# What was known at the time?

This essai inspects the information boundary of a fictional festival-shuttle review. It distinguishes an event, its observation, its arrival as a report, and the reference moment for a decision. It does not calculate a dispatch policy, estimate a psychological effect, or choose an operational reporting cadence.

## Governing framework

Warren B. Powell’s [*On State Variables, Bandit Problems and POMDPs*](https://arxiv.org/abs/2002.06238) supplies the state-based sequential-decision framing. The example makes the following distinctions within it:

| Element | Construction in this essai |
|---|---|
| State | The selected decision reference, received reports with event/observation/arrival times, and available event records. A configured freshness horizon is carried with the example. |
| Action | A review or a request for further evidence can be recorded. Drafting a request does not send it or record that it happened. |
| New information | A report becomes available at its arrival time. Its content still describes its event time. |
| Transition | Advancing the reference can admit later arrivals and make earlier events older. Editing or generating a trace is scenario authoring, not an operational transition. |
| Objective | Make a service decision using appropriate evidence. This page does not specify utility, action consequences or an optimisation problem; it examines a prerequisite information boundary. |

The readiness reports, timing rules and review flags are the author’s construction. They are not numerical results or an example attributed to Powell.

## Time and availability contract

All times are whole minutes after midnight on one fictional day, from 0 to 1439. The page does not read the computer clock, timezone or current date. There is no midnight rollover.

For each report, the validated order is:

`eventAt <= observedAt <= arrivedAt`

At reference `r`, a report is **not received** if `arrivedAt > r`. Otherwise its event age is `r - eventAt`; it is **fresh enough** when this age is at most horizon `h`, inclusive, and **older than horizon** otherwise. Arrival age, `r - arrivedAt`, is a separate displayed quantity. Observation latency and transport latency remain recoverable from the trace.

An arbitrarily broad horizon does not admit a report before arrival. A recent arrival does not refresh the event it describes. A common horizon between 5 and 120 minutes is deliberately simple; no claim is made that one horizon is suitable for every signal. Freshness neither establishes reliability nor ensures that the underlying situation has stayed the same.

The implementation object named `state` retains the full invented trace so that the reader can explore different references. It is a scenario-authoring container, not itself a leakage-safe decision state. `derive()` attaches availability labels to that full trace; its rows still contain later values. A future policy must explicitly project to received records rather than taking this object as its information input. Future actual and signal values are withheld in the decision panels. Future timestamps remain visible to explain why reports are unavailable. The explicitly labelled task-authoring form and downloaded full trace can reveal later values; neither is presented as the organiser’s information at the earlier reference. Editing replaces a synthetic task report rather than appending a revision history.

## Reports, flags and operations

- Twelve snapshots are fifteen minutes apart, from 09:15 to 12:00. Planned and actual readiness within a snapshot refer to its event time. Snapshots are independent invented summary reports; they are not computed as an average of task reports.
- Six tasks retain planned and actual percentage progress. Their gap is actual minus planned in **percentage points**, not relative percentage error. Both refer to the task’s event time. No projection or interpolation to the selected reference is attempted.
- Six signals retain value, unit and the three report clocks. Their values have no automatic operating limits in this example.
- The banner asks for a review if any eligible snapshot has an absolute gap of at least 20 percentage points, or any eligible task is more than 15 points behind plan. These counts are displayed separately and are not summed as independent evidence. Otherwise, stale or missing task/signal reports prompt a coverage check. Absence of flags is not a claim of safety or readiness.
- The seed gives reproducible generated reports. Generate replaces snapshots, tasks and signals while preserving reference, horizon and existing event records. Events are standalone records, not assertions about the newly generated sample. Reset restores the complete initial example.
- Slower arrival is an explicit counterfactual: each signal’s arrival increases by a seeded 5–30 minutes, capped at 23:59. Its value, event time and observation time do not change. Repeated trials accumulate delay. This is not a simulated refresh request or a new measurement.
- All state lives in page memory. The JSON download retains the full trace, chosen reference and horizon, provenance labels and derived classifications. Reloading resets the page.

## Event records

Seeded records are labelled **invented example**. Added records are labelled **reader-entered record**. Each record carries context, an action, action time, an outcome when observed, outcome time, severity and `recordedAt`.

The reference at explicit submission becomes `recordedAt`. The action must not be later than the reference. For an observed outcome, `actionAt <= outcomeAt <= recordedAt` is enforced. The entered whole-minute delay determines outcome time from action time. Zero delay is permitted. Pending records have no outcome time and no invented delay. Records whose `recordedAt` is later than the inspected reference are hidden from that decision view.

Opening a task or signal does not append an event. A signal can prepare a contextual draft, but the required action field is left empty: the reader must describe an action before adding it. No request is sent outside the page. Records are not silently capped or discarded.

The delay/horizon comparison is descriptive only. A short action–outcome interval does not establish causation; a long interval does not refute it. The record does not attribute a later result to the action.

## Supporting temporal-perception literature

This is a separate research line, subordinate to the state/information framing above:

1. Eagleman, D. M. (2008). [*Human time perception and its illusions*](https://doi.org/10.1016/j.conb.2008.06.002), *Current Opinion in Neurobiology* 18(2), 131–136. The review distinguishes aspects of experienced timing and restricts itself to subsecond perception. It does not justify carrying a perceptual window over to project timescales.
2. Cai, M., Stetson, C., and Eagleman, D. M. (2012). [*A neural model for temporal order judgments and their active recalibration: a common mechanism for space and time?*](https://doi.org/10.3389/fpsyg.2012.00470), *Frontiers in Psychology* 3:470. Its treatment of temporal recalibration distinguishes it from an intentional-binding explanation. This is not evidence that a project team shares that neural mechanism.

The useful design question is whether distinct time labels have been collapsed. The papers do not establish this page’s reporting thresholds, alert rules, causal judgements or any improvement in project performance.

## Verification and limits

`tests/temporal-observations-model.cjs` checks reproducibility, clock ordering, inclusive boundaries, the received-before-freshness gate, 648 independently classified time combinations, horizon monotonicity, arrival-only lag, task validation, explicit event ordering, pending/zero-delay outcomes, generation and export contents. Its whole-day format/parse round trip is a representation check, not 1,440 independent operational cases.

`tests/temporal-observations-ui.cjs` exercises the page through jsdom: all retained report sets, reference and horizon changes, withheld later values, keyboard tabs, dialog focus and cancellation, explicit authoring boundaries, task edits, draft-versus-record separation, invalid/future outcomes, safe rendering of text, lag, generation, download and reset.

These are bounded checks of the implemented rules. They do not prove state sufficiency for a real transport decision, validate the supplied reports, model correlated errors, establish cause and effect, or measure user benefit.
