Seal Phase 4AJ fixed APL timing hypothesis audit

models/male-cns-lif/apl-timing-lab-v1/PROTOCOL.md

@@ -0,0 +1,72 @@
+# Phase 4AJ: a single graded-APL timing hypothesis
+
+Independent branch authorized by the human. All earlier models and data remain
+read-only. The learning goal remains unmet; neither an audit nor fixed-CAL PASS
+constitutes usable learning.
+
+## Fixed intervention
+
+Use the exact sealed AG author-style float64 LIF and float32 legacy graded APL.
+Compare h applied at its generating tick with the identical generated operator
+applied 18 steps later. No other change: dt=0.1 ms, anatomical delay=18 steps,
+refractory=22 steps, tau_m=20 ms, tau_syn=5 ms, W_syn=0.275 mV, original signs,
+threshold/rest/reset, all canonical IDs/edges, legacy APL tau=20 ms and gain=0.01.
+The unchanged fast-APL zero mask includes 1,212 non-KC edges not replaced by the
+graded operator. The branch does not silently fix this limitation.
+
+This is engineering scheduling parity, not evidence of a measured 1.8 ms graded
+APL delay. Whole h is delayed, including KC feedback; it is not a PN-only
+intervention. APL still subtracts unconditionally at end-of-step, including
+refractory KC, whereas ordinary synaptic arrivals are guarded by refractory.
+Equal arrival time is not equivalent biological action.
+
+## Stages and gates
+
+1. Exact conditional reconstruction of all 420 metadata-defined ALLN in all
+   60 sealed CAL r0 records. Audit every simultaneous earliest non-design PN,
+   before first KC, and duration-matched never-active PN controls. PN→ALLN→PN
+   witnesses must be actual accepted timed edges; distinguish signed weights,
+   reset windows and zeroed edges. No shortest paths, unique-cause or necessity
+   claims. Unknown antecedents are retained and do not falsify numerical audit.
+2. Fixed stored h trace shift audit: no loss, duplication, amplitude arithmetic
+   or brain simulation. Initial +0 padding and chronological unconsumed tail
+   explicitly retained. No extra brain/APL tail advance or future-decay claim.
+3. Both arms: eight fixed real-edge fixtures, three fresh GPU repeats, independent
+   Brian LIF/own spikes/own FIFO, legacy APL component audit. Delay0 must reproduce
+   sealed AG scientific outputs bitwise.
+4. Both arms: the same two fullgraph first-seed CAL probes, independent Brian
+   comparison and APL component audit. Numerical failure stops later execution.
+5. Both arms: every original six-odor × two-mode × five-seed CAL, plus first-seed
+   repeat for each odor/mode. Exact source input files, not regenerated random
+   inputs. All results retained. Known baseline representation failure does not
+   skip treatment; no seed selection, threshold/gain search or gate relaxation.
+6. Recompute comparisons from raw evidence, old checksums and five frozen
+   reference regressions. Seal truthful outcome with SHA-256 and source version.
+
+## Interpretation limits
+
+Generated h depends on changed closed-loop spikes, so branch inhibition integrals
+need not match. Fixed-trace conservation is a transport check, not proof of
+equal closed-loop dose. Accepted request counts can also change with refractory
+state, despite byte-identical requested input schedules.
+
+Brian state integration is independent but APL arithmetic is intentionally a
+shared frozen component. Original operator AST replay and rounding audit remain
+separate; no universal CPU/GPU equivalence is claimed from sampled fixtures.
+
+CAL gate retains original 3–9% mean KC fraction, all-run limits, within/between
+Jaccard and Euclidean gates, and minimum odor-pair count in both input modes.
+Passing fixed CAL is not heldout generalization or associative memory. Further
+progress requires independent seeds/novel odors, credit specificity, isolated and
+whole-brain functional readouts, then conditioning with proper controls.
+
+No training, eta/threshold/sign tuning, plastic mask expansion, heldout dynamic
+access, visual input, decoder, RL, game, Web or new candidate in this experiment.
+
+## Literature scope
+
+Shiu et al. 2024, DOI 10.1038/s41586-024-07763-9, supplies the spiking-LIF
+scheduling reference, not a measured MaleCNS graded-APL delay. Amin et al. 2020,
+DOI 10.7554/eLife.56954, and Prisco et al. 2021 (PMC8741211) motivate local graded
+APL/calyx interaction but do not determine this millisecond delay value. Prior
+sealed source/literature audits retain versions and provenance.