Audit conditional PN recruitment and graded APL source timing

2026-09-30 10:45:48 +08:00
parent 9437bbff79
commit b2cd81c7b9
283 changed files with 159989 additions and 9 deletions
@@ -0,0 +1,63 @@
+"""Post-seal descriptive interpretation; no trajectory rerun or gate change."""
+import hashlib
+import json
+from pathlib import Path
+import sys
+import numpy as np
+import pandas as pd
+
+ROOT = Path('/data/digital-drosophila')
+MODEL = ROOT / 'models/male-cns-lif/pn-recruitment-audit-v1'
+EXP = ROOT / 'experiments/phase4ai-pn-recruitment-audit'
+REGISTRY = ROOT / 'registry/phase4ai-pn-recruitment-audit'
+sys.path.insert(0, str(MODEL))
+from context import read, save, digest, GROUPS
+from publication import verify_seal
+
+
+def main():
+    verified = {'model': verify_seal(MODEL), 'experiment': verify_seal(EXP)}
+    status = read(EXP / 'audit/final-status.json')
+    structure = read(EXP / 'audit/PN-odor-structure.json')
+    output = {}
+    for mode, aggregate in status['modes'].items():
+        rows = [r for r in read(EXP / 'protocol/inventory.json') if r['repeat'] == 0 and r['kind'] == 'calibration'
+                and read(EXP / 'trials' / (r['case'] + '-r0') / 'summary.json')['mode'] == mode]
+        assert len(rows) == 30
+        pn = pd.concat([pd.read_parquet(EXP / 'trials' / (r['case'] + '-r0') / 'per-pn.parquet') for r in rows], ignore_index=True)
+        chosen = pn[(~pn.is_design_target) & pn.first_spike_observed]
+        positives = {g: float(chosen[g + '_accepted_positive_before_first_V'].sum()) for g in GROUPS}
+        total = sum(positives.values())
+        source = {g: dict(pre_first_accepted_positive_V=positives[g],
+            share=None if total == 0 else positives[g] / total,
+            pre_first_accepted_negative_abs_V=float(chosen[g + '_accepted_negative_before_first_abs_V'].sum()),
+            mean_integrated_positive_V=float(chosen['first_integrated_V_from_' + g + '/positive_weight'].mean()),
+            mean_integrated_negative_V=float(chosen['first_integrated_V_from_' + g + '/negative_weight'].mean())) for g in GROUPS}
+        view_metrics = {}
+        for name in ('fixed_686', 'common_non_design'):
+            view = structure['modes'][mode][name]
+            view_metrics[name] = dict(dimensions=view['dimensions'],
+                metrics={m: dict(within=v['mean'], between=view['between'][m]['mean'],
+                    pair_mean_between_greater_than_within=sum(p['mean_between_minus_within'][m] is not None and
+                        p['mean_between_minus_within'][m] > 0 for p in view['pairs'])) for m, v in view['within'].items()})
+        output[mode] = dict(frozen_aggregate=aggregate,
+            active_non_design_PN_neuron_trials=len(chosen), unique_active_non_design_PN=int(chosen.neuron_idx.nunique()),
+            active_non_design_PN_prefirst_source_accounting=source,
+            source_shares_are_nominal_positive_increment_sums_not_necessity_or_instantaneous_current=True,
+            odor_structure_descriptive=view_metrics)
+    result = dict(status=status['status'], modes=output,
+        exact_PN_trials=status['exact_PN_trials'], repeat_accounting_pass=status['repeat_accounting_pass'],
+        source_script_sha256=digest(Path(__file__)),
+        model_manifest_sha256=verified['model']['manifest_sha256'],
+        experiment_manifest_sha256=verified['experiment']['manifest_sha256'],
+        descriptive_postseal_interpretation=True, new_scientific_gate=False, learning_achieved=False,
+        original_CAL_failure_unchanged=True)
+    path = REGISTRY / 'interpretation-metrics.json'
+    save(path, result)
+    with (REGISTRY / 'interpretation-metrics.sha256').open('x') as f:
+        f.write(digest(path) + '  interpretation-metrics.json\n')
+    print(json.dumps(result, ensure_ascii=False, allow_nan=False), flush=True)
+
+
+if __name__ == '__main__':
+    main()