Add fixed normal acquisition executor and independent evidence audit

models/male-cns-lif/normal-conditioning-v1/PROTOCOL.md

@@ -0,0 +1,91 @@
+# Normal olfactory acquisition: independent D18 branch
+
+This phase does not modify MaleCNS reference/runtime, Phase 1–4AQ data, the
+base graph, the 14,232 candidate edges, eligibility-v2, APL gain, cell thresholds,
+transmitter signs or any previously sealed failure. An engineering interface
+PASS is not evidence of learning.
+
+## Fixed science
+
+The original AO file (SHA-256
+`a7bda850ee87df4d9f647994e1a1a4074dfea5793184edb1822ab3d8c02e8216`)
+supplies master 20260930, all semantic seeds, ordered cues and control schedules.
+The old frozen evaluation policy supplies completed-epoch snapshot boundaries.
+There are 40 normal blocks: physiological/matched-energy × PAM15/PPL103 ×
+reciprocal reinforced odor × five training replicates. A is 4-ethylphenol, B is
+3-hexanone. Evaluation uses 80 independently registered input streams.
+
+All seven arms are real: paired, D=0, frozen, random reward, unpaired, CS-only
+and DAN-only. 30,000 complete fresh training/control epochs and 3,920 real
+fresh A/B probes are required. Even identical baselines are executed, not
+relabelled copies. Snapshots 0/1/2/5/10/20/50 never authorize outcome selection.
+Only the 50-pairing terminal outcome determines acquisition. Every arm finishes
+its entire original schedule, including trailing neutral or blank epochs.
+
+Epochs are 120 ms, CS is [0,100) ms, scheduled artificial US is at 90 ms.
+PAM15 and PPL103 use only their existing confirmed compartment-specific DAN
+identities. The positive/negative behavioral predictions are explicit model
+hypotheses, not a claim that every PAM/PPL neuron has that function. Scheduled
+D is an artificial dimensionless interface, not measured dopamine concentration
+or the number of accepted DAN spikes. US attempts/rejections and actual events
+remain in the original runtime artifacts. No resampling compensates for a
+rejected pulse.
+
+Fresh neural state is reset each epoch; only the separately authenticated
+plastic delta carries forward. Original E/U compute the end-of-epoch update,
+then a separate CPU process recomputes it from sealed actual KC events. Probes
+have no US, no E/update and no feedback. Brain math is untouched.
+
+## Readout and gates
+
+The frozen AQ `learning_gates.py` and original `metrics.py` are pinned before
+sampling. All six controls, raw signed MBON balance, functional count/latency
+change, MBON separation and both downstream readouts remain mandatory. The
+same replicate must satisfy the conjunction; at least three of five pass in
+each of the eight cells. PAM/PPL or input modes are never pooled to rescue a
+failure. Original H mean-control rules are reported separately for each DAN
+control source, both required. Delta change alone cannot imply learning.
+
+Readout populations are anatomical, never inferred from active/plastic subsets.
+The original H actual MBON downstream edge table defines legacyDownstream;
+nonMBONDownstream is its exact set difference with the full MBON population.
+Missing identity or an altered frozen source stops preparation.
+
+## Execution, resources and integrity
+
+Use only the registered 4090 UUID and serial execution. ComfyUI admission must
+be empty; never stop or evict a service. Source code is committed before
+registration. Static tables invoke the original L PN-rate transform, preserving
+all original PN columns for both modes. Shared CS aliases reuse saved input
+bytes only. The full 63-bit CUDA sampler, full matrix draw and CS window are
+unchanged; there is no CPU-RNG replacement.
+
+The engineering ceiling is 256 GiB of new outputs, at least 100 GiB disk reserve,
+1,800 seconds per child and an absolute 21-day registration deadline. The AQ
+linear estimate is only planning arithmetic from limited warm measurements,
+not a measured forecast across all conditions. Exceeding a limit means an
+incomplete preserved experiment, not shortened trials or changed parameters.
+
+Each child has an exclusive start record, actual parent-observed exit,
+dependency hashes and immutable output inventory. A stopped or uncommitted
+child is never silently retried. A persistent incremental ledger avoids
+rehashing every historical trial on every iteration; each invocation still
+rehashes its actual consumed dependencies. Full audits read every unique file
+and reject missing, extra, changed, linked or incomplete output. They are not
+claimed to be atomic filesystem snapshots. Mutable compiler cache is accounted
+separately. CPU/source tests are not described as GPU numerical regression.
+
+A detached supervisor keeps the serial controller alive across SSH disconnects.
+Administrative logs live in `/data/digital-drosophila/logs/phase4ar-normal-conditioning/`,
+outside the immutable per-trial ledger. The final outer auditor must retain
+these logs and its actual exit receipt. A launch PID alone proves no completion.
+
+All original event/state/queue/update artifacts are retained; no dense wholebrain
+time-by-neuron output is added. Complete historical sealed integrity is checked
+at the boundaries. The five historical reference trials are preserved evidence,
+not claimed as newly rerun here.
+
+Normal acquisition is the only authorized scientific scope of this registration.
+No delayed/reversal/retention assay starts automatically, and no candidate is
+issued from normal acquisition alone. There is no game, action decoder, RL,
+visual input, training of base weights, web service or 3D interface.