# Enough to hand over? — model and verification

[Read the essai](handover.html) · [Earlier shuttle model](EVIDENCE_METHOD.md)

## Question and fixed comparison

Can an incoming organiser continue from the review number and current demand belief, or does a survey spending limit require one more state variable? Separately, can the same summary explain which evidence was obtained?

This is an authored extension of the existing fictional festival shuttle. It applies a known sequential-decision principle; it is not a discovery that the earlier model was incorrect. The six existing presets, terminal shuttle payoffs, information timing and conditionally independent observations are held fixed. The added contract allows total survey fees of at most twice the price of one survey. Ordinary waiting costs still count in the objective but do not use this allowance. Free surveys remain free, with no artificial two-survey count limit.

The fixed bank includes every feasible history before commitment, including histories a best policy would never choose. A decision state at the final review is included because commitment has not yet occurred. Histories are compared only within the same preset and contract; fixed model parameters remain known to the recipient. No distribution over earlier discretionary action choices is assumed, so history counts are not probabilities or population estimates.

The protocol compares full history with `(review, belief)`, then `(review, belief, remaining allowance)`. A retained history accompanies the last summary for the separate audit question: **which evidence was obtained, and what survey fees were already incurred?** This question is fixed before the run. The experiment does not test whether an arbitrary unforeseen question can be answered.

## State, choices and objective

Let `t` be the current review, `p` the posterior probability of high demand and `b` the unspent survey allowance. State under the changed contract is `(t,p,b)`, together with the preset's fixed parameters. Choices are commitment to one bus, three buses or no shuttle; waiting for a fresh booking observation; or buying a fresh survey if its fee fits and its result arrives by the final review.

The hidden high/low demand state remains fixed. Observations update `p` by Bayes' rule, advance `t` by their stated duration, and reduce `b` only when a survey fee is paid. A policy chooses using information already received. Expected terminal service value, less future survey and waiting costs, is maximised. Earlier costs are sunk in that continuation objective; remaining spending permission can still constrain future choices. Neither the policy nor the summary sees hidden demand.

The original model omits `b` because it has no cumulative survey cap. Its `(t,p)` state already supports a full sequential policy, including tests whose value is earned through later choices. This essai does not equate sufficient state with a label for the next action.

## What counts as a sufficient handover

For each group of histories sharing a summary, ask whether **one entire observation-contingent continuation policy** is feasible and optimal for every history in that group. All optimal ties count; a difference in an arbitrary tie-break is not a failure.

For the reduced `(t,p)` groups, the only hidden difference relevant to future control is remaining allowance. The feasible policy sets are nested: every policy affordable with the smallest allowance is affordable with a larger one. Future observations and payoffs conditional on a policy are otherwise identical. Therefore:

- the common feasible set is the set affordable at the smallest allowance;
- its best value is `V(t,p,min b)`;
- the least possible worst-case regret of a common feasible policy is `max_b V(t,p,b) − V(t,p,min b)`.

This checks complete contingent policies, not merely whether recommended next-action labels match. It allows the incoming organiser to carry the common plan forward and track subsequent evidence and spending. It does not demand that the organiser repeatedly forget their own actions. For a fixed exact allowance, all histories give identical future feasible actions, transitions and rewards. Retaining that allowance restores the state under this declared model; it is not claimed to be the unique or smallest possible summary.

The audit comparison is different. Even two histories with identical state and optimal continuation can report different observations or their order. Keeping a compact control state does not by itself license deleting the underlying record. The test concerns explicit evidence traces, not an organisational accountability system.

## Run protocol and checks

The implementation uses Python standard-library rational arithmetic: posterior collisions, value comparisons, feasibility and ties are exact. Decimal display values accompany exact fractions in the saved JSON. The run has no random sampling or parameter search.

The witness is selected by greatest exact worst-case regret, then a stable preset/group order. All groups and histories remain available in the downloadable results. The run also identifies histories reachable under any original optimal policy, including ties. Where a mixed group includes both optimal and non-optimal earlier histories, that does not establish ambiguity between two handovers both generated by an optimal policy.

Before interpreting the result, the following controls must hold:

1. Under the original contract, `(t,p)` preserves all continuation values.
2. With the allowance, `(t,p,b)` preserves all continuation values.
3. An independent enumerator of deterministic contingent policies agrees with the exact solver. It evaluates each policy conditional on high and low demand, then mixes those payoffs at the belief; it does not reuse the Bellman optimiser.
4. The six unconstrained roots reproduce the existing JavaScript model.
5. Rebuilding the results produces the checked-in file exactly.

A zero loss for all `(t,p)` groups would mean this bank had not shown a consequential need for the extra field. Failure of either state control would require debugging or a better state, not a positive interpretation. Absence of differing audit answers would leave the proposed trace distinction unsupported. The added cap is an assumption; its usefulness in a real programme and any benefit to a human reader remain untested.

## Sources and relation to other essais

Powell, *Sequential Decision Analytics and Modeling: Modeling with Python* (2022), §1.4, defines state through the information required for subsequent costs, decision constraints and transitions; physical/resource, other information and belief state have distinct roles. His [discussion of state variables](https://arxiv.org/abs/2002.06238) develops the distinction between policy-specific and optimisation definitions. The earlier shuttle essai also uses the information-dependent lookahead in §12.4.3. This implementation and its allowance example are independently written, not a reproduction of Powell's numerical experiments.

Three outward references locate the question without merging the experiments:

- [Knowing Enough, second essai](https://experimentsin.ai/knowing-enough/essai-02/): the available actions affect what must be distinguished for a choice.
- [Rotunda's open-interface test](https://experimentsin.ai/bricklaying-a-rotunda/apps/constructive-state/open-interface-method.md): preserving one declared behaviour is weaker than preserving everything an extended interface may need.
- [Enough for the next question](https://experimentsin.ai/picture-and-judgement/essai-03/): a shared verdict can conceal different reasons needed for a later question.

These are comparisons between bounded constructions, not evidence for one universal representation or a common end-to-end system.

## Result of the fixed run

Across the six presets there are 172 original histories and 164 histories that respect the survey allowance. The latter form 49 `(t,p)` groups. Nine groups hide different remaining allowances; only one requires different optimal continuations. Adding allowance makes 61 groups and leaves no continuation-value loss. The original contract has 51 `(t,p)` groups and no loss.

The consequential collision is in the three-review preset. Two opposite surveys return belief to one half at review 2 after spending the full one-point allowance. Two routine observations also leave belief at one half at review 2, but spend no survey allowance. With no survey funds, the best continuation reserves one bus for 15 points. With funds remaining, a survey followed by a demand-contingent reservation earns 19.5 points after its fee and waiting cost. The best common feasible policy earns 15: maximum regret is 4.5 points. This is a continuation-value comparison; it does not compare total performance of the two earlier histories.

The waiting history is feasible but not chosen by an original optimal policy. Restricting handovers to histories known to have arisen under an original optimal policy leaves **zero harmful groups**. That limits the claim: the witness concerns handovers after different admissible ways of working, not a demonstrated ambiguity between two original-optimal paths. The bank and allowance were fixed before the run; this additional subset interpretation was recorded after inspecting the initial result.

The allowance itself lowers the three-review preset's optimal starting value from 21.44 to 20.05; the other five roots are unchanged. That 1.39-point loss is the effect of a tighter resource constraint, distinct from the 4.5-point worst-case handover loss within the mixed history group.

The augmented state determines past survey spending because the initial allowance is known. It still cannot recover the ordered observation trace in 31 of its 61 groups. For example, positive-then-negative and negative-then-positive surveys have the same final belief and spending. A retained record answers the trace question; it does not improve the optimal continuation value once the state is sufficient.

## Reproduce

Python 3.10 or later, standard library only:

```text
python3 scripts/handover_experiment.py
python3 scripts/handover_experiment.py --check
python3 tests/handover-independent.py
```

The first command regenerates `handover-results.json`; `--check` verifies that the saved result is current without writing it. `pnpm run test:handover` additionally compares all 172 unconstrained history values with the existing JavaScript solver. `pnpm test` includes this suite alongside the earlier model and interface tests; `pnpm run build:check` verifies the unchanged generated companion applications.

Serve the repository over HTTP and open `handover.html` to inspect the saved run. The browser does not rerun the Python experiment or claim to optimise arbitrary new inputs; its controls explore the recorded contracts, histories and contingent policies.
