# When reassurance misleads

21 October 2026 · one new essai in the fictional mountain-festival shuttle scenario.

## Question and contribution

A dashboard can become reassuring because a resource genuinely improves, because its reporting target moves, or because adverse updates stop arriving. This essai makes those mechanisms explicit, then asks what the organiser should do next: commit three buses, commit one bus, cancel, or buy one independent check before deciding.

The opening examples share a zero displayed gap and the label **reassuring**. Their raw scores, targets, received observations, physical distributions and contractual commitments are shown separately. They are not three identical histories, nor evidence that equal dashboard summaries imply equal situations. Continuation decisions use the full declared information state rather than that summary.

## Sources and authorship

- The original [foray prompt](source/foray_prompt.md) separates scheduled work, management decisions and uncertainty-reduction effort. The existing [evidence method](EVIDENCE_METHOD.md) supplies the festival-shuttle setting and demand-dependent service values. The original climate, evidence and handover models remain separate.
- Warren Powell's [*On State Variables, Bandit Problems and POMDPs*](https://arxiv.org/abs/2002.06238) (2020) discusses state representation, partial observability and decisions that affect what is observed and learned. It supplies a useful organising framework for this enquiry; it does not supply the shuttle mechanisms, numerical assumptions or results below.
- [Gimmer Crag](https://github.com/lawrencerowland/gimmer-crag) applications 4 and 6 provide the starting contrasts: reference changes can improve displayed discrepancy without physical progress; actions can affect resources, commitments and the information subsequently received. The source audit found that application 6 did not establish its claimed persistent commitment-induced reduction in learning. This successor therefore declares and implements a specific action-dependent reporting rule, rather than inheriting that claim.

Powell is the flexible, integrative spine: state, decision, information, transition and objective make the components inspectable together. The particular causal mechanisms, likelihoods, contracts, payoff extension and finite solution method are ours. Other theories may suggest better mechanisms or richer representations that fit this spine. This is not a claim that the framework alone completes the substantive modelling or provides a general psychological explanation.

## State, actions, information, transition and objective

| Component | Implemented meaning |
|---|---|
| State | Review time `t`; the full four-state probability distribution `b`; public commitment `open` or `reserved-three`; whether the one independent check has been used; and fixed model settings. Reporting target and retained status are separate public display/history fields. |
| Decision | First choose one declared mechanism. After its source report, choose a terminal shuttle commitment or one feasible independent check. The implementation compares initial mechanisms; it does not optimise a larger sequence of initial interventions. |
| Information | A noisy initial readiness report, possibly withheld by the selected contract; then, if purchased, a noisy two-part independent check of demand and readiness. Missingness is an observation when it is caused by the selected reporting rule. |
| Transition | A real backup intervention may convert unready to ready; a target change affects only the display reference; a provisional reservation changes commitment and the report channel. A received observation updates `b` by Bayes' rule. A final decision locks three buses, locks one bus, or leaves no shuttle. |
| Objective | Maximise expected terminal net service value, less the initial intervention charge, any purchased check charge, and any reservation exit charge. The original objective stays fixed when the reporting target changes. |

The latent states, in code order, are `HR` (high demand, ready), `HU` (high, unready), `LR` (low, ready), and `LU` (low, unready). Readiness refers to the additional capacity needed for the three-bus service. One bus is assumed available regardless of that readiness variable.

The initial prior factorises: if `h=P(high)` and `r=P(ready)`, then `b=[hr,h(1-r),(1-h)r,(1-h)(1-r)]`. Defaults are `h=.75`, `r=.5`. The solver nevertheless retains a full four-element belief and accepts correlated beliefs; it does not replace the joint distribution with two marginals. Demand remains fixed. After the initial intervention, readiness also remains fixed throughout the short decision window.

No realised hidden state or future signal enters policy selection. The solver accepts belief, public commitment, time, check-used status and the known model. World-specific values appear only in integration over possible worlds and in the separate policy evaluator. “Actual value” in the interface means expectation under the declared data-generating law, not observed transport delivery or human benefit.

## Three initial mechanisms

All initial actions and their source reports are treated as completed at review `0`, before the next decision. They do not consume review time in this fixture. Each begins with a retained green display whose historical origin has no diagnostic evidential weight.

1. **Real resource improvement.** Pay `improvementCost` (default 6). A ready supplier stays ready; an unready supplier becomes ready with probability `improvementRate` (default `.8`). Demand is unchanged. The physical transition matrix, with rows and columns in `HR,HU,LR,LU` order, is `[[1,0,0,0],[a,1-a,0,0],[0,0,1,0],[0,0,a,1-a]]`. The report channel remains open and the target stays at 1. This is a probabilistic physical intervention, not a guarantee of successful improvement. At defaults it raises pre-report readiness probability from `.5` to `.9`.
2. **Lower the reported target.** Physical state, evidence channel and contractual commitment are unchanged. The reporting target changes from 1 to 0, with organiser authority recorded in the history; the original target remains 1. An adverse score of 0 now meets the reported target. There is no charge for this display-only action in the fixture. The same adverse receipt and belief yield exactly the same continuation values under either reference.
3. **Reserve, then withhold adverse updates.** Pay `reservationCost` (default 2), set public commitment to `reserved-three`, and switch the portal to favorable-only updates. A favorable raw report is released. An adverse raw report produces no update by the review; the dashboard retains its preceding green score. Physical readiness does not improve. Moving from the provisional three-bus reservation to one bus or no shuttle later incurs `exitCost` (default 12). Final commitment to three buses incurs no exit charge.

The third mechanism is an authored institutional rule. It is not a claim that real suppliers necessarily withhold adverse information after reservation, nor that commitment automatically changes beliefs. The provisional reservation and its reporting rule are a declared package in this experiment. The fee and exit liability expose its costs; any practical benefit of a reservation beyond those declared effects is outside this fixture.

## Observation laws and belief updates

For the open report channel, let `q=reportQuality` (default `.8`). A positive raw readiness report has likelihood `q` when ready and `1-q` when unready. A negative report has the complementary likelihood. Reports carry no direct demand information. The raw reporting process is the same before and after selecting the filtering contract; only release changes:

| Received outcome | Open channel: ready / unready | Filtered channel: ready / unready |
|---|---:|---:|
| Positive | `q / (1-q)` | `q / (1-q)` |
| Negative | `(1-q) / q` | `0 / 0` |
| Missing | `0 / 0` | `(1-q) / q` |

For every received observation `o`, `P(o)=sum_w b(w)L(o|w)` and `b'(w)=b(w)L(o|w)/P(o)`. Zero-probability outcomes are omitted; no posterior is invented for them. Under the filtering contract a missing update carries the same likelihood as the withheld adverse raw report. Calling it reassuring on the retained display does not remove that evidence.

The independent check returns one of four pairs: high/ready, high/unready, low/ready, or low/unready. Each bit has accuracy `checkQuality` (default `.85`). The two bits and the earlier raw report are conditionally independent given the post-intervention world. For a world `w` and reported pair `o`, the check likelihood is the product of the two correct-bit or incorrect-bit probabilities. These are modelling assumptions, not empirical calibration. The check changes information only; it cannot repair readiness, remove an exit liability, or undo a final commitment.

## Payoffs, charges and time boundary

The terminal values below include the illustrative service benefit and ordinary hire cost, in arbitrary net value points. The additional high-demand/unready penalty extends the earlier two-state shuttle fixture. It is authored for this essai.

| Terminal action | High, ready | High, unready | Low, ready | Low, unready |
|---|---:|---:|---:|---:|
| Commit three buses | 60 | -50 | -40 | -40 |
| Commit one bus | 25 | 25 | 5 | 5 |
| No shuttle / cancel | 0 | 0 | 0 | 0 |

Every path subtracts the initial mechanism cost once, the independent-check fee once if that check is purchased, and the exit charge once if a provisional reservation ends in one bus or cancellation. A purchased check costs `evidenceCost` (default 4). No cost is charged for an unavailable or unchosen check. Initial fees are sunk at the continuation decision and are included again only as an accounting subtraction from the continuation value, not as a repeated charge on each node.

These are additive value charges, not cash balances or an affordability constraint. There is no waiting charge, discounting, refund, shortage budget, parallel survey, or further repair action. No parameter for such a quantity is silently implied by the interface.

One review represents one day. A check is feasible exactly when it has not already been used and `t + latency <= deadline`. Default latency and deadline are both 1. A result arriving on the deadline is usable; a later result is excluded before purchase. The supported deadline is an integer from 0 to 4; latency is an integer from 1 to 4. At most one check can be bought, even when more calendar time remains. There are no interim signals or decisions while it is pending and no separate wait action.

A terminal action ends the model. Cancel is implemented as a real terminal transition to no commitment, including the reservation exit charge when applicable. It does not reopen a later decision. One bus strictly dominates cancellation under this payoff table because both incur the same applicable exit charge and one bus has positive service value in every world. Cancellation is consequently an inspectable comparator, not an optimally selected outcome in this fixture.

## Finite policy calculation and fair baseline

At each continuation state the solver compares the three terminal values with `-check fee + sum_o P(o) max_a E[terminal value(a) - exit charge(a) | o]`, when the one check is feasible. This exhausts the specified decision horizon: without a check there are 3 deterministic policies; with a check there are those 3 plus `3^4=81` mappings from the four check receipts to a terminal choice. “Exact” means finite enumeration subject to JavaScript floating-point arithmetic. A `1e-10` tolerance breaks ties in favour of one bus, then three buses, cancellation, and the check.

The comparator is a **missing-at-random baseline**, not a display-following caricature. It uses the same prior, physical transition, payoff table, known reservation cost, exit charge, check accuracy, deadline and exact optimiser. For the filtered initial report only, it mistakenly assumes a normalized process in which half the reports are absent independently of readiness, and the other half follow the ordinary report channel: `L(missing|w)=.5`, `L(positive|w)=.5 L_open(positive|w)`, `L(negative|w)=.5 L_open(negative|w)`. Thus missingness leaves its prior unchanged. Its positive-report posterior agrees with the action-aware posterior. The fixed `.5` missing rate matches the default overall absence rate; it is not recalibrated as settings change.

After each actual source receipt, each policy is chosen using its own belief. Both fixed policies are then evaluated under the **same actual conditional world distribution and independent-check law**. The baseline does not get to assess its own success using its mistaken missingness model. Ex-ante values weight every feasible source history by its actual probability and subtract the initial cost. The highlighted witnesses are conditional teaching cases, not an estimate based on selecting only favorable histories.

The action-aware policy cannot have lower expected value than another feasible policy under the exact, correctly specified finite model, apart from numerical tolerance. The interest is in the concrete changed action and cost, not that mathematical inequality. Real calibration uncertainty or an incorrectly specified action-aware model could reverse an empirical comparison.

## Reproducible witnesses

At the default settings all three highlighted histories have displayed gap zero:

| History | Posterior readiness | Action-aware next action | Total expected value, including initial charge |
|---|---:|---|---:|
| Improvement, then positive report | `36/37`, about 97.3% | Check | 28.729730 |
| Lowered target, then negative report | 20% | One bus | 20.000000 |
| Reservation, then missing report | 20% | One bus, paying exit charge | 6.000000 |

For the missing-report witness, the baseline keeps readiness at 50% and buys a check. Its actual continuation value is `3.91525`, or `1.91525` after the initial reservation fee. The action-aware continuation value is 8, or 6 after that fee: an advantage of `4.08475` points for this conditional history. The baseline's own assumed continuation value is `13.103125`; it is labelled separately from evaluation under the actual law.

Across all initial report outcomes, default total values are `26.421625` for improvement, `24.270000` for changing the target, and `13.145500` for reservation/filtering under the action-aware policy. The baseline's corresponding filtered value is `11.103125`. These comparisons include the mechanisms' different fees and contractual effects; they are not isolated estimates of a single causal coefficient. The reference-only counterfactual and the report likelihood table provide the narrower invariance and channel checks.

The positive control is not anti-commitment: increasing the check fee to 6 makes both policies commit three buses immediately after the default improvement/positive-report witness. A genuine resource improvement can therefore be accepted without further evidence when the additional check is not worth its cost.

## Verification and limits

`node tests/reassurance-model.cjs` was independently rerun unchanged on 21 October 2026. It passed 1,458 decision states, 83,106 independently evaluated complete policies and 196 feasible source histories. The policy evaluator in the test restates the declared payoff and signal laws and sums outcomes directly; it does not call the production Bellman solver, transition/filter functions or policy evaluator to obtain expected policy values.

Checks cover every feasible root choice in the finite test bank; Bayes examples; normalized actual and baseline report laws; posterior martingale identities; genuine physical transitions; reference-only invariance; the positive control; baseline evaluation under the actual law; probability and fee conservation; inclusive deadline and late exclusion; terminal cancellation/recourse; rejection of hidden/future fields; deterministic replay; and browser-module loading. The check covers varied and degenerate beliefs and accuracies, not every point in the continuous parameter space. No material model defect was found in the independent review. Presentation and browser checks are reported separately from this model verification.

The result establishes a bounded authored calculation and a reproducible teaching contrast. It does not establish that practitioners understand the contrast, change decisions, deliver a better shuttle service, or transfer the reasoning to a real project. No human-user outcome has been measured. Demand dynamics, uncertain reporting rules, strategic agents, correlated checks, cash constraints, multiple interventions, legitimate changes to objectives, and richer theories of commitment or learning would require explicit extensions and fresh verification.
