Add KV dynamic-state causal result note

kv_dynamic_state_causal_results.md

@@ -0,0 +1,81 @@
+# 3B K/V Dynamic-State Causal Revision Results
+
+## Run
+
+- Date: 2026-08-15 UTC.
+- Model: `Qwen/Qwen2.5-VL-3B-Instruct`.
+- Dataset: TAP-Vid-DAVIS local artifact.
+- Diagnostic tensors: `artifacts/dynamic_diag_3b_train_v2.pt` and
+  `artifacts/dynamic_diag_3b_test_v2.pt`.
+- Output: `outputs/kv_dynamic_state_causal.json`.
+- Rows: 2,592 condition rows from 108 held-out causal episodes.
+- Substrates: value primary, key replication.
+- Evidence conditions: correct, wrong, shuffled.
+- Methods: append-only, full structured edit, random rank-4, task-aligned rank-4.
+- Norm cap: 2.0.
+- Optimization steps: 40.
+
+## Structural validation
+
+The result file is complete over:
+
+```text
+108 episodes x 2 substrates x 3 evidence conditions x 4 methods = 2592 rows
+```
+
+Additional validation:
+
+- unique condition rows: 2,592;
+- duplicate condition rows: 0;
+- append-only bad rows: 0;
+- rows over norm cap: 0.
+
+The file SHA-256 after writing the embedded result hash is:
+
+```text
+8b46b24aa7fda3da33d3eae20082f8d4e8829205fd7395cb528cd0c1b0a61723
+```
+
+The runner-recorded pre-finalization SHA-256 is:
+
+```text
+90c3f468464078042036c5fab5cda2a565e2ccb39a75a231c5c6fad3e458417a
+```
+
+These differ because the final JSON stores the recorded hash inside the file.
+
+## Aggregate directional effects
+
+Mean directional log-probability gain for correct evidence:
+
+| Substrate | Method | Visibility delta | Moving delta | Mean edit norm |
+|---|---:|---:|---:|---:|
+| value | full structured | -0.0345 | -0.1161 | 0.1751 |
+| value | random rank-4 | 0.4715 | -0.9399 | 1.7819 |
+| value | task rank-4 | -0.0227 | -0.0779 | 0.1970 |
+| key | full structured | -0.2263 | -0.1712 | 0.1673 |
+| key | random rank-4 | 0.1100 | -0.3565 | 1.6944 |
+| key | task rank-4 | -0.2960 | -0.1625 | 0.2387 |
+
+Append-only is the zero-edit baseline and has zero directional delta by
+construction.
+
+## Interpretation
+
+Under the preregistered decision table, the current aggregate means do not
+support a shared revisable dynamic-state substrate. The value-cache task-aligned
+rank-4 edit, which is the primary method, has negative mean directional gains
+for both visibility and moving/static under correct evidence.
+
+The random rank-4 value edit improves visibility on average but strongly harms
+moving/static. That pattern does not satisfy the shared-state criterion because
+one finalized edit is expected to help both frozen probes.
+
+The current result is therefore best reported as:
+
+> The pooled diagnostic K/V state is editable under the fixed interface, but
+> this run does not show semantically directional revision from future RGB
+> evidence.
+
+No claim should be made about identity, trajectory belief, motion axis, general
+world belief, or arbitrary VLM memory.