A sequential-decision essai · Powell’s framework

How old is the
picture you act on?

Toy scenario

A mountain festival is preparing its shuttle service. A weather report arrives, the forecast changes, and a decision is made. Those three things happen at different times.

3

Keep the times separate:
when it happened,
when you learned,
when you could act.

An authored simulation. One tick is an abstract planning interval; every index, cost and rule is illustrative. This is not a transport safety assessment.

01 / Watch information catch up

A new report.
An older world.

Step forward, introduce a disruption, then watch its report arrive. Review the queue to commit a preset response.

The organiser carries a state, chooses a response, receives new information and updates the state through a transition. An explicit objective lets us compare the resulting paths.

Review 0

Latest received reportWaitingNo report has arrived.
Current estimated readiness0%Projected to the current review.
Next committed effectNoneNo preset action queued.
Mountain festival shuttle preparationVillage depot, winding mountain route and festival. A moving ring marks the largest received discrepancy. The scene shows the current model, not hidden world state. Village depot Festival Crew Weather Mountain routeCURRENT MODEL · ring = largest received discrepancy
The scene follows the organiser’s estimate. It does not reveal hidden simulated truth to the action rule.

Four discrepancies to review

Ranked by absolute error × an authored signal weight × work-stage coupling. A large error is not automatically a high risk.

    The report belongs back here

    Inspect the report, its matching forecast, and the correction

    SignalReportedPrior for sample timeAbsolute errorRank score

    Each packet is compared with the revised prior for its own sample time, before that packet is used. Earlier report arrivals may already have revised this prior; it is not necessarily the forecast originally held at that historical review. A fixed blend corrects that old estimate. The model then replays its expected transitions and known action effects up to the present. The table is the discrepancy before correction; the readiness above is the current estimate after correction.

    Change the delays and authored assumptions

    Applying settings restarts the run. A delay is fixed when a report or action is created; it never changes halfway through its journey.

    A proposed review window

    This inherited heuristic is displayed for discussion only. It does not schedule reports, change decisions, or estimate a human temporal binding window.

    02 / Hold the surprises fixed

    Does faster action help
    under these rules?

    Two fresh 24-interval runs share your applied seed and report delay, the same noise draws, and disruptions at intervals 4, 8 and 12. Both use the same automatic queue policy. One keeps your decision delay; the other uses zero. Your manual path above remains available.

    Run the comparison to see readiness, exposure, action costs and the illustrative objective together.

    Why this is a comparison, not an optimum

    The rule only reacts to large, adverse discrepancies in the four selected signals. It can miss a dangerous state that the model already predicted. It does not evaluate all possible actions or account for every unobserved state. Acting earlier can also spend more or interrupt more work.

    The end-of-run objective is 100 × mean work-package completion − 25 × the larger exposure/disruption index − action costs. These are declared preferences and model points, not monetary benefits or measured safety outcomes. The chart shows one paired seed, not a statistical effect. Common random draws control one source of variation; action-dependent states and work paths can still diverge.

    03 / Keep the episode inspectable

    What was known?
    What was committed?

    A capture keeps the received report’s timestamp, current estimate, pending commitments and effects already executed. It does not turn a later observation into proof that an action caused it.

    No episodes captured.

    Inspect captured episodes
    No episodes captured yet.
    Open the event record
      Work packages: keep task, place, people and constraints together

      The six packages are serial in this deliberately small model. Each object below keeps its attributes attached; this is an information-design choice, not evidence of a cognitive effect.

      The distinction

      An aligned picture is still a model.

      A discrepancy can mean that the world changed, the forecast was weak, or the observation was noisy. Comparing a delayed report with a current forecast adds another source of disagreement: the two numbers describe different times.

      Keeping their timestamps restores a meaningful comparison. It does not make the observations complete or the response wise. The organiser still needs a state that remembers commitments, a decision rule, a transition model and an objective.

      Powell’s framework, the model construction and its limits

      The five parts

      State
      Current estimate, timestamped reports, stored predictions, pending actions, costs and remaining time. The simulator separately holds hidden world state.
      Decision
      Review the top four errors; commit at most two eligible preset responses, or wait.
      New information
      Noisy packets with sample and arrival times. Manual surprise controls intervene in the simulated world; they do not give the policy an observation.
      Transition
      Work and exposure drift, delayed action effects, then one-time correction at the report’s timestamp and projection forward.
      Objective
      Final preparation completion less exposure and action costs under a stated illustrative scoring rule.

      What this construction establishes

      Powell’s framework organises sequential decisions. This example adds its own dynamics, fixed blending weights, queue and action rules. It is a hand-built simulation and policy, not reinforcement learning or an optimised solution.

      The forecast is projected to the present between reports. It does not literally run ahead of time. The model treats action effects as known and deterministic, while its weather and deterioration rules are imperfect.

      Powell: On State Variables, Bandit Problems and POMDPs

      Read the full model and verification · Inspect the model

      Preset response rules

      No queue entry is a recommendation for a real service. In particular, the exposure indices are dimensionless authored variables, not probabilities of an accident.

      Optional research lens: what does “binding” mean here?

      Here, binding means keeping related project attributes and times together in an external record. Human perceptual binding and intentional binding are distinct research subjects. This simulation does not test either phenomenon, and a team is not a single perceptual system.

      Ideas retained as questions

      Feature integration: could keeping task, place, people and constraints together make an agenda easier to inspect? The work-package cards illustrate that design question.

      Prediction and correction: forecasts and exceptions can be represented together. The fixed blend here is an engineering rule; its numbers are not neural parameters.

      Shared workspace and event records: a queue makes selected items visible; a capture preserves a dated episode. Neither implements or validates a theory of consciousness, a team attention mechanism, or stimulus–action binding.

      Timing phenomena should stay distinct

      Cai, Stetson and Eagleman (2012) distinguish temporal recalibration from intentional binding. Their experimental distinction is a reason to keep this project’s timestamp alignment separate from claims about perceived time or agency.

      Cai, Stetson & Eagleman: A neural model for temporal order judgments and their active recalibration: a common mechanism for space and time?

      The optional analogy suggests questions for later research. The calculated results above come only from the disclosed simulation rules.