Digital Drosophila Archive
71 lines
4.1 KiB
Markdown
71 lines
4.1 KiB
Markdown
# AO — bounded two-epoch conditioning-interface validation
|
||
|
||
This independent branch does not replace the frozen reference, runtime or any
|
||
Phase 1–AN result. It tests the installation interface for the existing frozen
|
||
eligibility-v2 rule. It is not a new dynamics hypothesis or a learned model.
|
||
|
||
## Fixed, pre-outcome selection
|
||
|
||
- Master seed: `20260930`; no seed search or outcome-based retries.
|
||
- First canonical block: physiological / PAM15 / reinforce_A / rep01.
|
||
- First normal paired epoch requesting a US, and its immediate successor.
|
||
- Both selected epochs are odor A (4-ethylphenol), not an A/B discrimination
|
||
test. Exact recorded L glomerular/PN mapping and PN column order are retained.
|
||
- Each epoch is 120 ms: 100 ms CS plus 20 ms tail, US at tick 900.
|
||
- D1, D0 and frozen arms receive identical saved CS and DAN requests. The
|
||
scheduled artificial D scalar is not measured dopamine concentration and is
|
||
not multiplied by actual DAN request acceptance. Delivery is reported apart.
|
||
- Two executors (reference and light), two fresh epochs per arm: 12 calls,
|
||
14,400 timesteps. This does not authorize the full conditioning inventory.
|
||
|
||
## Immutable neural semantics and authority
|
||
|
||
Each epoch resets V, g, refractory counters, APL state and both delay queues;
|
||
the zero-initialized eligibility trace is computed from actual sealed spikes.
|
||
Only the float32-mV plastic delta carries between epochs. No update occurs
|
||
inside an epoch. Original graph, base weights, transmitter signs, numerical
|
||
precision, APL parameters, frozen v2 rule and complete 14,232-edge mask remain
|
||
unchanged. Full graph is 166,700 neurons and 25,582,938 edges.
|
||
|
||
Registration pins source snapshots, actual graph/APL/coefficient files, frozen
|
||
L static resources, H mappings, the complete generated semantic plan and the
|
||
12 selected call descriptors before CUDA sampling. Sampling receipts bind the
|
||
two generated request arrays before any brain call. Every input file is shared
|
||
byte-for-byte between arms and executors. The full 63-bit input seed is separate
|
||
from the inherited conservative 31-bit neural reset-control value.
|
||
|
||
A fixed first-seed CUDA check executes only the four original L sampling AST
|
||
statements, compares sparse requests, and compares full63-bit versus low32-bit
|
||
uniform matrices. This is a sampler test, not a neural trial or seed search.
|
||
Two production and three diagnostic matrices are drawn; each is 1200×PN count,
|
||
at most 2,000,000 uniform values. No continuous-time Poisson claim is made.
|
||
|
||
Each neural output is immediately sealed. Outside that seal, the coordinator
|
||
binds original inputs, checkpoint events, actual v2 recomputation and the new
|
||
delta snapshot. `parent_snapshot_sha256` means the previous **overlay NPZ file**,
|
||
not its manifest or the neural checkpoint. Epoch 2 must install that exact
|
||
snapshot; D0/frozen deltas must stay positive-zero bitwise.
|
||
|
||
## Finite gates and reporting limits
|
||
|
||
Required: all 12 processes complete; six reference/light pairs have identical
|
||
sparse events, counts, final V/g/ref and terminal queues/APL; deltas are exact;
|
||
D0/frozen identity holds; epoch 2 actually installs epoch 1 delta. D1 must have
|
||
nonzero actual update coverage or the result is `INCOMPLETE`, never rescued by
|
||
another input. Nonzero delta is interface coverage, not evidence of memory.
|
||
|
||
The light executor does not record intermediate floating states; no all-tick
|
||
bitwise claim is inferred. This validation has only one training replicate, two
|
||
A epochs, no independent behavioral evaluation, and no acquisition/reversal/
|
||
retention success claim. The post-update retention boundary is still unsupported.
|
||
|
||
## Resource and failure policy
|
||
|
||
Use only RTX4090 UUID GPU-79f3a85d-adce-7833-7003-6488140ca596; admit a child only
|
||
when all three ComfyUI queues are empty. Never stop or evict services. Cap new
|
||
output at 4 GiB plus 10 GiB free reserve and each child at 1800 seconds. Preserve
|
||
partial output and stop on any failure; no automatic brain retry. End by
|
||
rechecking AN and AM seals and the inherited 28-root integrity inventory. Five
|
||
old frozen reference results are retained from sealed AM evidence, not claimed
|
||
to have been rerun by this small interface gate.
|