digital-drosophila/registry/phase4ao-conditioning-interface/source-snapshot/README_DRAFT.md

Olfactory conditioning lab — unregistered preparation

This directory is an independent, unfinished model-branch implementation. Its 21 engineering sources are registered and sealed by Phase 4AN; those files must not be edited. Phase 4AM and the finite Phase 4AN engineering gate passed. The scientific conditioning proposal and additional planning modules remain unregistered, have not run training, and have not produced a learned candidate. Engineering PASS is not learning PASS.

Engineering evidence: registry/phase4an-engineering-validation in the project repository, commit aeccb7d0ccdcf6788e0a8de60e7ab96f70dbe734. Its exact 21-source inventory is in source-identity.json; this README is not one of those registered sources. The experiment manifest SHA-256 is e8952a48cd3ea6a93c811a1183a1f7b79e19646299924e7688e5f393737a9aa2.

Existing code

• runtime_light.py: derives a fixed-weight trial runner from the SHA-pinned AJ source. Its full-recording reference retains the entire original tick loop. The light variant removes an explicit diagnostics whitelist while preserving the ordered mathematical AST. It retains sparse events, stimulus acceptance, terminal neural state and delay queues. It supports an explicit longer admission bound, not hidden time truncation or approximate delays.
• seed_registry.py: explicitly separates training, evaluation and schedule design random streams. Declared same-split aliases can give controls common random inputs; aliases do not count as independent replicates. Duplicate keys, cross-split aliases and seed collisions stop preparation. A full manifest is independently reconstructible.
• runtime_continuation.py: carries full terminal state, spike/APL delay queues and the absolute clock across fixed-weight segments. It protects every ancestor checkpoint against output/build writes. The explicit arrival-admission change and reporting hooks are AST-audited. The finite AN gate passed 32 actual GPU calls and 74 exact comparisons, including pending arrivals across a split. This does not verify every future 120/180 ms training epoch or the unexecuted 1/10 s retention protocol.
• biological_contract.py: binds observed canonical identity, exact DAN/MBON compartment tokens and real graph edges without filling unknown relations.
• trial_plan.py: constructs and independently validates the proposed complete 40-block training/control matrix and explicit random-stream aliases. It has not selected a formal master seed or generated scientific brain inputs. Complete-protocol measured storage and runtime budgets remain unfinished.
• evaluation_plan.py: requires explicit evaluation timing, reversal points and per-control snapshot bindings. It reconstructs the parent inventory and retains 80 independent A/B evaluation streams, paired across snapshots within each training replicate. It does not bind actual state files or authorize a run.
• evaluation_policy.py: supplies the explicit draft policy in EVALUATION_POLICY_DRAFT.md: 100 ms CS, 120 ms probes, seven curve points, actual ordered-CS-prefix control alignment and separate US/time accounting. Its conditional inventory contains 9,440 probes, not executed trials.
• conditioning_inputs.py: consumes a full 63-bit semantic seed in the original discrete float32 Bernoulli CUDA sampler, with explicit device/draw budget and CS window. CPU/mock tests do not establish CUDA random-sequence equivalence or the influence of high seed bits. It never substitutes a CPU random generator.
• conditioning_update.py: uses the SHA-pinned AM/J bridge to compute a single epoch-end delta from supplied actual sparse events. It never installs weights; scheduled D, actual DAN activity, attempted increments and actual changed edges remain distinct. Synthetic validation does not establish a trained brain.
• metrics.py: pure, explicitly draft functional gates; keeps the historical downstream population and the non-MBON difference as distinct readouts. Caller-bound receipts and anatomical membership are still required.
• test_*.py: CPU synthetic/structural checks. These cannot establish real GPU equivalence, biological validity or successful learning.
• EVIDENCE_DRAFT.md and PROTOCOL_DRAFT.md: evidence and protocol proposals; neither is an approved or frozen experiment.

Required before registration or training

1. Preserve completed AM/AN engineering evidence. AM completed its 144 zero and 144 fixed-overlay runs, five old frozen reference trials, and integrity checks. AN's 32-call continuation gate passed its finite exact comparisons and 28-root integrity checks; it did not rerun the five old reference trials. Neither gate executed conditioning.
2. Review and lock the complete experimental matrix, true executed controls, separate training/evaluation stream table, data budget and stop rules. Report any additional gates as new model hypotheses, not inherited facts. The generic seed registry alone does not prove protocol coverage or semantic independence: preparation must require every planned stream and permit aliases only within the same training replicate and input role, for named matched controls/pre-post comparisons. Never alias two training replicates or CS and US roles. Save generator/version, target ordering, rates, timing, consumed stream ID and actual requests hash. Do not truncate a 63-bit seed to 32 bits or derive a new seed by adding the epoch number downstream.
3. Bind all graph arrays, APL bundle, coefficient JSON, base-candidate vector, full 14,232-edge mask, DAN/MBON identity tables and input files to the registered manifest. A runtime's self-consistent spec or source hash is not sufficient authority for these data.
4. Preserve the completed finite real-GPU light/reference comparisons, including exact ordered events, per-neuron counts, terminal V/g/ref, external-request acceptance, APL terminal state and both pending delay queues. No tolerance may excuse a spike divergence. Benchmark encoding, loading, simulation, validation and output time separately and together.
5. Validate the actual retention protocol before claiming retention. The separate continuation module passed the finite AN GPU gate. V/g/ref, spike-ring phase, APL state/pending phase and global tick must survive; segmented and monolithic executions must agree. A checkpoint seal binds fresh state bytes when created; it cannot prove pre-seal immutability.
6. Verify every learned snapshot's parent chain, input events and exact frozen eligibility-v2 update. The single-trial runtime verifies the bounded overlay and parent-hash declaration, but does not prove the coordinator actually computed that update. Within each neural trial the delta is fixed. The original base is never written.
7. Record the realized wall-time/output budget before the full acquisition suite. A failed acquisition gate must not be rescued by tuning eta, thresholds, APL gain, trial count, mask membership or selected odors.

Interpretation limits

PAM15 and PPL103 are specific, statically mapped artificial US interfaces; PAM/PPL cluster names are not general valence labels. The fixed eligibility rule is a single-direction causal-LTD engineering hypothesis, not a complete model of dopamine, bidirectional plasticity, forgetting or consolidation. The proposed fixed MBON type contrast is offline analysis, not a decoder.

No old sealed artifact, raw connectome, transmitter sign, eligibility-v2 or base weight is modified here. No game, reinforcement-learning policy, classifier, action decoder, visual encoder, web service or 3D interface is introduced. A future learning PASS requires functional stimulus-specific responses and executed controls, never merely a changed delta array.