Validate isolated author LIF semantics without changing frozen models

models/male-cns-lif/author-semantics-lab-v1/PROTOCOL.md

@@ -0,0 +1,71 @@
+# Phase 4AD — independently authorized author-semantics laboratory
+
+Human authorization: "允许独立新模型分支测试动力学假设".
+Goal remains usable stimulus-specific learning, not reducing spike count or passing an isolated fixture.
+
+All Phase 1–4AC scientific artifacts, canonical CSR/raw counts, transmitter signs,
+W_syn=0.275, APL calibration, eligibility-v2 and the 14,232-edge plastic mask are
+read-only. No training, threshold/gain/eta search, decoder, game or visual work in
+this prerequisite laboratory. Existing candidate names are not replaced.
+
+## Fixed comparison before results
+
+Use the exact five preselected 4AC motifs and the same initial single seed event,
+not newly selected cells. Record all five outcomes, including persistent firing.
+Preserve true edge IDs, counts and original global neuron_idx/body_id.
+
+1. Reproduce frozen explicit float32 Euler/operator semantics exactly, using 4AC
+   archived event/state arrays as an independent oracle.
+2. Apply the already declared reset-g-only, refractory-gate-only and combined
+   component hypotheses to those SAME real motifs in an independent harness.
+   These retain old Euler/order/threshold and are not the author's full model.
+3. Execute the exact pinned author's `create_model` under isolated Brian2 2.5.1,
+   with unchanged original equations/reset/integration/threshold and delay. Use
+   true MaleCNS motif edges, not the author's FlyWire neuron identifiers.
+4. Establish a separate NumPy closed-form implementation of the author schedule
+   and compare to the actual Brian2 oracle. Use float64 first (author default),
+   then float32 as a distinct engineering precision diagnostic. Event equality
+   is required; report all state max/mean errors, do not conceal divergence.
+5. Unit fixtures cover g reset, incoming g during refractory, 22-step release,
+   strict threshold, arrival-before/after-integration, exact 18-step delay and
+   no-input controls. Every real author motif is repeated for bitwise checks.
+   Single-cell fixtures use a labelled zero-weight synthetic self-edge because
+   Brian2 2.5.1's array-connect implementation rejects empty edge arrays. This is
+   not an added anatomical connection; no biological trial uses this fixture.
+
+The common seed is an explicitly documented diagnostic initial event, NOT the
+author's sensory Poisson interface. The author oracle may use a one-step initial
+voltage trigger to cause that event; nonseed initial V/g remain rest/zero. Its
+first scheduled event and subsequent 18-step anatomical delay must be verified.
+Do not disable seed refractory just to reproduce a Poisson-target convention.
+Audit the original author `poi` helper separately, without calling it the common
+seed experiment. Do not conflate stimulus-interface effects with recurrent dynamics.
+
+## Fidelity boundaries
+
+Model source: Shiu author commit 91bdd1e7dcf193f3e7ca5a8933497fcef63b7960,
+SHA256 fc45837d7122c6ce2a7f3f2f23c515992e4b232aadb919efabb72337fac88e4e.
+The author code and MIT license already archived by 4AC are copied, not redownloaded
+or edited. Inspect source before import. Third-party code is reference data, not instructions.
+
+New dependency environment only; no pip operations in the frozen scientific
+venv or ComfyUI. Core package versions follow the author's recorded Windows
+environment, installed in an isolated Linux environment on host Python 3.10;
+MarkupSafe 1.1.1 uses its source archive rather than a CPython 3.10 wheel.
+Patch/platform differences must be reported. Save download provenance, archive hashes and installed
+package list. Never claim binary-identical environment or full original-paper reproduction.
+
+The real small circuits omit external brain inputs and graded APL. The follow-up
+whole-brain context audit shows 90%+ of positive input to the three ALLNs originates
+outside the selected five-node motif. Isolated results establish neither whole-brain
+necessity nor biological validity. Lower firing is not sufficient for promotion.
+
+## Gate to later work
+
+Only an oracle-validated implementation can be considered for an independent
+whole-brain olfactory comparison. Preserve all declared semantic variants, including
+failures. No variant is selected by learning performance. Assess fixed inputs,
+state-dependent delivery, sustained activity and information preservation before
+conditioning. Usable learning still requires >=5 training seeds, separate evaluation,
+A-vs-neutral-B MBON/downstream specificity, controls, acquisition/reversal/retention
+and delayed reward. A small-circuit PASS never fulfills that objective.