digital-drosophila/models/male-cns-lif/olfactory-conditioning-lab-v1/PROTOCOL_DRAFT.md

# DRAFT — fixed-v2 olfactory conditioning in the independent D18 branch

**Status: DRAFT / NOT REGISTERED / NOT APPROVED FOR EXECUTION.**

This document is a proposal for review. It does not authorize a run, record a
successful gate, freeze a protocol, or generate a learned candidate. It was
written from frozen H/J/L source and the declared AM interface contract, not
from selecting favorable AM outcomes. The coordinator must approve the final
scope, budget, source/input inventory, and exact gate implementation before
registration. No training or remote action is performed by this document.

## 1. Purpose and entry conditions

The next scientific endpoint should be actual acquisition under the unchanged
TemporalEligibility-v2 rule, followed, if its registered acquisition gate passes,
by fixed reversal, delayed-reward, and real-time retention assays. File validity,
static 25% sensitivity, or a change in delta alone is not learning.

Entry requires the **complete** AM whole-brain interface gate, including its
registered zero-overlay comparisons, nonzero internal/repeat validation, old
reference regression, and frozen integrity checks. A partially completed zero
suite is not this gate. Continue to use the independent D18 branch; do not edit
or overwrite Phase 1–AM artifacts. The original base CSR, anatomical counts,
transmitter signs, W_syn, LIF coefficients, APL calibration, D18 timing,
14,232-edge candidate mask, and eligibility-v2 remain unchanged.

New long-run/continuation plumbing must preserve the exact existing timestep
operation order. Validate zero-delta and fixed-delta continuation, including a
split across pending synaptic/APL arrivals, against an unsplit run before using
it for retention. This is an implementation prerequisite, not another tunable
scientific endpoint. No game, decoder, RL, visual branch, or brain-gain search
is in scope.

## 2. Fixed population and replicate matrix

Use the original externally selected 4L odors without re-selection:

* ODOR_A: **4-ethylphenol**.
* ODOR_B: **3-hexanone**.
* Both frozen encoder modes: `physiological` and `matched_energy`.
* Both reciprocal assignments: A reinforced/B neutral, and B reinforced/A
  neutral.
* Both exact DAN/type-compartment branches below.
* Five independent training replicates in each cell.

Thus there are **2 modes × 2 DAN branches × 2 assignments × 5 replicates =
40 blocks**, or eight experimental cells with five training replicates each.

| Branch identity | Exact raw compartment string | Proposed signed prediction |
| --- | --- | --- |
| PAM15 | `y5B'2a` | S > 0 |
| PPL103 | `y2a'1` | S < 0 |

These identities are drawn from old H/J configuration, not from a rule that all
PAM is appetitive or all PPL1 is aversive. Preparation must materialize and hash
the actual DAN body IDs/neuron indices, exact annotation values, confirmed
DAN→MBON/compartment mapping, and affected existing plastic edges from the
sealed H mapping. Missing, ambiguous, or unmatched mappings remain `unknown`;
they cannot be completed by cell-name inference.

The type-level signed predictions above are **new, falsifiable engineering
hypotheses**. Evidence specifically establishing PAM15's proposed functional
direction is limited. Correspondence of type labels across fly datasets is not
proof that these individual MaleCNS bodies have been functionally validated.
Neither branch is labeled a biologically demonstrated reward/aversion channel.

### Seed inventory

Before outcomes, generate and save the complete RNG lineage and explicit seed
table for replicate, cell, arm, epoch, CS events, US ordering, and evaluation.
Use distinct replicate streams, for example an explicitly registered
`SeedSequence` hierarchy with a newly audited master seed. Test expanded seed
identities for collisions. Merely choosing five different base integers is
insufficient: H/J/L used consecutive bases plus the trial number, producing
overlapping nominal replicate streams.

Matched counterfactual arms may deliberately reuse the same input event table;
list each such common-random-number alias. They are paired controls, not extra
independent replicates. All evaluation RNG streams must be disjoint from all
training/US/control RNG streams. Within a replicate use the same registered
evaluation events for pre/post/control comparisons. If more than one evaluation
stream is used per replicate, report/aggregate it within that replicate;
**five evaluation streams never substitute for five independent training
replicates**. The minimal budget below uses one independent A/B evaluation
stream pair per training replicate, reused for paired contrasts.

## 3. Frozen update and explicit US interface

Use the exact frozen J/L v2 implementation and operation order:

* KC trace float32, tau = 100 ms, cap = 1.
* Every timestep: separate multiply, add actual KC spikes, then clamp.
* At the registered US tick, sample after that tick's multiply/add/clamp;
  include its KC spikes. Do not insert an anatomical-delay shift into eligibility.
* eta = 0.00275 mV, compartment polarity = −1.
* Additive update only on the already-confirmed compartment mask, with the
  old float32 floor `−0.25 × base_candidate_mV`.
* Update once at the end of the training epoch; it affects subsequent epochs,
  not an earlier part of the same epoch. No recovery, potentiation, or new mask.
* Install via the verified AM bridge: immutable base SI float64 plus the
  float32 delta widened to float64 and converted from mV to V. Preserve all
  untouched weights and fast-APL handling exactly.

The first proposal preserves the old **scheduled dimensionless D = 1** artificial
unconditioned-signal interface. It is neither measured dopamine concentration
nor an actual-DAN-spike count. At the same scheduled tick, request spikes in the
fixed matched DAN population. Record requested, accepted, rejected, and actual
DAN spikes, including zero-delivery epochs, separately from D. Do not resample,
boost, or silently discard failed delivery to obtain a preferred result.

Multiplying D by actual acceptance, changing D to a spike count, or adding a
measured dopamine kinetic model would be a distinct, explicitly registered
interface hypothesis; none is hidden in this draft. If neural DAN delivery is
absent, any update still demonstrates scheduled artificial gating, not that
actual DAN activity caused plasticity. This limitation must remain visible even
if an engineering output gate passes.

## 4. Epochs, timing, and all actual controls

Training epochs retain the old trial-boundary reset policy: start V/g/ref,
spike-delay queues, APL state/queue and KC trace from their declared initial
values; carry only plastic_delta between epochs. Log the state policy. This is
not continuous natural behavior. Retention deliberately uses the different,
explicit continuous-state policy in section 7.

Normal epochs last 120 ms. CS occupies 0–100 ms, and a paired US is at 90 ms.
Delayed epochs last 180 ms; the identical CS occupies 0–100 ms, and US is at
150 ms, i.e. **50 ms after CS offset** and 60 ms later than the normal US.
All delayed controls use that same 180 ms duration and 150 ms US slot, including
the neutral epochs; do not reuse 90 ms controls to attribute delayed effects.

The frozen L encoder mapping/rates and original A/B identity are retained.
However, old `phase4l.py::run_rates` generated PN events throughout its passed
120 ms duration. The present 100 ms CS plus a no-CS tail is an explicit new
experimental timing choice aligned with validated AK input duration, **not** a
claim of bitwise identity to old L's full training input. `matched_energy`
matches the frozen encoder's expected total rate, not an artificially edited
actual Poisson event count; report actual requests and acceptance by condition.

Within each arm, present 50 reinforced-cue epochs and 50 neutral-cue epochs in
a pre-generated balanced order. The **50 reinforcement opportunities** come
from L's fixed formal conditioning count, not from a new magnitude search.

| Arm | Actual execution in each normal or delayed suite |
| --- | --- |
| paired | 50 CS+ with its matched DAN/US and 50 CS− without US |
| D=0 | Identical input events and US schedule to paired; D=0 in updates |
| frozen | Identical input events and US schedule; no delta updates |
| random reward | Same 50 A/50 B and 50 US opportunities; exactly 25 A and 25 B get US, selected by a pre-generated random balanced schedule independent of responses |
| unpaired | Same 100 CS epochs and same 50 US events in 50 additional separate, no-CS epochs; preserve total event inventory, declare 150 rather than 100 epochs |
| CS-only | Same 100 CS epochs, no DAN/US; no scheduled D |
| DAN-only | Same 50 US epochs and 50 matched blank epochs, no CS |

Every control is genuinely executed; neither identity reasoning nor a text label
replaces its trajectory. The unpaired arm is specifically **separate-epoch,
reset-trace unpairing**, not a claim about a long continuous biological
interstimulus interval. Actual DAN-only network activity may recruit KC and
produce eligibility; it must be measured, not assumed zero. Blank and US-only
epochs retain the corresponding 90/150 ms timing slot and 120/180 ms duration.

## 5. Offline signed readout — do not discard contrary signs

Preparation defines complete, exact-annotation MBON12 and MBON01 populations
from MaleCNS metadata, with all body/index membership and side information. No
best responding neuron, hemisphere, compartment subset, or evaluation pair is
selected after outcomes. Missing either population makes this specific readout
`NOT-TESTABLE`; do not substitute another type or silently use an empty mean.

For an evaluation stimulus X and a fixed 0–120 ms probe window, define:

`B(X) = mean_spike_count(MBON12, X) − mean_spike_count(MBON01, X)`.

This is **offline analysis only**, not an action decoder, a feedback signal,
or proof of approach/avoidance. Also save both population means separately,
their sizes and counts, all-MBON output, genuine downstream output, and latency.

Keep two distinct downstream inventories. The frozen H actual-edge table's
complete target union is `legacyDownstream` (11,342 neurons in the verified
local data), and contains **all 97 MBONs**. Do not relabel or silently edit that
old population. Define `nonMBONDownstream` as the exact set difference of that
union and the complete MBON population (11,245 here). Report both. Requiring
an effect outside the MBON population is an additional pre-outcome hypothesis,
not an already inherited H gate; it prevents a purely MBON-local change from
being counted twice as evidence of broader propagation.

With C denoting that block's reinforced cue and N its neutral cue:

`S_arm = [B_arm(C) − B_arm(N)] − [B_pre(C) − B_pre(N)]`.

Always emit the raw signed B and S, pre/post components and control S. For PAM15
the proposed relation is **S > 0**; for PPL103 it is **S < 0**. Never replace raw
S in the report with `abs(S)`, `polarity × S`, a magnitude-only score, or a
renamed “positive learning score.” An opposite sign is a counterexample to the
prediction, even if a magnitude or unsigned separation statistic increases.

The expected-control relation is equally explicit:

* PAM15: `S_paired > S_control` for each registered control.
* PPL103: `S_paired < S_control` for each registered control.

Use matched evaluation inputs and each arm's matched pre baseline. All probe
updates and US inputs are off. Evaluation begins from the same frozen initial
state for acquisition/reversal/delayed comparisons; retention uses its saved
continuation state and time-matched frozen control as specified below.

## 6. Acquisition and strict reversal gates

### 6.1 Acquisition — newly proposed signed gate

For each of the eight cells, a successful training replicate must satisfy all:

1. The raw S has the registered direction (PAM15 >0; PPL103 <0).
2. It lies in that direction beyond **each of the six actually executed
   timing-matched controls**, not just the easiest control.
3. The inherited L functional requirement holds: conditioned-cue MBON
   `Delta_g_max_abs` exceeds neutral-cue `Delta_g_max_abs`, and the conditioned
   cue has a nonzero MBON spike-count or first-latency change versus its own pre
   probe. Undefined latency is separately reported, not silently zero-filled
   into proof of a latency effect.
4. The inherited H MBON and complete `legacyDownstream` A/B Euclidean
   separation gains are positive. Independently require positive separation
   gain in `nonMBONDownstream` as a newly declared specificity/transfer gate.
   Report all three underlying count vectors and gains, never select whichever
   downstream definition is more favorable after outcomes.

Require **at least 3 of the 5 independent training replicates to satisfy the
entire conjunction in the same replicate**. Also calculate the inherited H
mean-control rule separately. This exact new combined signed gate is a proposal,
not a claim that H/L already implemented it.

Record acquisition at fixed update counts 0, 1, 2, 5, 10, 20 and 50. Only the
registered terminal count 50 determines its gate. Do not stop at the best
earlier count or add pairings after failure.

### 6.2 Reversal — new hypothesis, not H's permissive sign test

After normal acquisition, fork its exact delta into two arms:

* **Reversal:** 50 pairings of the former neutral cue with the same specific DAN
  branch, and 50 formerly reinforced cue epochs without US.
* **No-reversal continuation:** 50 further original-cue pairings and 50 original
  neutral epochs, with matched epoch totals and registered independent input
  streams. This is an additional real control, included in the budget.

Never reset the learned delta between acquisition and reversal; do not restore
depressed old-cue edges. Keep the old epoch-reset neural state policy in both
arms and evaluate from common fresh states. The pure-LTD/no-recovery rule may
prevent reversal; that is a falsifiable model limitation, not permission to add
potentiation, recovery, or more trials.

For strict reversal, retain the original C−N orientation when calculating
`S_acquisition`, `S_reversal`, and `S_continuation` against the **same original
pre-training baseline**. At the fixed terminal reversal count 50 require:

| DAN branch | Acquisition | Reversal | No-reversal continuation |
| --- | --- | --- | --- |
| PAM15 | `S_acquisition > 0` | `S_reversal < 0` | `S_continuation > 0` |
| PPL103 | `S_acquisition < 0` | `S_reversal > 0` | `S_continuation < 0` |

Additionally require a nonzero change in the former neutral cue's MBON spike
count or defined first latency between acquisition and reversal, and a nonzero
change in its actual-downstream event/count output. Save both cue vectors;
counts alone can conceal a shift confined to the original cue. The same
replicate must satisfy all conditions, with a 3/5 majority per cell. These are
**new terminal reversal hypotheses**, fixed before the learning results.

Zero is failure of a strict sign-crossing gate, not reversal. Report H's original
`sign(acquisition) != sign(reversal)` separately as a legacy statistic; its
nonzero→zero success case does not satisfy this new gate. Do not claim that
these engineering sign changes establish reversal of real appetitive/aversive
behavior.

## 7. Delayed reward and genuine retention gates

### Delayed-reward test — new hypothesis

Run delayed acquisition from zero delta, using the same fixed 50 reinforcement
opportunities and all six controls in the **180 ms / US-at-150 ms** suite. Sample
the unchanged trace at 150 ms after that tick's update. Do not extend tau or
redefine the trace to force delayed learning.

At count 50, require the same signed, functional, downstream and control
conjunction as section 6.1, with at least 3/5 independent training replicates per
cell and raw S direction intact. This explicitly tests whether the unchanged
rule supports this particular delayed schedule. H did not supply an independent
delayed-reward success threshold; this is a new pre-outcome gate.

Report delayed and immediate signed S side by side and their per-replicate
difference. Do not require a invented retained percentage, impose a monotonic
amplitude ratio, pool their seeds, or call an opposite delayed sign a successful
larger absolute response. Delayed controls cannot be substituted by normal ones.

### Retention — new continuous-state implementation and gate

Fork the terminal normal-acquisition checkpoint and the matched terminal normal
frozen-control checkpoint. Each includes V/g/ref, pending spike-delay state,
APL state, the complete D18 APL queue, absolute tick, delta, and relevant RNG
state. Run real no-input/no-update dynamics continuously to 1 s and 10 s; the 1 s
checkpoint is on the same idle trajectory used to reach 10 s. Do not reset,
discard pending events, analytically decay only delta, or fabricate elapsed time
by changing a seed.

At 0, 1, 10 s, fork independent A/B probes from each checkpoint; probes do not feed
back into the idle continuation. Use the same registered evaluation streams
for trained/frozen comparisons at every elapsed time. The probe lasts 120 ms,
without US or updates. Keep delta constant throughout idle; the model does not
contain molecular consolidation or recovery, and a pass cannot prove either.

Define signed `S_trained(t)` and `S_frozen(t)` using the original C−N orientation
and their registered pre baselines. Require, in the **same replicate at both 1 s
and 10 s**:

* PAM15: `S_trained(t) > 0` and `S_trained(t) > S_frozen(t)`.
* PPL103: `S_trained(t) < 0` and `S_trained(t) < S_frozen(t)`.
* Conditioned-cue MBON output has a nonzero spike-count or defined first-latency
  difference from the time-matched frozen probe; actual-downstream event/count
  output also differs. Save neutral-cue differences and the complete signed
  specificity contrasts so a generic global response is not presented as memory.

Require 3/5 independent training replicates per cell, without a fitted amplitude
tolerance or an arbitrarily chosen “percent retained.” This is a new functional
retention hypothesis, not the old H reset-and-retest result. Report an optional
predeclared fresh-state/persistent-delta probe separately as a weight-storage
diagnostic; it cannot satisfy the real-time gate.

## 8. Legacy gates, new hypotheses, and candidate limits

Retain the following old predicates verbatim as separately labeled audit fields
where their data definitions apply:

* H: minimum seed majority 3; positive MBON and downstream Euclidean separation
  gains; the smaller mean gain of its two DAN branches exceeds the maximum
  mean gain among D0/random/frozen controls.
* H: legacy reversal sign-inequality and last-retention separation>pre. Disclose
  the old retention had no actual idle dynamics and is not direct validation of
  the new real-time protocol.
* L: conditioned `Delta_g_max_abs`>neutral plus a functional spike/latency change,
  in at least 3/5 seeds. Preserve the exact old implementation's legacy result as
  well as the new explicitly defined handling of missing latencies.

There was **no completed inherited delayed-reward gate**, no proven signed
MBON12−MBON01 valence mapping for these bodies, and no complete L reversal or
retention assay. The strict signed gates in sections 6–7 are new proposals, not
historical passes and not thresholds obtained from AL/AM responses.

For the strongest proposed engineering candidate claim, all eight cells must
pass their registered acquisition and subsequent assay gates, old required
integrity/control gates, and numerical safety checks. This all-cell conjunction
is itself an explicit proposed scope of claim; it is not a pre-existing H/L
aggregate gate. Emit a complete 8-cell×assay table, including contrary signs,
failures, `NOT-RUN`, `NOT-TESTABLE`, and actual delivery limitations. Do not pool
one mode/compartment/assignment to rescue another, or silently retain only
successful cells as the candidate.

Suggested staged execution: run normal acquisition and all six normal controls
for all 40 blocks first. If the approved overall acquisition gate fails, seal a
real acquisition failure and stop, recording subsequent assays as `NOT-RUN`.
Otherwise execute the already registered complete delayed, reversal,
continuation and retention suites. No gate changes, tuning or new pair choices
occur between stages. A scientific failure is not a reason to delete evidence;
an integrity/numerical/resource failure stops execution without a scientific
success claim.

Even a full engineering pass would not establish odor preference behavior,
motor learning, natural dopamine concentration, appetitive/aversive valence,
long-term memory, or readiness for a game. Delta change alone, a file/checksum
pass, static-overlay sensitivity, or signed output readout alone is insufficient.

## 9. Transparent worst-case runtime and data budget

The original estimate of 38,000 epochs omitted the new no-reversal continuation
and a full timing-matched delayed-control suite. They cannot be silently added
without revising the budget.

| Per block | Epochs | Biological seconds |
| --- | ---: | ---: |
| Normal paired + five 100-epoch controls + 150-epoch unpaired | 750 | 90 |
| Delayed paired + five 100-epoch controls + 150-epoch unpaired | 750 | 135 |
| Reversal | 100 | 12 |
| No-reversal continuation | 100 | 12 |
| Total training/control/continuation | 1700 | 249 |

Across 40 blocks this is **68,000 epochs and 9,960 biological seconds**. Real 10 s
idle progression of trained and frozen retention branches adds 800 s, totaling
**at least 10,760 s biological time** (107.6 million dt=0.1 ms steps), before
evaluation probes, implementation validation, and old regressions. Checkpoints
at 1 s are reused within the 10 s trajectory, not separate additional 10 s runs.

Relative to 38,000: add 4,000 continuation epochs and 26,000 delayed-control epochs.
All delayed neutral/control epochs are 180 ms, not shortened to 120 ms. The table
includes this duration correction. Any approved reduction of scope must be
decided and documented before registration, not after results or a resource
overrun.

Let measured AM walltime per 100 ms be `t100`. A preliminary linear conversion is
`107,600 × t100`, plus input generation, updates, evaluation, I/O and validation.
If `t100=1 s`, this base is 29.9 h; if 5 s, 149.4 h. These are arithmetic scenarios, not
claimed measured performance: learning may change spike load. Measure and
reserve a conservative actual budget before registration; enforce a resource
stop rather than changing the protocol during a run.

Full generated+applied APL traces alone use approximately
`2 × 4064 × 4 × 10000 = 325,120,000 bytes/biological-second`, hence **more than
3.18 TiB** for the above minimum. Do not silently carry the small AM experiment
disk budget into this suite. Per-epoch full neuron counts also total about 45 GB
at 68,000×166,700×4 bytes, before events or full checkpoints.

The proposed recording policy to approve before registration is: all spike
events, actual input/delivery logs, per-epoch counts, signed metrics, every delta
update and full terminal V/g/ref/APL/queue state. Every timestep retains a GPU
finite-value latch and first-failure tick, **not a CPU copy/hash of all state**.
At epoch boundaries hash the files and bind the complete chain of parent,
input, delta-before, events/state/queues, eligibility/D and delta-after. Full
APL/per-tick reference recordings are restricted to a fixed pre-outcome
validation inventory, not a result-selected subset. That subset's exact
comparison must not be described as per-tick regression of all 68,000 epochs.
Compute capacity from measured event/log
sizes plus a stated safety reserve, and register an exclusive-output storage
cap. Do not publish a guessed final compressed size or omit encoder/update/I/O
latency from walltime accounting.

The frozen L/J metric called `MBON_g_max_abs` is the maximum **across neurons
of the terminal post-minus-pre g difference**, not a temporal maximum. Save
`MBON_final_delta_g_max_abs` with explicit SI units and the legacy alias when
needed. Complete terminal g and sparse events suffice to reconstruct that
metric; adding a dense time-by-neuron recording is not required by its name.

Even without full histories, all 68,000 terminal state/ring payloads would use
roughly 507 GB uncompressed, before events and evaluation. Use lossless
compression and measure real complete-episode outputs before setting the hard
storage cap. Do not infer final compressed usage from sparse activity alone.

## 10. Evidence, provenance, and unresolved biology

Local source basis (read-only, relative to the project):

* `23-ai-gpu/source/models/male-cns-lif/phase4h.yaml` and `phase4h.py`:
  exact DAN configuration, old control/majority predicates, old nominal retention.
* `23-ai-gpu/source/models/male-cns-lif/phase4j.yaml` and `phase4j.py`:
  frozen float32 trace order and `apply_v2`.
* `23-ai-gpu/source/models/male-cns-lif/phase4l.yaml` and `phase4l.py`:
  original odor mapping, 50 formal pairings, input timing, and acquisition gate.
* `23-ai-gpu/source/wholebrain-plastic-interface-v1/PROTOCOL.md`:
  unchanged-base bridge, scheduled-US semantics and later-learning boundary.

Relevant primary evidence, with narrow use:

* [Hige et al., Neuron 2015, author institution](https://www.janelia.org/publication/heterosynaptic-plasticity-underlies-aversive-olfactory-learning-drosophila):
  compartmental KC→MBON plasticity and odor-specific effects motivate local LTD;
  this does not measure this MaleCNS model's eta, tau or DAN concentration.
* [Aso and Rubin, eLife 2016, author institution](https://www.janelia.org/publication/dopaminergic-neurons-write-and-update-memories-cell-type-specific-rules):
  cell-type-specific DAN learning/update rules oppose a universal cluster-valence
  assumption. They do not functionally identify every MaleCNS body here.
* [Aso et al., eLife 2014](https://elifesciences.org/articles/04580):
  MBON population outputs motivate population-level readout. The exact
  MBON12−MBON01 signed hypothesis and the two predictions in this draft remain
  explicit cross-dataset/model assumptions, not a demonstrated behavioral scale.

The exact scalar D interface, fixed trace/eta/floor, absence of recovery,
trial-boundary resetting, simplified signed readout, and 100/120/180 ms timings
are declared model/engineering choices. In particular, the 1.8 ms anatomical
synaptic delay is not an eligibility time constant. Physiological applicability,
PAM15 functional direction in these bodies, continuous intertrial learning,
behavioral valence and long-term retention remain unresolved.

Before formal registration the coordinator must approve and hash the completed
population/seed/input inventory, new readout/gate code, exact continuation
semantics, storage policy/cap, resource budget and final candidate wording.
Until then this remains **DRAFT — NOT REGISTERED — NO EXECUTION AUTHORITY**.