Preserve context development diagnostics and paired training timeout

docs/NATIVE_CONTEXT_GATE.md

@@ -9,6 +9,26 @@ No cloud workload, download, private-index change or remote telemetry operation
 started by this change. Existing model, tokenizer v4, checkpoint, training, API,
 telemetry, context ceilings and batch policy are not changed.
 
+## Observed failure and development evidence (September 30, 2026)
+
+The [observed development diagnostic](evidence/context/development-factorial96-20260930.md)
+and [portable receipt](evidence/context/development-factorial96-20260930.json)
+retain the fixed v8 marker failure: 3/18 against the unchanged 80% requirement.
+This differs from the earlier v6 6/18 historical failure. The separate 96-case
+short/medium readout scores Choice 79/96, Score 90/96 and Noul 48/96; every
+Noul prediction is true. It does not establish exact-limit correctness, a
+representative quality result, or a passing native-context-v3 gate.
+
+The label readback finds no mismatch. Target correlations and wording/rubric/
+distractor differences leave multiple explanations for these failures. The
+separately frozen paired144 attempt subsequently timed out with 96/144
+durable updates; see the [timeout receipt](evidence/context/paired144-timeout-20260930.md).
+No completed repair, calibration or readout result is claimed. A new run initialized from weights resets
+current-run exposure; resumed training preserves it. A separately fixed
+exact-limit continuation is required before final validation.
+Protocol ceilings, numerical tolerances, marker threshold and issue status
+remain unchanged.
+
 ## What changes in the acceptance report
 
 The protocol is now `native-context-v3`; preserve prior v2 reports as historical,

docs/RELEASE_READINESS.md

@@ -1,5 +1,29 @@
 # Release readiness — intentionally incomplete
 
+## Observed synthetic context decisions (September 30, 2026)
+
+The [development diagnostic](evidence/context/development-factorial96-20260930.md)
+and [portable receipt](evidence/context/development-factorial96-20260930.json)
+record 96 short/medium generated cases on the fixed separately calibrated v8
+checkpoint. Choice scores 79/96, Score 90/96 and Noul 48/96. All Noul predictions
+are true; all six Score errors have gold zero. Choice first-ranked gold scores
+18/32, compared with middle 30/32 and last 31/32. Label/option/target readback
+finds no mismatch. Correlated generated targets, marker wording, rubrics and
+distractors limit causal interpretation; false and zero labels existed in prior
+training. Positive temperature calibration cannot change argmax.
+
+The original v8 native-context-v3 marker result is 3/18 against the unchanged
+80% threshold. The earlier v6 result of 6/18 remains a separate historical
+failure. The short/medium development scores establish neither a passing
+32768/65536 gate nor representative deployment quality or Jev benefit.
+The separately frozen paired144 attempt subsequently timed out with 96/144
+durable updates; see the [timeout receipt](evidence/context/paired144-timeout-20260930.md).
+No completed repair, calibration or readout result is available. A new run
+initialized from weights resets current-run exposure; resuming a run preserves
+its exposure. A later fixed exact-limit continuation is required before final
+context acceptance. Existing numerical receipts and
+measured rank-0 resume/resource failures remain unchanged; issues #3/#4 stay open.
+
 ## Verified numerical/source delivery (September 29, 2026)
 
 [PR #56](https://github.com/Jev-Engineering/layev/pull/56) merged at

docs/acceptance-status.json

@@ -1,4 +1,23 @@
 {
+  "context_development_diagnostic_2026_09_30": {
+    "scope": "Observed synthetic short/medium development diagnostic; not acceptance, representative quality or Jev comparison",
+    "receipt": "docs/evidence/context/development-factorial96-20260930.json",
+    "explanation": "docs/evidence/context/development-factorial96-20260930.md",
+    "private_report_sha256": "331b353f58d6225e4d42ef266557d6f345d109c042df9500c86330fba42f3fa7",
+    "checkpoint_sha256": "c37e133a21c10054584b345140d767c91bd4beb63faecda28f524ef1339f0cb9",
+    "cases": 96,
+    "tasks": 288,
+    "choice_correct": 79,
+    "score_correct": 90,
+    "noul_correct": 48,
+    "failure_patterns": "All 96 Noul predictions true; all six Score errors gold zero; Choice first/middle/last gold 18/32, 30/32, 31/32",
+    "causal_limits": "No label mismatch; prior training includes false/zero; target correlations and wording/rubric/distractor differences prevent assigning one cause",
+    "original_v8_marker_gate": "3/18 below unchanged 80%; failed; separate from historical v6 marker 6/18",
+    "paired_training": "Subsequent paired144 attempt timed out at 5400.990688599995 seconds, exit -15; durable96/144, unsaved updates unknown; no complete repair/calibration/readout result. See docs/evidence/context/paired144-timeout-20260930.json",
+    "new_training_exposure": "A new run initialized from weights resets current-run exposure; resumed training preserves it. Historical parent exposure cannot establish exact-limit exposure for the new run",
+    "issue_3": "open",
+    "issue_4": "open"
+  },
   "canonical_fp32_graph_delivery_2026_09_29": {
     "pull_request": "https://github.com/Jev-Engineering/layev/pull/56",
     "source_commit": "4ffd9f1f1263d225e4f130f174cf6070c635d235",