# How old is the picture you act on?

Model version 1.0.0. A sequential-decision essai in Powell’s framework, using a fictional mountain festival shuttle. Every number below is authored. Exposure indices are dimensionless variables, not estimated accident probabilities. One tick is an abstract interval, not a calibrated day or minute.

## The decision problem

The organiser is preparing six serial work packages. Weather, route condition, fuel, fatigue and two exposure indices evolve while noisy reports arrive late. The organiser can wait or review a ranked queue and commit preset responses. Actions take time; some interrupt progress and all incur illustrative costs. The experiment compares a selected decision delay with zero delay while holding the report delay, external draws, other assumptions and policy fixed.

Powell’s state–decision–information–transition–objective framework governs this construction. The cognitive research at the end is an optional source of questions; it does not supply the queue, numerical weights, transition equations, objective or claimed effectiveness.

## State and information boundaries

The simulator holds a **hidden authored world**, a **current organiser estimate**, timestamped forecast history, report packets, pending actions, executed effects, committed costs, a received-report history and optional episode captures. The policy reads only received discrepancies and known commitments. The display reveals hidden truth only in the retrospective paired evaluation, explicitly labelled as such.

Initial state:

| Variable | Initial value |
|---|---:|
| Weather exposure | 0.35 |
| Route readiness | 0.92 |
| Fuel stock | 1.00 |
| Crew fatigue | 0.15 |
| Travel exposure index | 0.18 |
| Disruption index | 0.22 |
| Six package completions | All 0 |
| Reduced-pace interval end | −1 |

All six scalar state variables and package completion values are clipped to [0,1]. The model initially shares the known starting state. Subsequent model predictions omit external random disturbances and use somewhat gentler fatigue and route deterioration coefficients. The forecast is maintained **at the current review time**, not literally ahead of the world.

The packages are: inspect route; fuel and prepare vehicles; position vehicles and crew; trial shuttle; open boarding points; complete launch check. Base durations are [2,3,3,2,2,1]. Only the first incomplete package can advance in an interval. The world and estimate apply this rule separately. Blended completion estimates can therefore revise a previously estimated completion; they are not an authoritative physical completion certificate.

## Clock and transition order

At initialisation a single noisy packet is sampled at tick 0. It arrives at `0 + reportDelay`. There are no fabricated negative-time samples or repeated placeholder corrections. With zero report delay it is processed immediately.

For each later review t:

1. Evolve the world and estimate from t−1 to t, respecting any reduced-pace periods already in force.
2. Execute actions whose declared effect time is t, applying their deterministic effects to both world and estimate. These effects have phase `before-sample`.
3. Apply any explicitly scheduled external disruption to the world only.
4. Store the forecast for t and sample a packet with `sampledAt=t`, `arrivesAt=t+reportDelay`.
5. Deliver the packet due at t, if one exists. Compare it with the revised prior for **its sample time**, not the current estimate. Earlier packets may have updated that historical prior by replay; it is not necessarily the original forecast held at the historical review. Store that pre-correction comparison permanently in the received history.
6. Blend the packet into the estimate at its own timestamp. Replay expected transitions and known action effects forward to t. Every packet is processed once.
7. If automatic review is enabled, review a newly received packet when its scaled mismatch exceeds 0.62. Otherwise the reader may review manually. Zero-delay actions take effect now, **after the sample and report processing**, with phase `after-sample`.
8. Record the current state for inspection.

Manual surprises occur after the current sample. They neither rewrite that sample nor directly update the estimate. Their first possible evidence is the next sampled packet, received after its own delay. The reader sees the intervention control and log; the policy still learns only through the simulated report stream.

A chosen action records `chosenAt`, `decisionAt = chosenAt + decisionDelay`, and `effectiveAt`. Fuel adds `supplyDelay` to the latter. Other actions take effect at `decisionAt`. Costs are committed once, when chosen. Existing action times do not move when settings change: applying settings starts a fresh run.

### Retrospective correction with known effects

Stored prediction history is revisable internal model state. Original sampled packets, received pre-correction comparisons, action choices, review-time history snapshots and captured episodes are retained separately and are not rewritten. The table therefore labels the comparator as a prior for sample time, rather than an original historical forecast.

At a report’s sample time s, the stored estimate already contains effects executed **before that sample**. After blending, replay applies any known `after-sample` effects at s. For each following interval, replay runs the expected transition, applies `before-sample` effects, stores the revised estimate for that sample time, then applies `after-sample` effects. This preserves both immediate and scheduled action causality when an old report arrives.

All channels have one fixed delay and packets remain ordered. The implementation is a small fixed-lag smoother for this ordered stream, not a general solution to missing or out-of-order asynchronous sensor data. It assumes action effects are known exactly.

## Authored world and forecast dynamics

Let w be weather, f fatigue, r route readiness, u fuel, v travel exposure and q disruption. All updates below use the newly updated values when they precede another equation. `clip` limits a scalar to [0,1]. The displayed model source is the executable specification.

* Weather: `w ← clip(w + noise − k(w−0.35))`, with k=0.03 in the world and 0.02 in the forecast. World noise standard deviation is 0.05; forecast noise is zero.
* Pace: `clip(1−0.45w−0.35f−0.30(1−r)−0.35(1−u), 0.25, 1.05)`.
* Fatigue: add `k×pace−0.03(1−pace)`, where k=0.05 in the world and 0.04 in the forecast. Subtract 0.025 during a reduced-pace interval. World noise standard deviation is 0.01.
* Route: subtract `k×w + j×q`, with world k=0.02 and j=0.015; forecast k=0.015 and j=0.008. World noise standard deviation is 0.005.
* Fuel: add `clip(0.09−0.006×supplyDelay, 0.01, 0.10)−0.03×pace`. World noise standard deviation is 0.02. Supply lead time therefore affects both routine replenishment and the arrival of an extra chosen load; it is deliberately not a pure delay-only experiment.
* Hazard speed is `clip(1.2−hazardTime/10, 0.35, 1.2)`. Lower hazard timescale means faster drift; the minimum speed plateau is explicit.
* Travel exposure adds `speed×(0.08w+0.06f−0.04(r−0.8))`; disruption adds `speed×(0.06w+0.03f−0.03(r−0.8))`. Each receives world noise of standard deviation 0.01. The forecast uses the same drift but no noise.
* Recompute pace after those updates. Progress rate is `clip(pace×sensitivity×(1−0.20v−0.10q)/baseDuration, 0.02, 0.55)`, halved during a reduced-pace interval. Package sensitivities are [0.80,0.95,1,0.90,1,1].
* In the world only, a keyed uniform draw below `0.04×fatigue` removes `0.06×fatigue` completion from the active package. This is an authored rework rule, not a measured error probability.

The process may continue after all preparation packages complete. This is a finite-horizon preparation illustration, not a live shuttle schedule, passenger-flow simulation, engineering feasibility check or service operating plan.

### Reports and blend

Reports contain all six scalar variables plus all six completion values. Scalar sensor noise standard deviations are weather 0.05, route 0.03, fuel 0.06, fatigue 0.04, travel exposure 0.04 and disruption 0.04. Completion reporting has no added noise. All reported scalars are clipped to [0,1].

The correction is `corrected = prior + alpha × (reported−prior)` at the packet’s sample time. Alpha is 0.1125 for weather, 0.0875 for route/fuel/fatigue, 0.089375 for each exposure/disruption index, and 0.18 for completion. These fixed weights preserve the prototype’s weak blended-correction idea; they are neither Bayesian posterior weights nor fitted neural learning rates.

## Queue and response policy

For six scalar signals and the active package selected from **received completion**, compute absolute error against the forecast for that same sample time. No hidden-world dependency is used to select the active package. If all reported packages are complete, the final package is used.

Signal weights are weather 0.9, route 1.3, fuel 1.0, fatigue 0.9, travel exposure 1.35, disruption 1.2 and completion 1.1. Each is multiplied by a declared package coupling factor:

* Route and travel exposure: `clip(0.9+0.12×stage, 0.9, 1.5)`.
* Disruption: `clip(0.95+0.08×stage, 0.95, 1.35)`.
* Weather: `clip(0.95+0.10×stage, 0.95, 1.40)`.
* Fuel: 1.25 during package 2 or 3, otherwise 0.95.
* Fatigue and completion: 1.

Stage is zero-based. The queue shows the four largest `absoluteError × weight × coupling` scores. The moving scene ring points to the first. Scaled mismatch is `clip(1.55 × sum of all seven scores)`. The queue is an explicit display selection rule, not a validated attentional mechanism or a complete risk register.

Within those four selected signals, the response rules are:

| Response | Adverse discrepancy required | Effect | Cost |
|---|---|---|---:|
| Pause and inspect route | Route report >0.10 below forecast | Route +0.12; fatigue +0.03; half pace for next two intervals | 3 |
| Extra fuel load | Fuel report >0.12 below forecast | Fuel +0.18 after decision delay plus supply lead | 4 |
| Sheltered preparation | Weather or travel exposure above forecast and score >0.16 | Fatigue −0.07; travel exposure −0.05; half pace for next two intervals | 2 |
| Crew rest | Fatigue report >0.12 above forecast | Fatigue −0.15; half pace for next two intervals | 2 |
| Service contingency | Disruption report >0.15 above forecast | Disruption −0.10; half pace for next interval | 3 |

The policy takes at most two distinct eligible responses, in queue order. An already-pending response of the same type is excluded. Each packet can be reviewed only once. Two different packets may still lead to repeated responses after an earlier one executes: this simple rule does not infer whether repeated old news belongs to one persistent cause. It can overreact, spend too much or interrupt work unnecessarily. A high-risk state already predicted accurately may generate little error and no response. This is why the experiment does not call the rule optimal.

Manual review has no overall mismatch gate; it still requires the directional thresholds. Automatic review additionally requires a newly received packet with scaled mismatch >0.62. It is deterministic; the original random auto-triage gate is removed.

### The displayed review-window heuristic

The inherited window idea is retained as `clip(4+0.8×reportDelay+0.6×hazardTime−0.4×decisionDelay, 1, 8)`. Unlike the previous prototype, its direction now agrees with the hazard-timescale control: faster hazards narrow it. It is **display-only** and does not govern the policy. It is not an estimate of a human temporal binding window or an evidence-based recommended meeting cadence.

## Objective and paired comparison

At horizon 24, score:

`100 × mean(package completion) − 25 × max(travel exposure, disruption) − committed action costs`.

Action costs include pending commitments at the horizon; future benefits are not credited. This boundary is intentional and inspectable. These points express hypothetical preferences, not money, transport benefits or validated safety results.

The paired comparison starts two fresh runs. It holds the applied seed, sensor delay, hazard timescale, supply lead and automatic policy fixed; one arm uses the selected decision delay, the other zero. Both receive a weather-front intervention at tick 4, route disruption at 8 and fuel disruption at 12, all before sampling. The displayed manual run is unaffected.

External draws are keyed by **seed, tick and channel**, so extra decisions cannot consume or shift the other arm’s random stream. Noise draws and external interventions match. Actual observations need not match: actions change the worlds being observed. The random rework threshold also depends on current fatigue, so identical uniform draws need not cause identical rework events after paths diverge.

For the default seed 2718, report delay 2 and decision delay 2, the comparison produces:

| | Selected delay 2 | Zero delay |
|---|---:|---:|
| Committed action costs | 38 | 46 |
| Final objective | −13.0000 | −13.5480 |

Zero delay scores about 0.55 points lower in this constructed path. The result is a useful warning against claiming that less delay must help a simple intervention rule. A single paired seed is not a statistical estimate, and the example does not search over policies or calibrate real project settings.

## Inspectable records and controls

Run/pause, single step, surprise, restart, manual review, settings presets, capture and JSON export are available. Keyboard shortcuts outside interactive controls: Space steps; P runs/pauses; S introduces a surprise; R restarts. Input/button focus is respected. Automatic playback stops at 120 intervals; no background computation is implied after the page is closed.

The JSON export includes applied settings, every original received report and pre-correction comparison, event times and phases, pending and executed actions, current-history snapshots, captures, final score and the full paired runs if calculated. Captured details are inspectable on the page as frozen tables and action lists as well as in the JSON download. Captures distinguish **context and known commitments** from effects already executed. They do not label an observed state as a causal action outcome.

## Supporting research and boundaries

* [Powell, On State Variables, Bandit Problems and POMDPs](https://arxiv.org/abs/2002.06238) provides the state-based sequential-decision framework. This essai supplies its own policy, dynamics and objective. It does not reduce Powell’s general framework to reinforcement learning.
* [Cai, Stetson & Eagleman (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) distinguishes the modelled temporal recalibration from an intentional-binding explanation. This is a boundary for the analogy, not evidence validating a project queue or fixed correction weight.
* [Eagleman (2008), Prediction and postdiction: Two frameworks with the goal of delay compensation](https://doi.org/10.1017/S0140525X08003889) motivates separating forward prediction from retrospective interpretation as a conceptual question. Retrospective model correction here never changes the original evidence or historical action record.
* [Eagleman (2008), Human time perception and its illusions](https://doi.org/10.1016/j.conb.2008.06.002) concerns perceptual timing, with the reviewed phenomena restricted to subsecond scales. Transfer to project timing requires independent justification.

The predecessor invoked feature integration, predictive coding, global workspace and event-file theories. Their useful design questions are retained in work-package attribute cards, a ranked shared queue, prediction/correction and episode records. The implementation does not instantiate neural or cognitive mechanisms and has not measured any effect on human or organisational performance.

## Verification

The model is separated from the interface in `model.js`. Source tests are held outside the served folder:

* `tests/temporal-binding-model.cjs` independently calculates blended correction and expected projection, including both before-sample and after-sample action replay. It checks packet timing and one-time assimilation, immediate/delayed effects, delivery lead time, directional triggers, hidden-state independence, immutable captures, no-action invariance across decision delays, exact paired reproducibility, a nonzero default action contrast, objective arithmetic and bounded states.
* `tests/temporal-binding-ui.cjs` exercises the page in a DOM environment: dated reports, top-four queue and seven-signal table, settings draft/application boundary, validation, keyboard guards, run/pause, captures, JSON receipts and the paired-result display.

These checks establish implementation behaviour under the declared construction. They do not validate the scenario against festival data or establish cognitive benefits. Desktop and mobile rendered-browser review remains a separate check.
