Seal Phase 4AH conditional KC drive audit and source accounting

models/male-cns-lif/kc-drive-audit-v1/PROTOCOL.md

@@ -0,0 +1,71 @@
+# Phase 4AH — sealed-trajectory KC drive accounting
+
+This diagnostic was specified after Phase 4AG's fixed CAL failed. It is not a
+new hypothesis trial, a preregistered discovery, a new brain simulation, a
+parameter search, a held-out test, or training. All observations are conditional
+on the existing logged trajectories. A priori here means before this accounting
+is computed, not before the original CAL outcomes were known.
+
+## Scope and fixed inputs
+
+Read all eight Phase 4AG-r2 fixture r0 records and all 60 CAL r0 plus 12 first-seed
+r1 records. No outcome-based subset choice. Read actual static canonical metadata,
+all real incoming edges of the relevant KC, the exact frozen coefficients and SI
+weights, requested and actual events, recorded float32 APL inhibition, and final
+states. Keep all previous artifacts read-only. Do not import a GPU runtime or
+advance a new whole-brain state. Use at most four CPU threads. No external data
+download is required.
+
+## Exact target-state reconstruction
+
+Compute KC spikes independently, never force the logged KC spike/reset sequence.
+Upstream arrivals use the logged whole-brain source events at exactly t-18.
+At each tick decrement refractory, update available V/g in the exact separate
+float64 operation order, apply accepted request override, strict threshold,
+reset and ref22, then sequential canonical-order incoming edges only if post-reset
+ref==0. Finally widen stored float32 h to float64, multiply by float64 .001 and
+subtract unconditionally. In particular preserve the stored reset coefficient
+bits; do not substitute hand-typed -0.052 or recompute exponentials.
+
+Verify all target spikes/counts and final V/g/ref byte-for-byte. For fixtures also
+verify saved pre-APL and post-APL dense target histories at every tick. Full CAL
+has whole-brain hashes only: these cannot verify each-tick KC subset states.
+Never claim all-tick independently observed CAL state equivalence. On any failure,
+save diagnostic failure and stop without adjusting tolerances or publishing a
+mechanistic PASS.
+
+## Accounting (not counterfactual causality)
+
+Per KC and source class report actual due/accepted/rejected edges, original
+contact counts, nominal signed SI increments (positive and negative separately),
+and zeroed-APL arrivals. Keep actual anatomical population and dopamine annotation
+distinct; unavailable metadata is unknown, not a guessed identity.
+
+At each tick ledger the actual g changes due to available-state decay, reset,
+sequential arrivals and graded APL write. Sum these differences in extended
+precision only as a non-feedback accounting device; report reconstruction residual
+and an explicit roundoff bound. The actual replay always stays float64.
+Nominal per-source increments do not exactly sum to applied floating increments;
+report that rounding residual rather than altering the source labels or dynamics.
+
+Record APL writes during ref versus available state, pre-first-spike APL writes,
+negative-g exposure, threshold margin, PN input acceptance, and active versus
+silent KC summaries. KC-active/silent strata are descriptive and outcome-defined,
+not an experimental comparison. Cumulative increments and state-time integrals
+must retain distinct units and must not be called instantaneous conductance or
+a causal percentage of spike production.
+
+## Gate and interpretation
+
+Gate: all 80 prescribed records reconstruct exactly under the stated observable
+scope, inputs/checksums stay unchanged, all accounting is finite, and no missing
+records are silently skipped. Repeated-record metrics must agree excluding wall
+time. Pure adversarial tests are required before registration.
+
+Passing is PASS-CONDITIONAL-KC-DRIVE-AUDIT, not a learning/candidate gate. Do not
+infer that removing APL would restore sparse/separable coding, that refractory
+APL accumulation is the cause of CAL failure, or that a larger drive is justified.
+An actual counterfactual would require a separately declared new-model hypothesis.
+No gain/threshold/eta changes, new edges, mask expansion, held-out tuning, rewards,
+games, decoder, visual input or Web. Preserve failed records. New output budget
+2 GiB; source and experiment get SHA-256 manifests at completion.