# WI-003 — Moving target: post-commit delay and the four receipt-state bindings (oracle)

Status: Built → Contract-verified → Measured (oracle) → ledger entry in LEDGER.md

```text
work_item_id:        WI-003
question_id:         Q3 (also Q5 second intervention)
hypothesis_id:       H3.1, H3.2 (instrument validation; oracle controllers)
study_id:            S3.1 (setting: predictably moving target), S3.2, S5.1 (state-at-B vs newest-at-D)
current_project_belief:
    The harness scores application-time validity correctly for a STATIC target. Whether it
    handles post-commit change (target moves after D) and can host the four S3.2 bindings with
    matched B, D, A — state at B; newest at D; predicted at A; true at A — is untested. The
    paper also needs the EXACT operation by which each binding uses evidence (§9.4 forbids
    merging "coordinate correction" with "same-decision update").
exact_uncertainty_to_reduce:
    U7. Does the evaluator's application-time validity and spatial error at A match the closed
        form for a target moving at constant velocity, over a post-commit delay sweep (S3.1)?
    U8. With B, D, A matched across bindings, do the four bindings order as H3.2 predicts on
        the moving target (true_at_A >= predicted_at_A >= newest_at_D >= state_at_B), tie on
        the static target, and does a mis-estimated delivery delay degrade predicted_at_A
        linearly (S3.2)?
    U9. Can the trace distinguish "coordinate correction at D" (newest_at_D / predicted_at_A)
        from a fixed-input decision, without calling either a same-decision update?
claim_or_decision_enabled_by_this_work:
    The operational definitions of the four bindings; the delay–value curve shape needed by
    S4.2; whether "newest at D" must be traced as a separately identified correction (yes).
intended intervention or measurement:
    scenario moving_target (velocity explicit, px/s); oracle controllers with `binding` in
    {state_at_B, newest_at_D, predicted_at_A, true_at_A}; sweeps over delivery delay and
    velocity; predictor fed expected_delivery_delay_ns with deliberate error.
matched conditions:          bindings share seed, frames, B, D, A exactly (simulated clock).
variables held fixed:        manifest, compute_ns, delivery delay per cell, action space, scorer.
primary outcome:             valid_at_receipt; spatial_error_px at A.
required trace fields:       §13 + `coordinate_correction` record (binding, observation used,
                             proposed-before / committed-after coordinates) for correcting bindings.
competing explanation ruled out:
    "Prediction helps because it gets more time or newer frames" — ruled out by matched B, D, A
    and identical admitted frames.
possible result patterns and their interpretations:
    U7/U8 (a) closed form holds and ordering holds -> instruments valid; definitions pinned.
          (b) mismatch -> scorer or binding implementation wrong; fix before any agent study.
project_delta_expected_from_each_result: as above.
project_decision_unlocked:   S3/S5 can run with real agents on this scenario; S4.2 has its
                             delay–value relation for the moving-target case.
generalist_reader_starting_point:
    "If the thing you are aiming at keeps moving after you decide, where should you aim?"
one_sentence_reader_takeaway:
    Aiming at where the target IS when you decide is wrong by velocity x (delay until the click
    lands); aiming at where it WILL be fixes that only if you know the delay.
reader_confusion_or_misconception_prevented:
    "Using the newest frame at D" and "predicting the state at A" are both coordinate
    corrections applied at commitment, not same-decision updates, and neither helps a static target.
positive acceptance test:    tests/test_moving_target.py (closed form; ordering; perturbation).
negative contract test:      a correcting binding whose trace lacks the coordinate_correction
                             record is flagged; a correction that changes action KIND is rejected.
stop condition:              any closed-form mismatch.
what this work does not establish:
    that any real agent can estimate velocity or benefits from prediction; anything physical.
what remains unknown afterward:
    real-agent behaviour on the moving target; the non-oracle predictor's error.
```