expectation. Truncating it is a different acquisition.

A later desk that wants more exploration can use upper confidence bound with a stated beta-t. That is legal. The sequence must be stated before the budget. Searching beta to make x-star land on a favorite is a shuffle. We default to expected improvement because it does not require the sequence. We still require a random probe every k steps, because both closed forms probe nonsense when the cheap model is a lie.

The incumbent is the best observed value. It is not a target. It is not a cash number. It is not a peak occupancy. Resetting it to a hope makes both terms larger and is a different acquisition. When several observations share the best value, f-star is that value, not an average of near-bests chosen to smooth the field. Smoothing f-star is a story.

If the cheap model is bad, both terms are a lie. The first term hunts a false mean. The second term hunts a false variance. The random member is the defence. The board shows the two terms at x-star, or it shows the sum and the posterior scale. A star without a scale is a point contest. Point contests are not this acquisition.

125. How a decision board is allowed to talk about cash, occupancy, and silence

Cash speech names the series, names thirteen weeks when the horizon is thirteen weeks, names CRPS on the full predictive, names pinball at the median and at the tail quantiles, and draws the tail. It shows last-window weights and the dissent. It shows the reconciled parent if a tree was named. It does not speak a single percent. It does not speak a Sharpe. It does not speak a do-operator. It does not speak an acquisition star.

Occupancy speech names the site or the aggregate, names peaks, names the later winter than the fit, and speaks MAE on those peaks. It may show Tuesday error as a diagnostic. It does not speak Tuesday error as skill. It does not speak a thirty-day percent. It does not refit a compartment model in the room. Incidence, if used, is a dated column with a source paper. A testing-regime break is marked. The dissent is shown, because the loser on a first freeze is often the speech a ward can hear.

Equity speech in this room is a refusal: not scored here. A leftover column from the equity paper may enter a path member. The leftover is not rescored as CRPS. The leftover is not turned into a contrast. The leftover is not a reason to probe a box.

Contrast speech is a graph and a gate, or silence. Silence is allowed. Silence is preferred to a costume. If a number is spoken, the criterion is named, the window is pre-declared, and the placebos or the first stage are shown. Weak first stages empty the cell. Moving placebos empty the cell. Pre-trends that want the gap empty the cell. Attribution plots do not speak.

Probe speech is a budget, a kernel, a length-scale panel, an expected-improvement star, and a mark when the last probe was random. If length-scales collapsed, that is spoken. If the process loves a corner, that is spoken. When the budget is gone, x-star is spoken as a probe. It is not spoken as a truth. A room that asks for a blended percent after all of that speech has asked for the July draft. The September revision’s answer is no.

126. Columns that may enter a path, and columns that may not become the target

Sentiment leftovers and political polarity may enter as columns. They enter after the obvious tape has been removed, and only if they are dated. A leftover that is restated after the weight window is not written back. A leftover that is a renaming of the target is refused. Collinearity with a seasonal block is a reason to drop the leftover, not a reason to keep both and let the neural sleeve split a story.

Incidence may enter as a dated regressor from a compartment fit. The compartment paper owns the fit. This paper does not refit SEIRD inside the ensemble. A restated incidence series is not written back. A testing-regime break is marked. Incidence is not a parent in a cash tree. Incidence is not a shock unless a graph closes, which it usually does not.

Equity residuals may enter as finished leftover columns from the equity paper. They are not scored here. They are not CRPS. They are not a Sharpe. They are not a reason to speak a contrast about a named firm. This paper does not print a name as a recommendation.

Calendar features may enter the neural sleeve and the decomposer. Calendar features that are the target in costume are refused. A holiday dummy that is the cash print is a costume. A reporting-day dummy that is the occupancy print is a costume. The test is whether the feature remains defined if the target is permuted. If it does not, it was the target.

Drift in the ensemble residual is not a column. It is an alarm. The alarm is a boolean. Adding the residual of the mix back into the mix as a fourth member is a different product. This paper has three members. The alarm may trip a pattern note. It does not trip a do-operator. It does not trip an acquisition unless an experiment is separately funded.

127. Hierarchical trees, diagonal W, and the parent a treasury is allowed to see

A tree is named before the window. Cash trees and occupancy trees are different objects. They do not share a summing matrix. They do not share a W. They do not share a skill percent. A notebook that stacks both trees into one MinT is a notebook that has invented a blend. The blend is refused.

S maps bottom-level paths to the full tree. The incoherent vector is the stack of member-mix paths at every node, or the stack of a single member if a member is being reconciled for a diagnostic. Operating speech reconciles the mix. Diagnostic speech may reconcile a member so the board can see whether the neural sleeve or the decomposer was the break. Diagnostic speech is labeled diagnostic.

W is diagonal unless a later window says otherwise. Diagonal W is a default, not a theorem that errors are independent. A full W estimated on the skill window is a shuffle. A full W estimated on the weight window and frozen is a later-desk variant that must be named. Shrinkage toward the diagonal is a later-desk variant that must be named. Unnamed dense W is a costume for a prettier parent.

The parent a treasury is allowed to see is S times the reconciled bottom vector. The parent a ward is allowed to see, when occupancy is hierarchical, is the same identity on the occupancy tree. An unreconciled neural parent is not allowed. An attention plot that “explains” the break is not allowed as a substitute. The break is reconciled or the sleeve is dropped.

After reconciliation, CRPS and pinball are computed on the reconciled series if that series is the spoken object. Computing them on the unreconciled series and speaking the smaller number is a costume. The dissent among members is still shown on the unreconciled members, because dissent is about who lost, not about the summing map. If reconciliation changes who lost on the parent, both stories are shown. Hiding the change is hiding a break.

128. The Gaussian closed form, the energy form, and the crime of mixing them unnamed

The Gaussian closed form for CRPS in the loss convention is a function of y, mu, and sigma. It is legal when the predictive is named Gaussian. The ARIMA–GARCH member is named Gaussian in the operating default. The seasonal decomposer is named Gaussian when its leftover is wrapped in a residual scale and that wrap is named. The neural sleeve is named Gaussian only if its head is a Gaussian head. A neural sample head is not Gaussian. A quantile head is not a full Gaussian law.

The energy form is legal when the predictive is a sample, or when the mixture is represented as a sample. Two independent draws from the same law give the pairwise term. A single draw against y gives the first term. Finite-sample energy on a small sample is biased; the sample size is named. Using ten draws and speaking the number as if it were the integral is a different product. Using a thousand draws is the usual later-desk choice. The four-date toy uses the Gaussian closed form because the toy members are Gaussian.

Mixing the two without naming which is how a member that lost on the integral wins on a ten-draw energy, or the reverse. The board names the form. A table that puts closed-form CRPS and energy CRPS in one column without a form mark is the July blend in costume. The September revision requires the mark.

The energy identity is not a second definition. It is Lemma 2.2 of Baringhaus and Franz, or identity (17) of Székely and Rizzo, used for computation. Strict propriety is the integral argument: expected integrated squared CDF error is a Bregman divergence and vanishes only when the CDFs agree almost everywhere. We do not prove propriety twice and we do not speak two scores.

If a later desk wants a threshold-weighted CRPS, the weight function is declared before the skill window. Tail-weighted CRPS is a later-desk cousin of tail pinball. It is not a license to blend cash and occupancy. It is not a license to skip named pinball on cash tails. This paper’s operating cash speech remains CRPS on the full law plus pinball at named quantiles, forms named, windows ordered.

129. Upper confidence bound as a later-desk acquisition, not as a silent default

Upper confidence bound acquisition is mu plus square-root-of-beta-t times sigma. Srinivas, Krause, Kakade and Seeger give a cumulative regret bound under a bounded RKHS norm when the beta sequence is stated. The sequence is the price of the bound. This paper does not pay that price by default, because a short board cannot audit a sequence as cleanly as it can audit a closed-form expected improvement and a random member. A later desk that wants the bound will state the sequence before the budget and will not search it after seeing a pretty corner.

UCB without a stated sequence is a slogan. A sequence that is “large enough to look exploratory” is a story. A sequence that is tightened until x-star lands on a favorite is a shuffle. All three are refused. The legal later-desk path is a named sequence, a named kernel, a named nugget, a random probe every k steps, and a board that shows length-scales.

UCB and expected improvement both probe nonsense if the cheap model is bad. UCB’s extra term is not a defence against collapse. Collapse makes sigma small near observations and large far away, so UCB will wander or glue depending on beta. That wander is not the random member. The random member does not ask sigma. A later desk that treats UCB as a substitute for the random member has dropped the only defence named in the acquisition section.

We do not run both UCB and expected improvement and then pick the prettier x-star. Picking after seeing the values is a shuffle. We do not average the two acquisitions. Averaging acquisitions is a blend. The search ensemble is one named acquisition plus the random member. The path ensemble is three named members plus last-window softmax. The two ensembles stay in their cells.

When the budget is gone, x-star is the answer whether the acquisition was expected improvement or UCB. It is a probe. It is not a contrast. It is not a cash path. It is not occupancy skill. A room that speaks a regret bound as if it were a peak MAE has mixed cells. Mixed cells are the failure mode.

130. What the July draft still said, and what the September revision deleted

The July 2026 draft still carried a blended accuracy across asset classes. That sentence is the failure mode this paper exists to stop. It treated cash CRPS, occupancy peak error, and an equity leftover as if they shared a percent. They do not. The 12 September revision deletes the blend. The skill table is indexed by series and horizon. Speech is per series and per horizon.

The July draft was also willing to let a pretty attribution sit near a contrast. The September revision sends attribution back to the path member. Feature attribution is a plot. It is not a back-door. It is not a front-door. It is not an instrument. The contrast cell is empty unless a criterion closes.

The July draft was willing to speak a mixture without a dashed stroke. The September revision requires the dissent. Last-window softmax keeps a positive weight on the loser. Stacking may zero it; the zero is shown. Negative weights remain refused.

The July draft was willing to treat expected improvement as a truth if the field looked decisive. The September revision treats x-star as a probe, requires a random member every k steps, and treats length-scale collapse as a broken cheap model. A corner-loving process is named, not celebrated.

The July draft was willing to keep a weak-instrument number with a footnote. The September revision empties the cell. Tau equals Cov of z and y over Cov of z and x remains the functional. A weak first stage remains a refusal. Placebos remain gates. Pre-trends that want the gap remain deletions.

This note is the long form of those deletions. It is not a new product. It is not a fourth cell. It is the same three cells, written so a later desk cannot claim the short paper was unclear.

131. How not to read the worked toys as skill, and how to read them as arithmetic

Four dates and three Gaussian members are arithmetic for last-window softmax versus stacking. Mean CRPS 0.19, 0.28, 0.41 at lambda 4 gives weights near 0.52, 0.36, 0.12. Stacking on the same dates gives a zero on the third member. The toy is small enough that the zero is visible. It is not a cash-path skill table. It is not occupancy skill. It is not a reason to speak a percent.

Six demeaned points are arithmetic for the IV ratio. Inner products 2.35 over 2.09 give about 1.12. An independent z of the same length gives a denominator near zero and a jumping ratio. The toy is small enough that the jump is visible. It is not a causal claim about a named shock. It is not a reason to keep a weak first stage with a footnote.

One Gaussian posterior is arithmetic for expected improvement. Incumbent 0.50, mean 0.62, scale 0.20 gives about 0.154. Scale 0.05 at the same mean gives about 0.120. The wider posterior wins. The toy is small enough that the exploration term is visible. It is not a truth about an experiment. It is not a cash path. It is not a reason to drop the random member.

Median pinball 0.125, CRPS 0.19, and pinball at nine tenths 0.09 on the first toy member are arithmetic for the refusal to speak one number for a median, a law, and a tail. They are not a cash skill table. They are the reason cash speech draws the tail.

A later desk that puts these toys in a lead paragraph as “results” has left this paper. Toys are arithmetic. Skill is a later window, per series and per horizon. Contrast is a closed graph or empty. Probes are budgets. The toys exist so the board can see the arithmetic of dissent, refusal, and exploration without pretending a four-date series is a winter or a treasury tree.

132. What remains when the three cells are named, and what is still refused

When the series and the horizon are named, the path cell can speak. The three members fit on the expanding window. Last-window softmax weights form. The dissent is shown. CRPS is spoken in the loss convention, energy identity used as the same number, Gaussian closed form used only for named Gaussians. Pinball is spoken on cash quantiles. Occupancy is spoken as peak MAE on a later winter. MinT holds the tree. Tails are drawn at thirteen weeks. No blend is spoken.

When a graph closes, the contrast cell can speak. Back-door, front-door, or instrument. Tau equals Cov of z and y over Cov of z and x when the instrument is the strategy, and only if the first stage is not weak. Placebos are quiet. Pre-trends do not want the gap. Attribution is not used. Unnamed shocks are not plotted. Empty remains legal and is preferred to a footnote.

When an experiment is funded, the probe cell can speak. A Gaussian process with a named kernel and a named nugget updates. Expected improvement’s closed form proposes a location. A random probe fires every k steps. Length-scales are shown. Collapse is named as a broken cheap model. A loved corner is named as a habit. When the budget is gone, x-star is a probe, not a truth.

What is still refused is the mix of those speeches. A CPI print is not allowed to be a path and a shock and a probe in one breath. A neural attribution is not allowed to be a contrast. A stacking zero is not allowed to be hidden. A weak F is not allowed to be footnoted. A mean-week occupancy error is not allowed to be skill. A cash median is not allowed to stand in for a tail. A blended percent is not allowed to stand in for the table.

If those cells cannot be named, the stack is not in use. That sentence is the short paper’s ending and this note’s ending. The long form exists so the ending cannot be read as a taste. It is a protocol. Path, contrast, probe. Members, dissent, scores. Identification or empty. Placebos. Kernels. Collapse. Peaks, not mean weeks. Thirteen-week tails. Board speech without a blend.

133. Seasonal decomposer hygiene: breaks, calendars, and leftover scales

A seasonal decomposer that does not mark a calendar break will invent a season from the break. A quarter-end that moved, a reporting calendar that slipped, a holiday that landed on a Friday in one year and a Sunday in the next, are breaks. They are marked. They are not smoothed. Smoothing a break is how a decomposer becomes a story about a week that was a rule change. The mark is a dated dummy or a split of the seasonal block. The split is declared on the expanding window. It is not declared on the skill window after a pretty residual.

The leftover scale of the decomposer is estimated on the same expanding window as the seasonal block. A scale that is borrowed from the ARIMA–GARCH member is a borrowed scale. Borrowed scales are refused unless a later note names a shared-scale experiment and keeps the members otherwise separate. Shared scale is how two members become one member in costume. This paper’s members stay separate so dissent stays meaningful. If the decomposer and the GARCH member share a scale, a loss on scale cannot be attributed.

Weekly seasonality and yearly seasonality are different blocks when both exist. A cash series may have both a Friday and a quarter. An occupancy series may have both a Tuesday and a respiratory year. Fitting one block and calling it both is how a decomposer hides a lost week inside a pretty year. The board is allowed to see both blocks when both were named. A single seasonal stroke that claims to be both is a blend inside one member.

Missing dates are marked. Interpolating a missing Friday with a Thursday and a Monday is a story about a Friday that was not observed. The decomposer does not invent Fridays. A gap in occupancy during a reporting outage is a gap. Filling the gap with the mean week is how a peak disappears into a Tuesday. Peaks are the operating occupancy speech. Mean weeks are diagnostics.

The decomposer remains a path member. It does not become a contrast because a seasonal dummy sat near a policy date. It does not become a probe because a seasonal leftover was large. Large seasonal leftovers are dissent candidates. They are not instruments. They are not acquisition stars.

134. ARIMA orders, GARCH persistence, and the explosive-scale refusal

ARIMA orders are frozen before the quoted skill window. An information criterion computed on the expanding window may propose orders. The proposal is accepted or refused before the skill dates are seen. Searching orders on the skill dates is a shuffle. A later window that finds the orders stale re-declares them and moves the skill window later still. There is no private refit on the spoken block.

Differencing is a decision. Over-differencing a cash series that already lost its week can turn a short memory into a leftover that GARCH will try to scale. Under-differencing a series that still has a year can leave a season in the mean that belongs to the decomposer. This paper keeps the decomposer as a separate member so the ARIMA mean does not have to be the season. If a later desk drops the decomposer, the ARIMA mean will be asked to do two jobs and dissent will be less readable.

GARCH persistence is a warning. A pair that sits near the covariance-stationary boundary is marked near-boundary. A pair that sums through one is marked explosive and is dropped from the mix for that series and horizon. Dropping is visible. A dropped GARCH member is dissent at weight zero for a reason the board can name, which is different from a stacking zero that names only redundancy. The reason is spoken: explosive scale, not pretty CRPS.

The operating predictive is Gaussian with the ARIMA mean and the GARCH scale. CRPS uses the Gaussian closed form. A t-scale, if used, is declared before the skill window. An unnamed fat tail is a costume. Using GARCH to widen a cash tail after seeing the skill tail is a shuffle. Using GARCH to move the mean is a different product. Scale is a warning. Mean is a mean.

This member does not speak a contrast. A GARCH spike after a policy date is a path leftover. It is not an impulse. A GARCH spike is not a reason to empty or fill the contrast cell. Identification failures have their own gates. This member also does not speak a probe. A loud scale is not a budget.

135. Neural sleeve heads, dated events, and the costume test for features

A neural seasonal sleeve may carry a CRPS head, a Gaussian head, or pinball heads at named quantiles. The training loss is named. A head that is trained on pinball at one half and spoken as a full law is a costume. A head that is trained on CRPS and spoken only as a median is incomplete speech, not a costume, and the board should still show the law. Incomplete speech is repaired by drawing the band. Costume speech is refused.

Dated events enter as features that remain defined if the target is permuted. That is the costume test. A holiday dummy remains defined if cash is permuted. A dummy that is “the day the print was large” does not. A testing dummy remains defined if occupancy is permuted. A dummy that is “the day occupancy crossed the peak rule” does not. Features that fail the costume test are refused. They are the target in costume.

Retraining on the expanding window is allowed. Retraining on the skill window is not. Tuning depth, width, or dropout on the skill window is not. A later desk that wants a new architecture re-declares it and moves the skill window later. The September revision does not retune to fatten a later CRPS. Attribution that moves after a legal retrain is expected. Attribution that is used as a contrast is refused.

Incidence columns come from the compartment paper. They are dated. They are not refit inside the sleeve. Sentiment leftovers are dated and residualized against the obvious tape. Political polarity is a column, not a target. Equity leftovers are finished columns from the equity paper and are not rescored here. None of these columns becomes a parent in a cash tree or an occupancy tree.

The sleeve is reconciled or dropped. Adding-up is not optional because the sleeve is nonlinear. Nonlinearity is why MinT exists here. Attention is not MinT. A pretty attention roll over a quarter-end is a plot. Plots stay in the path cell.

137. Three cells that must not mix

A path, a contrast, and a next measurement are three products. Mixing them into one forecast is how a desk double-counts a CPI print. The July 2026 blended accuracy across asset classes is deleted. The September revision keeps the deletion.

Spoken, name the series and the horizon first. Equity residuals are scored on the residual-class protocol, not as a CRPS percent here. Incidence is a dated regressor from a compartment fit, not a refit SEIRD inside the ensemble.

Misread, a blended percent is a result of this paper. It is not. A single percent that covers cash, occupancy, and a residual class is the failure mode.

Refused: filling a contrast cell with a path, filling a path cell with an attribution plot, and hiding the dissent member.

Sentiment leftovers and political polarity may enter as columns. They do not replace the target. Drift in the ensemble residual is an alarm that the mix is stale, not a new causal graph.

138. The three path members

The seasonal decomposer owns a calendar. The ARIMA–GARCH member owns a short memory and a clustering variance. The neural seasonal sleeve is a competitor on the path and is reconciled or dropped when it breaks an adding-up identity.

Spoken, weights follow last-window CRPS. A member that was pretty on the training year and dead last year receives a small weight. The board shows that small weight.

Misread, a GARCH coefficient is an identification strategy. It is not. Attention is a reconciliation. It is not. The parent sees the reconciled identity, not the raw neural sum.

Refused: a print eaten twice as seasonal and as a shock, and an unreconciled prettier parent.

Hierarchical cash and hierarchical occupancy are different trees. They are not reconciled to each other. A number that claims to be both is the blended accuracy that was deleted.

139. Softmax dissent versus stacking zeros

Last-window softmax weights keep a small positive weight on the loser. Constrained stacking on a short board can put a zero on the loser. The four-date toy produces about 0.52, 0.36, 0.12 for softmax and 0.61, 0.39, 0.00 for stacking.

Spoken, the 0.12 is the dissent. The zero is what we refuse to hide. If a later desk wants stacking, the simplex constraint stays and the zero is shown.

Misread, negative weights are more honest because they cancel a loser. They hide who lost. That is why they are refused.

Refused: hiding a zero, using the skill window to set the weights, and treating the mix as a contrast.

The mix is not a back-door. Feature attribution on the neural member is a plot of that member. It does not move into the contrast cell by being plotted next to the mix.

140. CRPS in the loss convention

CRPS is the integral of squared distance between the forecast CDF and the indicator of the realization. The energy form, expected absolute deviation minus half an expected pairwise absolute deviation, is the same number, not a second score.

Spoken, the Gaussian closed form is used for the Gaussian member. The energy form is used for a sample ensemble. Mixing the two without naming which is a different product.

Misread, a point contest that a sharp wrong member wins is the same as CRPS. It is not. A wide honest member can win CRPS and lose a point contest. That is the point of CRPS on paths.

Refused: a reward-convention CRPS quoted as if it were this loss, and a blended CRPS across asset classes.

The score is strictly proper for laws with a finite first moment. The expected loss is uniquely minimized when the forecast law is the true law.

141. Pinball is not CRPS

Pinball is strictly consistent for a named quantile. We quote it on cash quantiles, including a tail quantile that is not the median. Median pinball is half the mean absolute error of the median. CRPS integrates Brier scores over all thresholds.

Spoken, a distribution can have a good median and a bad tail. The two scores will then disagree. That disagreement is information.

Misread, one number covers both. It does not. Member 3 on the four-date toy is worse on CRPS because its scale is wrong even when the median is not.

Refused: a pinball-optimal neural sleeve that violates adding-up and is explained with attention, and a single percent for pinball and CRPS.

A thirteen-week cash horizon must be named. A blended percent across thirteen weeks and across names is not a result.

142. Empty contrast cells

If the graph cannot be drawn without a cycle, the cell is empty. If no back-door set, no front-door mediator, and no instrument closes, the cell is empty. A weak first-stage F empties the cell. An instrument that is the target in costume empties the cell. A pre-trend that already wants the gap empties the cell. A placebo as large as the treated path empties the cell.

Spoken, empty is a result. A number with a footnote is not a result. Writing a do-operator in a notebook does not fill the cell.

Misread, attribution is a contrast. A pretty synthetic-control line without placebos is a contrast. They are not.

Refused: keeping a gap with a caveat, and a national series with no donor.

The defended graph is instrument, treatment, confounder, outcome. An unnamed shock is a residual. Plotting it as an impulse is a different product.

143. Back-door and front-door by name

A set W satisfies the back-door criterion when it blocks every back-door path from treatment to outcome and contains no descendant of treatment. The adjustment is then the mixture of the observational conditional over the law of W.

Spoken, an unblocked hidden common cause makes the functional not interventional. The front-door criterion uses a mediator that intercepts all directed paths and is unconfounded given treatment.

Misread, typing do is identification. It is not. We write the names of the criteria so a notebook cannot claim identification without one of them.

Refused: a contrast that names neither criterion nor an instrument, and a back-door set that contains a descendant of treatment.

Sharma and Kiciman are cited for the names of the criteria, not as a filled cell.

144. Weak instruments

The population IV functional is Cov(z,y) over Cov(z,x) when exclusion holds and the first stage is nonzero. In a sample the ratio is inner products after demeaning. When the first stage is weak, the denominator is a mean-zero form and the ratio does not have a Gaussian limit.

Spoken, Stock and Yogo tabulate critical values. A first-stage F below those values empties the cell. We do not footnote a weak F.

Misread, writing do repairs a weak F. It does not. The estimator is biased toward the OLS functional.

Refused: quoting a Cauchy-type ratio as tau, and the six-point independent-z draw as a number.

The printed six points give about 1.12 with a healthy denominator. That is a check, not a named shock.

145. Synthetic control gates

Spoken, if they are not quiet, the gap is deleted. A pretty line is a pretty line.

Misread, a single national series with no donor can still be synthetically controlled. It cannot. The method is refused before placebos are drawn.

Refused: fitting weights on a window that includes post-treatment outcomes, and quoting a gap without both placebo families.

A moving placebo as large as the treated path is the same as a pre-trend that already wants the gap: the cell is empty.

146. Acquisition as a probe

Expected improvement for a Gaussian posterior is (μ minus f-star) Φ(z) plus σ φ(z). The exploration term likes a higher variance. A second point with the same mean and a smaller variance can lose.

Spoken, the budget is stated. When it is gone, the current x-star is the answer. f-star is the best observed value, not a hope written to make improvement look large.

Misread, x-star is the truth. It is a probe. A Gaussian process that loves a corner of the box is probing the cheap model.

Refused: one more probe after the budget, a length-scale collapsed to the grid sold as a discovery, and replacing posterior variance with distance to the nearest point.

A random probe every k steps is a member of the search ensemble. When EI and UCB agree on a nonsense corner, the random member is the defence. UCB is the Srinivas–Krause–Kakade–Seeger choice with a stated β_t. EI is the default because it needs no extra sequence.

147. Kernels and collapsed length-scales

The default kernel is squared exponential plus a nugget. Matérn-5/2 is allowed when the cheap model should be less smooth. Length-scales have log-normal priors and are not set to make the acquisition look decisive.

Spoken, a length-scale that collapses to the grid spacing is a broken cheap model. The posterior variance is the Schur complement.

Misread, a heuristic distance to the nearest evaluated point is the posterior variance. It is not. The nugget is invertibility, not trust.

Refused: setting length-scales after seeing the acquisition field, and trusting a posterior that has collapsed.

A regime the donors do not contain will grow the dissent. The board shows the dissent. If the dissent is the only member that has seen this regime, the mix is stale and the pattern note should trip.

148. Occupancy peaks and later winters

MAE on occupancy is quoted on peaks, not on the mean week, on a later winter than the fit. A model that is good on Tuesdays and late on the first freeze looks skilled on a mean-week MAE and useless to a ward.

Spoken, incidence may enter as a dated regressor. A refit SEIRD inside the occupancy member is refused. The compartment paper owns the compartment.

Misread, a single occupancy percent is a thirty-day season. That speech is forbidden.

Refused: scoring occupancy on the training winter, and blending occupancy MAE with cash CRPS.

The later winter is the test. The winter it was trained on is not a test.

149. Cash paths at a named horizon

Cash paths use CRPS on the full predictive and pinball on named quantiles including a tail. A thirteen-week horizon must be named. The weight window is earlier than the skill window.

Spoken, using the skill window to set the weights is a shuffle. A neural sleeve that is beautiful at one half and dead at nine tenths has a bad tail. The tail is the information.

Misread, a blended percent across thirteen weeks and across names is a result. It is not.

Refused: a tail averaged away on a human board that still owns the path, and a weight window that includes the skill window.

On a human-owned board the overlay is to draw the tail. That overlay is not a CRPS weight and not a contrast.

150. Decision-board speech

Fit the three members on the expanding window. Form CRPS weights. Show the dissent. Draw the DAG. If no identification strategy closes, leave the contrast blank. If an experiment is funded, update the Gaussian process and take x-star from the named acquisition. Speak only the metric in the skill table for that object.

Spoken, once series and horizon are named, a path with a band and a dissent is the minimum speech. Silence in the contrast cell is allowed. Silence in the path cell is not after naming.

Misread, a board that speaks a blended percent is this paper. It is not. A board that fills contrast with attribution has left this paper.

Refused: hiding the dissent, and a fourth cell invented at the moment of a drift alarm.

The pattern note trips a stale mix. Decay toward equal and refit is the written action. This paper does not invent a fourth action.