Add post-exit normal acquisition audit with frozen gate regression

2026-09-30 21:38:56 +08:00
parent 1e429fd50d
commit 3fd58f2cf8
10 changed files with 1346 additions and 0 deletions
@@ -0,0 +1,61 @@
+# Phase 4AR read-only progress and post-run audit
+
+These tools are separate from the 16 registered scientific source files. They
+never launch a brain, change plastic delta, resample inputs, resume a partial
+epoch or authorize another assay. Tests use synthetic temporary data only.
+
+## While the experiment runs
+
+`progress_normal.py` reads the small committed ledger records. It distinguishes
+completed CPU-verified epochs, completed probes, pending processes and mere
+result files. It is not a whole-tree integrity check or a learning gate.
+
+`normal_finalize.py check` makes a real Linux process query and examines the
+actual terminal receipts. It intentionally refuses a live or incomplete run,
+even when a completion-looking file exists. An absent PID alone proves nothing.
+No final-audit directory or result is created by this check.
+
+## After all registered work has ended
+
+An operator must first provision this exact, empty, separately owned directory:
+
+`/data/digital-drosophila/experiments/phase4ar-normal-conditioning-final-audit/`
+
+Then run the committed `normal_finalize.py audit` with the existing science
+environment and `-B`. The driver refuses uncommitted audit code, a nonempty
+destination, missing/failed exit evidence, a live runtime or a stopped phase.
+There is no automatic retry after a partial finalization.
+
+The terminal ledger's authentic original PASS was written before
+`audit/execution-complete.json` existed. A fresh unmodified ledger audit will
+therefore report that exact one extra file. `terminal_ledger.py` retains both
+original reports unchanged. It accepts this boundary only if every other field,
+the full order, head, payload and compiler-cache identity agree, and both byte
+totals differ by exactly that completion file's actual length. It never deletes
+the file or changes a failure flag to PASS. Any other difference is rejected.
+
+The terminal inventory includes all experiment payload, registered external
+inputs and all six administrative supervisor files. Startup identities were
+independently bound while the run was live. The driver additionally checks real
+process quiescence before and after analysis; file existence is not a lock or
+an atomic filesystem snapshot.
+
+The frozen `normal_metrics.analyze` then runs in a fresh CPU process against all
+33,920 task records and 3,920 probes. Its real exit and file bytes are recorded.
+The old frozen data are reverified in a separate integrity process. Full
+terminal inventory is read again after analysis. All new reports and their
+manifest/checksums live outside the original experiment.
+
+`normal_outcome.py` checks the internally consistent fixed terminal report:
+all eight cells, the same-replicate conjunction, all controls and both H control
+source mappings. Incomplete/invalid evidence is not a scientific negative result.
+It cannot authenticate a caller's evidence flags itself; the driver must supply
+the actual verified files and process evidence.
+
+Even `NORMAL-ACQUISITION-PASS` is **not usable learning** or a model candidate.
+Delayed reward, reversal and retention remain unrun and are not automatically
+authorized. No parameter, seed or endpoint is chosen using intermediate results.
+
+Historical integrity checks do not rerun the five old reference trials and are
+never labeled as new numerical regression. Git stores code and selected small
+evidence, not a second copy of the bulk experiment dataset.