Digital Drosophila Archive
86 lines
5.1 KiB
Markdown
86 lines
5.1 KiB
Markdown
# Phase 4AM — D18 whole-brain plastic interface
|
||
|
||
This is a separate, non-trained branch. All Phase 1–4AL files remain read-only.
|
||
The user authorized independent new-model dynamics hypothesis tests. No base
|
||
CSR, transmitter sign, W_syn, neuron threshold, APL calibration, eligibility-v2,
|
||
or 14,232-edge mask is changed. No game, decoder, RL, or visual input is used.
|
||
|
||
## Fixed test order and scope
|
||
|
||
1. All 120 sealed AK inputs and their 24 predeclared repeats run with exactly
|
||
zero plastic_delta. Compare every recorded state bit, event, count, delayed
|
||
edge queue, request acceptance, and APL generated/applied/pending sample
|
||
against AK. All must pass before any nonzero overlay runs.
|
||
2. The same complete input inventory runs with one fixed diagnostic overlay:
|
||
the original float32 eligibility-v2 floor, `-.25 * base_candidate_mv`.
|
||
All 14,232 mask edges are included; all other weights remain identical.
|
||
Twenty-four technical repeats must be exact. This is not learned memory.
|
||
All presynaptic events are recomputed by the whole brain, unlike AL's
|
||
clamped-source MBON replay. No outcome selects an odor, compartment, dose,
|
||
parameter, or number of training pairings.
|
||
3. Pure-function compatibility tests validate the frozen eligibility trace and
|
||
original additive apply_v2 rule. No reinforcement or weight update occurs
|
||
in any of these GPU trials. Fixed delta is installed only before tick zero.
|
||
|
||
Report both cohorts, both modes, all odors and all seeds. Technical repeats do
|
||
not increase the statistical sample. Matched-energy retains its inherited
|
||
meaning: expected stochastic input, not exactly matched realized event counts.
|
||
Simulation remains 100 ms, including its unadvanced delay/APL tail.
|
||
|
||
## Explicit precision bridge — a new interface assumption
|
||
|
||
The existing D18 brain uses float64 SI voltage weights. Frozen eligibility-v2
|
||
uses float32 mV deltas and float32 candidate base amplitudes. The adapter keeps
|
||
the original D18 base bytes and adds `float64(delta32) * float64(.001)` using
|
||
separate non-fused operations, only on nonzero delta rows. Zero delta must not
|
||
reconstruct the base through float32. Positive original mask weights must
|
||
remain positive. This is not a claim of bitwise nonzero equivalence to the old
|
||
float32 brain, nor of precisely 25% of float64 D18 amplitudes. The base candidate
|
||
is reproduced from the pinned original Phase 4C NumPy expression, not converted
|
||
back from the float64 brain. The mask and all units are recorded.
|
||
|
||
Only the pre-loop weight-install hook is inserted in an AST-derived copy of
|
||
the SHA-pinned AJ function. Its timestep loop AST and original CUDA operators
|
||
remain identical. The old module is not patched and its original __file__ is
|
||
not fabricated. Inherited AJ source metadata describes the parent; additional
|
||
AM adapter provenance explicitly describes the new overlay. The original spike
|
||
delay queue stores spike identities and reads fixed weights at arrival. No
|
||
mid-trial update or implicit weight-snapshot delay rule is introduced.
|
||
|
||
## Interpretation and next experiment boundary
|
||
|
||
Record MBON, KC, DAN, actual MBON downstream, descending and motor output
|
||
changes separately, as well as first divergence and recruited/lost KC cells.
|
||
Static transfer is neither associative acquisition nor a maximal functional
|
||
effect bound. A weak static response does not prove all sparse overlays fail.
|
||
The full-brain response is not compared with AL as a strict arithmetic oracle:
|
||
AL clamps feedback and its float64 0.75 multiplication is a different fixed
|
||
precision intervention. Any side-by-side descriptive result states both.
|
||
|
||
Conditioning must be separately preregistered. Keep the original 4L static
|
||
ODOR_A (4-ethylphenol) and ODOR_B (3-hexanone), reciprocal cue assignments,
|
||
type/compartment-specific US and frozen eligibility-v2. PAM/PPL1 alone never
|
||
defines valence. The original D scalar is an artificial unconditioned-signal
|
||
interface, not measured dopamine or the number of actually emitted DAN spikes.
|
||
Log requests, acceptance, and actual DAN spikes separately. Multiplying D by
|
||
actual accepted spikes would be a different explicitly registered hypothesis,
|
||
not an unnoticed repair of eligibility-v2. No such change is made here.
|
||
|
||
Later functional learning requires independent evaluation seeds, actual paired
|
||
and control trajectories, MBON and downstream specificity, reversal and real
|
||
time retention. Neither file validity, delta change nor a static response
|
||
completes the learning goal. This stage creates no learned candidate.
|
||
|
||
## Integrity
|
||
|
||
Exclusive-create outputs, a pre-outcome source/input/config manifest, fixed
|
||
trial inventory, unit-test records, complete per-trial recording hashes,
|
||
pre/post frozen integrity, and the five unchanged deterministic-v2 reference
|
||
trials are required. Partial outputs are preserved and never overwritten.
|
||
No queued ComfyUI job is evicted. Only the declared RTX 4090 is used. The old
|
||
references are regression checks of the old runtime, not proof of D18 biology.
|
||
|
||
The biological limitations and primary-paper evidence are inherited unchanged
|
||
from sealed AL EVIDENCE.md; this precision/interface implementation introduces
|
||
no new single-cell physiological measurement or universal dopamine rule.
|