# Exact-interface recovery plan

## Objective

Test whether future RGB evidence can causally revise visibility and
moving/static variables only after proving that those variables remain
decodable on the exact cache interface being edited.

## Phase 1 — freeze and reproduce

1. Freeze the current 24-episode result, hashes, split, seeds, probe bundle, and
   failed competence report.
2. Reproduce the original two-frame diagnostic extraction and verify the
   published held-out probe scores.
3. Add an automated cache-interface signature covering prompt text, image
   count/order, chat template, generation suffix, pooled layers, token mask,
   and preprocessing configuration. Refuse mismatched probes by default.

## Phase 2 — choose one aligned design before held-out evaluation

Primary design:

- Build the editable base cache with the exact validated two-frame diagnostic
  prompt.
- Keep the probe evaluation on that unchanged token layout.
- Construct a separate teacher continuation that receives future RGB.
- Optimize only a bounded edit to the historical cache against a label-free
  teacher/student consistency objective.

Fallback design, used only if the primary design cannot preserve model-valid
continuation semantics:

- Extract train and test states using the exact three-image causal prompt.
- Train new probes on training videos only.
- Freeze them and require held-out competence before any causal analysis.

Do not select between designs using causal probe deltas.

## Phase 3 — competence gate

Run the unedited held-out cache through the frozen K/V visibility and
moving/static probes. Require lower 95% video-bootstrap bounds above 0.5 for
both balanced accuracy and AUROC. The primary causal target is V; K is a
replication.

If the moving/static probe fails, restrict the experiment to visibility. If
visibility also fails, stop: the causal target is invalid.

## Phase 4 — fixed causal test

- Model: `Qwen/Qwen2.5-VL-3B-Instruct`, frozen BF16.
- Dataset: TAP-Vid-DAVIS with the existing video-disjoint split.
- Evidence: correct, wrong, and fixed cross-video shuffled RGB.
- Methods: append-only, full KV edit, random rank-4, task-aligned rank-4.
- Primary statistic:
  `log p(y | KV') - log p(y | KV0)`.
- Required directional comparisons: correct greater than zero, correct greater
  than wrong, and correct greater than shuffled.
- Inference: paired video-level bootstrap with 10,000 resamples.
- Hardware: one RTX A4000; the measured configuration fits comfortably in 16
  GiB and does not require multi-GPU execution.

Run a small engineering pilot only to check finite gradients, loss reduction,
cache targeting, and norm caps. Do not tune using held-out probe movement.

## Phase 5 — claim and benchmark gate

- Visibility and moving/static both pass: shared revisable simple dynamic state.
- Visibility only passes: revisable perceptual state.
- Correct is not better than both controls: editable but not semantically
  revisable by future evidence.
- Competence fails: no causal claim.

Only after a valid positive mechanism result should the method move to native
model-answer benchmarks. Evaluate VSTAT first with the same frozen backbone and
append-only versus KV-revision inference. Use SpaMEM Level 3 as the second
world-state validation. Report same-backbone gains before any SOTA claim.