digital-drosophila/models/male-cns-lif/mb-credit-transfer-audit-v1/EVIDENCE.md

# Phase 4AL — credit specificity and conditional KC→MBON transfer

## Purpose and frozen boundary

This is a diagnostic audit of the already sealed Phase 4AK D18 branch, not
conditioning, training, or a new learning candidate. Phase 4AK established its
predeclared 100 ms KC representation gate; it did not establish appropriate
credit assignment, functional associative memory, or biological fidelity of
every omitted mechanism.

Use all 120 Phase 4AK primary trials and all 24 exact-repeat trials. Keep the
two cohorts and their two input modes separate. Repeats check reproducibility;
they are not extra independent observations in the five-seed statistics. Do
not replace seeds, select favorable odors, choose a best pair or compartment,
or inspect old held-out dynamic outcomes to change this design.
The inherited matched-energy mode matches expected input, not necessarily the
realized event count of every stochastic trial; do not relabel it exact
realized-count matching.

All earlier raw, processed, frozen/reference, model and experiment artifacts
remain read-only. Preserve the base CSR, anatomical counts, transmitter signs,
LIF parameters, D18 operator, APL settings, eligibility-v2, and existing 14,232
plastic-edge mask. All new records belong only to this new audit. No DAN
stimulus is injected; no reinforcement is delivered to a brain; no learned
`plastic_delta` is written. No decoder, action selection, game, RL, or visual
interface is introduced.

## Three distinct evidence levels

1. **Exact baseline reconstruction.** Reconstruct target MBONs from the actual
   frozen full-network events using real edges, original arithmetic/order,
   initial states, arrivals, refractory acceptance, resets, and applicable APL
   handling. Require bitwise identity of events, per-neuron counts, and final
   V/g/ref against Phase 4AK D18. A failed or incomplete reconstruction blocks
   scientific interpretation of that replay; it is never rescued by plausible
   plots or by agreement of aggregate firing rates. Complete-audit integrity
   requires all 144 baseline replays; do not silently drop failed cases and
   report the remaining subset as the complete predeclared audit.
2. **Virtual credit audit.** Replay the unchanged, existing float32 eligibility
   operator and inspect its state at the fixed tick-900 virtual gate, after
   that tick's original `mul → add → clamp` sequence. Tick-900 spikes are
   included. This is shadow accounting only, not dopamine delivery or weight
   updating; a before/after-update change is a different experiment. Do not
   treat virtual gate amplitude as measured dopamine concentration.
3. **One fixed conditional sensitivity test.** Compare the exact baseline with
   one temporary replay applying uniform 25% depression to the whole existing
   plastic-edge mask. This uses the pre-existing maximum allowed fractional
   depression as a single declared diagnostic intervention. It is not a
   fraction scan, an eligibility-trained delta, a selected successful memory,
   or permission to enlarge the mask. The base weights remain unchanged; only
   the diagnostic effective amplitudes change, preserving edge identity and
   transmitter sign. On an independent incoming-weight copy, apply
   `weights[plastic] *= 0.75` throughout the full 100 ms replay, not only after
   tick 900. All presynaptic spike trains, including those of other MBONs and
   DANs, remain the original recorded trains. Do not save the overlay as a
   learned or official model.

The 25% intervention is **not an upper bound on functional response**. LIF
threshold, reset, refractory gating, and recurrent influences make responses
nonlinear; the largest permitted fractional depression does not establish the
largest attainable change in spikes, latency, or downstream activity.

## Incoming-drive and overlap accounting

Decompose actual MBON incoming drive into six mutually exclusive groups:

- plastic KC→MBON;
- non-plastic KC→MBON;
- DAN;
- MBON;
- APL;
- other.

Use the frozen canonical IDs, mask membership and audited metadata, not guessed
names or transmitter-to-cell-role substitutions. Preserve unknown role
provenance inside the residual `other` group. Establish each edge's group once;
do not reclassify it to improve the result. Report accepted/rejected arrivals,
signed contributions, and roundoff/reset effects as the implementation makes
available. The model's variable `g` is a model state; do not rename it a measured
biophysical conductance or dopamine signal. A source's share of accepted drive
is not its necessity, sufficiency, or unique causal contribution to a spike.
Any ratio with a zero/undefined denominator must remain explicitly undefined.
The sole set-similarity convention is empty/empty Jaccard = 1 with an explicit
`empty_union` flag. Such pairs are counted separately and excluded from block
overlap aggregates; this convention is never evidence of useful credit.

For each cohort × mode, retain all 15 unordered pairs of its six odors: 60
pair records in total. Report two separate three-of-five-seeds memberships for
each odor: (1) a strictly positive frozen eligibility trace sampled after the
tick-900 update, including tick-900 spikes, and (2) a strictly positive spike
count anywhere in the full 100 ms trial. The primary virtual-credit grouping
uses the first definition. The whole-window activity grouping is auxiliary;
it cannot replace the primary grouping to rescue a result. These are
descriptive reliability rules, not biological thresholds, top-K filters, or
ways to delete cells/edges from replay. Preserve per-trial activity and
eligibility alongside both aggregate memberships.

Compute overlap on the actual A-active and B-active sets. In particular,
`Jaccard(A_active, B_active)` must use their real intersection and union; never
substitute the already disjoint sets `A_only` and `B_only`, whose zero
intersection is true by construction. Shared, A-only, B-only, and inactive
strata are useful descriptive partitions, not proof of stimulus-specific
causation. Shared representations can support biological generalization;
nonzero B effects are not automatically a defect, nor are they automatically
evidence of valid generalization.

Keep arithmetic/integrity checks separate from scientific results. Report all
pairwise and trial-level outcomes without inventing an overall 120-trial
biological PASS fraction, post hoc effect threshold, or pooled rescue gate.
An absence of MBON spike change may coexist with a subthreshold change and
does not mean the entire biological pathway is incapable of plasticity.

## What a conditional replay can and cannot establish

The baseline can establish that the saved events and the frozen operator
reconstruct the saved MBON trajectory exactly. Virtual eligibility can reveal
which real candidate edges would receive credit under that declared rule.
The single depression replay can measure how the selected MBONs respond to
that fixed intervention **with the external network event history held fixed**.

It cannot establish closed-loop whole-brain necessity/sufficiency or learned
behavior. If an MBON's replayed spikes change while its outgoing effects and
the other neurons' events remain clamped to the old trial, MBON→DAN, MBON→MBON,
and other feedback are not recomputed self-consistently. Their fixed incoming
traces must not be described as counterfactual feedback responses. If the
replay instead also clamps the target MBON's original reset/refractory gates,
its result is narrower still: conditional drive/state accounting, not an
independently recomputed functional spike response. State these scopes
explicitly for each reported quantity.

The anatomical D18 arrival delay is 18 × 0.1 ms = 1.8 ms. It is not the
eligibility lifetime, a measured dopamine latency, or the CS–reinforcer
association window. A 100 ms observation does not test memory retention,
consolidation, extinction, reversal, or behavioral recall. Do not promote any
successful replay to a dopamine-plasticity or learning candidate.

## Primary evidence and non-measured assumptions

- **Hige et al., Neuron (2015), DOI 10.1016/j.neuron.2015.11.003.** Pairing an
  odor with selected DAN activation produced odor-specific, compartment-local
  KC→MBON depression with order dependence. The reported induction did not
  require postsynaptic MBON spiking; overlapping KC representations predicted
  reduced specificity/generalization. This supports auditing the correct
  presynaptic and compartment credit, not requiring MBON spikes to define all
  plasticity. It does not specify MaleCNS per-edge learning parameters.
  [Author publication and paper](https://www.janelia.org/publication/heterosynaptic-plasticity-underlies-aversive-olfactory-learning-drosophila).
- **Handler et al., Cell (2019), DOI 10.1016/j.cell.2019.05.040.** DopR1/DopR2
  pathways support opposing plasticity associated with different cue–DAN
  timing. This establishes why time ordering matters, not that the frozen
  float32 trace is a measured biochemical state or that one rule applies to
  every compartment. [Primary paper](https://www.sciencedirect.com/science/article/pii/S0092867419306117).
- **Aso and Rubin, eLife (2016), DOI 10.7554/eLife.16135.** DAN cell types differ
  in memory formation, updating and persistence. PAM/PPL1 labels alone are
  not a license to assign a universal valence or a common temporal rule.
  [Primary paper](https://elifesciences.org/articles/16135).
- **Aso et al., eLife (2014), DOI 10.7554/eLife.04580.** Different MBON types
  bias approach/avoidance, with combined activation supporting population
  integration. Assessing a population response is more appropriate than
  equating greater weight or total firing with positive memory. The current
  audit nevertheless has no behavioral decoder or demonstrated behavior.
  [Primary paper](https://elifesciences.org/articles/04580.pdf).
- **Huang et al., Nature (2024), DOI 10.1038/s41586-024-07819-w.** Physiological
  measurements and a constrained model show heterogeneous, bidirectional
  PPL1 signaling and feedback interactions among memory units. This motivates
  explicit closed-loop limitations. Its fitted model weights and plasticity
  parameters are not MaleCNS anatomical counts or measurements to import into
  this frozen branch. [Primary paper](https://www.nature.com/articles/s41586-024-07819-w).
- **Manoim et al., Current Biology (2022), DOI 10.1016/j.cub.2022.09.007, and
  Manoim Wolkovitz et al. (2026), DOI 10.1016/j.cub.2026.01.014.** KC lateral
  axonal mAChR-B modulation contributes to stimulus specificity; subsequent
  work identifies voltage-dependent modulation relevant to active versus
  inactive KCs. Sparse coding plus a timed reinforcer is therefore not, by
  itself, proof of appropriate credit. These studies do not prove that an
  omitted receptor mechanism caused any result in this fixed model, and do
  not authorize a sign, gain, threshold, or receptor-parameter change.
  [2022 paper](https://www.sciencedirect.com/science/article/pii/S0960982222014518);
  [2026 paper](https://www.sciencedirect.com/science/article/pii/S096098222600014X).

The frozen eligibility operator, tick-900 virtual gate, three-of-five activity
summary, 25% uniform diagnostic depression, clamped event history, and chosen
observation window are engineering/model assumptions. None is relabelled as a
MaleCNS single-cell or synaptic measurement by citing the papers above.

## Deferred functional-memory controls — not executed here

A later explicitly authorized learning experiment would need pre/post CS+
and CS− responses, reciprocal cue assignment, independent evaluation seeds,
and reinforcement-free probes with updates disabled. CS-only, DAN-only,
unpaired/yoked, D(t)=0, and frozen/delta=0 conditions distinguish associative
effects from exposure, transient state, adaptation and unrelated modulation.
Real DAN-only stimulation is not universally expected to have zero effect;
that prediction depends on the declared model and compartment. Retention
requires its own observation interval, and any model-specific output effect
must remain distinct from a behavioral-memory claim. No such controls are
silently replaced by virtual credit or the 25% sensitivity test.

Google/DeepMind's [FlyBody work](https://www.nature.com/articles/s41586-025-09029-4)
provides body physics and artificial RL controllers, not evidence for this
KC→MBON credit rule. Google's [MaleCNS author resource](https://research.google/blog/a-connectomics-milestone-mapping-the-complete-male-fruit-fly-brain/)
supports the anatomical resource, not measured dopamine kinetics or successful
learning in this model. Neither permits bypassing the present mechanism audit.