Seal AO interface failure and verified isolated postprocessing diagnosis

registry/phase4ao-conditioning-interface/source-snapshot/EVALUATION_POLICY_DRAFT.md

@@ -0,0 +1,68 @@
+# Explicit evaluation policy — draft implementation decision
+
+Status: **DRAFT / NOT REGISTERED / NO SCIENTIFIC EXECUTION**.
+
+This is a pre-outcome implementation choice by the coordinator for the existing
+conditioning proposal. It does not change a frozen brain, eligibility-v2,
+anatomical mask, odor identity, five-replicate matrix or scientific pass gate.
+The AN engineering result remains separate; none of its 21 sources is edited.
+
+## Probe timing and snapshots
+
+- Every evaluation probe lasts 120 ms, with CS requests only in the half-open
+  interval [0, 100 ms). The remaining 20 ms is a no-CS tail. This matches the
+  proposed normal training CS duration; it is not the old L full-120-ms input.
+- Normal, delayed and reversal curves use 0, 1, 2, 5, 10, 20 and 50 scheduled
+  reinforcement opportunities. Intermediate snapshots are taken after the
+  **complete epoch containing** that opportunity and its boundary update,
+  not immediately at the US tick.
+- Terminal checkpoint 50 includes all planned epochs, including trailing
+  neutral or blank epochs. It is not an early snapshot chosen for response.
+- Each snapshot executes both A and B evaluation inputs. No baseline-output
+  deduplication is assumed. Repeated probes use the same independent evaluation
+  input within a training replicate; they are paired observations, not new
+  training replicates. There are still five independent training replicates
+  per experimental cell and 80 independent A/B evaluation streams overall.
+- Probe US and weight updates are disabled. A probe cannot update the training
+  or retention trajectory. Acquisition/reversal/delayed start from the common
+  declared reset state; retention forks the actual continuous saved state.
+
+## Intermediate control alignment
+
+For normal and delayed experiments, obtain the paired arm's complete ordered
+CS prefix at the chosen epoch boundary. Match actual cue, occurrence, odor,
+source random stream and full seed, not just the number of epochs.
+
+- D=0, frozen, random reward and CS-only use that identical CS prefix.
+- DAN-only uses the corresponding removed-CS/yoked epoch positions. It has
+  **zero CS exposure**, not equal CS exposure; this is an intended control.
+- Unpaired uses the earliest epoch boundary that has consumed the same ordered
+  CS prefix, including US-only epochs already inserted before that boundary.
+  Its terminal checkpoint includes the entire 150-epoch inventory.
+
+Unpaired intermediate US counts and elapsed time can differ from paired.
+Random-reward intermediate US counts can also differ. These differences must
+be reported rather than called fully matched exposure. All complete terminal
+inventories remain fixed by the original proposal. The terminal gate, not the
+most favorable intermediate point, decides the result.
+
+Reversal and no-reversal use each arm's own kth reinforcement epoch boundary.
+Their intermediate CS prefixes may differ because their reinforced cues differ;
+their complete 100-epoch inventories remain fixed. The signed readout always
+keeps the original acquisition cue orientation and original pre baseline.
+
+## Inventory, not approval
+
+Under these choices the pure planner is expected to enumerate 9,440 probes
+(1,132.8 biological seconds), in addition to the existing 10,760 seconds of
+training, controls, reversal and continuous idle. This arithmetic is not a
+measured wall-time or disk estimate and excludes engineering validation,
+encoding, output, regressions and reserves. Test fixtures use synthetic masters;
+none is adopted as a scientific master seed.
+
+Before any scientific run, complete the actual input/state/delta file bindings,
+exact event-to-eligibility update coordinator, implemented assay gates and a
+finite resource plan. Register the protocol and bounded implementation check
+before observing scientific outcomes. Only after those checks may the staged
+acquisition suite run. A failed acquisition is sealed as failure; it must not
+trigger more pairings, parameter changes, odor reselection or control removal.