digital-drosophila/models/male-cns-lif/lif-semantics-audit-v1/PROTOCOL.md

Phase 4AC — upstream LIF semantics audit

Goal context: usable stimulus-specific associative learning remains unachieved. This stage is a prerequisite diagnosis, never a substitute for that goal.

All raw, base CSR, old models/experiments and signs remain read-only. No eta, threshold, gain, mask or APL calibration is changed. No conditioning, games, decoder, visual stimulation, or new formal runtime is produced.

New evidence and hypotheses

The official Shiu paper links philshiu/Drosophila_brain_model. Commit 91bdd1e7dcf193f3e7ca5a8933497fcef63b7960 (2024-09-14) model.py explicitly resets g on a spike, declares both v and g (unless refractory), uses linear integration and strict v > v_th, and removes refractory time on Poisson-target neurons. Current frozen Phase 4B/P instead retains g, decays/accepts arrivals during refractory, uses explicit Euler, >=, and force requests respecting refractory. The Phase 4B config documents its refractory-g policy, so this is a documented adaptation difference, NOT evidence that deterministic CUDA regressed its oracle. Neither implementation is automatically a more accurate MaleCNS physiology model.

Scope of work allowed without replacing frozen brain dynamics

1. Preserve official code/license at a pinned commit, with hashes; do not execute downloaded code. Extract exact source statements into an evidence table.
2. Synthetic one-cell mathematical fixtures compare the frozen g policy with reset-only, refractory-only, and combined component ablations. Other arithmetic stays frozen. These are NOT a Brian2 reproduction (integration and scheduling also differ), and are not deployed to the MaleCNS brain.
3. Simulate only the UNCHANGED frozen dynamics on actual induced 2/3/5-neuron circuits selected from the already observed Phase 4Z path. One force request at t=0, then no input, for 1,000 ms. Include no-input and real-edge feedforward controls. Missing external network inhibition/feedback is explicitly excluded; a sustained isolated response is not proof of whole-brain necessity.
4. Validate every frozen CPU small-circuit event/state against the unchanged CUDA kernel, using exact float32 ordered accumulation and delay=18 steps.
5. Recheck the five archived deterministic reference trials and all old hashes.

All single-cell amplitudes derive from the complete unique contact-count set in the selected real subgraph, plus a one-contact unit fixture. Do not choose a successful amplitude after observing outcomes. Add a refractory-arrival fixture at t=0/1/2/2.2/2.3 ms and a sustained-input-then-off fixture (100 ms then100 ms). These describe operator consequences, not measured cell physiology or tuning.

The full learning gate still requires >=5 training seeds, independent evaluation, A-vs-neutral-B specificity in MBON/downstream, appropriate controls, acquisition, reversal, retention and delayed-reward tests. Existing transfer PASS is not that.

Next-stage authorization boundary

Testing alternative semantics or measured local physiology on the full MaleCNS network would be a NEW model experiment, not an exact runtime optimization. User confirmation is requested separately. This stage does not silently execute or promote that experiment, even if synthetic ablations reduce firing.

Caveats to record

W_syn=0.275 is the original paper's free parameter selected on its feeding circuit, not a universal measured MaleCNS contact efficacy. g reset by itself does not make the model equivalent to the author's Brian2 implementation. Old reference PASS means consistency with our old oracle, not matching the original author software.