Prepare fixed-checkpoint sensitivity aggregate gate

AGENTS.md

@@ -333,6 +333,17 @@ record-zero replacements after review, merge, and a new exact-commit owner GO.
 The successful seeds must not be rerun, and aggregation remains separately
 unauthorized. See
 `docs/f2_f3_fixed_checkpoint_batch_stability_repair.md`.
+The separately authorized issue-#194 replacements then completed seeds 3407
+and 3409 with exit code zero and zero restarts. Together with the three
+issue-#192 completions, all five unselected-checkpoint seed pairs are now
+available privately. No aggregate was computed during monitoring. Issue #196
+prepares one separately gated CPU-only audit that will validate the 20 unique
+selected/unselected prediction files and retained checkpoint identities before
+assembling fixed-epoch-1, fixed-epoch-2, and dev-selected policy statistics.
+Preparation and merge authorize no Kubernetes object, PVC or prediction read,
+metric, inference, training, checkpoint selection or reselection, retry,
+QALB-test access, diagnostic, prompt/parser change, or XG. One aggregate Job
+requires a fresh owner GO naming the exact merged commit.
 Issue #116's separately authorized, corpus-text-free runtime probe subsequently
 completed on phone-verified Kaggle account `thgh15`. It confirmed exactly one
 Tesla P100, CUDA 12.8, and importable preinstalled PyTorch 2.10.0+cu128, but

PLANS.md

@@ -910,3 +910,18 @@ or XG.
 
 Issue #190's single authorization is consumed. Do not rerun the audit or use
 the result to change the primary parser, predictions, or score.
+
+## Issue #196 — prepare fixed-checkpoint sensitivity aggregation
+
+- [x] Freeze the three completed source attempts and executable commits.
+- [x] Reuse 20 unique selected/unselected prediction files across three policies.
+- [x] Require exact training checkpoint, prediction, count, schema, and alignment gates.
+- [x] Prepare one CPU-only, write-once, no-retry aggregate Job.
+- [x] Pass focused and full repository validation plus Kubernetes dry-runs.
+- [ ] Obtain review and merge the repository-only preparation.
+- [ ] Obtain a fresh exact-commit owner GO before one aggregate Job.
+
+Issue #196 authorizes repository preparation only. It authorizes no Kubernetes
+object, PVC or prediction read, metric, inference, training, checkpoint
+selection or reselection, retry, QALB-test access, diagnostics, prompt/parser
+change, or XG.

docs/f2_f3_fixed_checkpoint_aggregate_protocol.md

@@ -0,0 +1,49 @@
+# F2/F3 fixed-checkpoint aggregate protocol
+
+Status: issue #196 preparation only; no execution is authorized.
+
+## Purpose
+
+This CPU-only audit assembles the post-hoc five-seed sensitivity table for
+fixed epoch 1, fixed epoch 2, and the frozen common-development checkpoint
+selection rule. It performs no inference and does not change any checkpoint.
+The original seed-3407 P100 comparison remains the primary experiment.
+
+## Frozen private sources
+
+- selected checkpoints: issue #183 attempt `5155890101`, evaluation commit
+  `e004e625a00c9c1c6fac7e2dbc0e7bc450fbad17`;
+- unselected checkpoints for seeds 3408, 3410, and 3411: issue #192 attempt
+  `5157509573`, executable commit
+  `6b77efafd53660d2b98557b93cff983e91dbbf27`;
+- unselected replacement checkpoints for seeds 3407 and 3409: issue #194
+  attempt `5158062318`, executable commit
+  `3b2a30aa994071d5a51a51f62ee31df6cd13d958`.
+
+The audit validates the retained training completion records and exact adapter
+hashes without exposing development losses. It then validates 20 unique
+prediction files: two arms, two epoch checkpoints, and five seeds. The three
+reported policies reuse those files by exact checkpoint identity.
+
+## Validation and output
+
+Before computing any statistic, the audit requires exact source attempts and
+commits, complete 511-row counts, the frozen test hash, prediction hashes,
+schema, paired order, batch-16 RTX 3090 identity, pre-test gate success, and
+the absence of retry, training, QALB-test use, prompt/parser change, or exposed
+development values. It recomputes each source summary's counts and paired
+statistics from private rows.
+
+The public result contains only corpus-text-free per-seed counts, accuracies,
+policy means, sample standard deviations, F3-minus-F2 differences, source
+hashes, and validation flags. Predictions, record IDs, corpus text, model
+responses, adapters, private logs, and development values remain on the PVC.
+
+## Execution boundary
+
+The manifest contains one write-once, no-GPU Job with `backoffLimit: 0` and a
+one-hour deadline. Preparation and merge authorize no Kubernetes object, PVC
+read, prediction read, metric, model loading, inference, training, checkpoint
+selection or reselection, retry, QALB test, diagnostics, prompt/parser change,
+or XG. One execution requires a fresh owner GO on issue #196 naming the exact
+merged commit and confirmation `AGGREGATE_F2_F3_FIXED_CHECKPOINT_RESULTS`.

results/f2_f3_fixed_checkpoint_aggregate_gate_audit.md

@@ -0,0 +1,36 @@
+# F2/F3 fixed-checkpoint aggregate gate preparation audit
+
+## Outcome
+
+Issue #196 prepares, but does not execute, one CPU-only write-once aggregate
+audit for the completed issue-#183, issue-#192, and issue-#194 sources. It
+validates 20 unique private prediction files, the retained checkpoint
+identities, source hashes, 511-row contracts, and record alignment before
+assembling fixed-epoch-1, fixed-epoch-2, and dev-selected policies.
+
+The proposed public output is corpus-text-free. It retains only aggregate
+counts and statistics, source hashes, checkpoint names, and validation flags.
+No private prediction, record ID, corpus text, response, adapter, log, or
+development value may leave the PVC.
+
+## Authorization
+
+Preparation and merge authorize no Kubernetes object, PVC read, prediction
+read, metric computation, GPU use, model load, inference, training, checkpoint
+selection or reselection, retry, QALB test, diagnostic, prompt/parser change,
+or XG. A fresh exact-commit owner GO on issue #196 is required before exactly
+one CPU-only aggregate Job.
+
+## Validation
+
+- `python -m unittest tests.test_f2_f3_fixed_checkpoint_aggregate -v`: 3 passed.
+- `python -m unittest discover -s tests -v`: 316 passed, 2 optional data skips.
+- `python -m compileall scripts`: passed.
+- focused Ruff and `git diff --check`: passed.
+- generated manifest JSON parsing: passed.
+- Kubernetes client and server dry-runs: one CPU-only Job passed both.
+- read-only cluster queries before and after dry-run: zero persisted issue-#196
+  Jobs.
+
+No Kubernetes object, PVC read, prediction read, metric, model load, inference,
+or training occurred during preparation.