Digital Drosophila Archive
82 lines
4.6 KiB
Markdown
82 lines
4.6 KiB
Markdown
# Phase 4AP: exact, process-isolated conditioning interface
|
|
|
|
This independent engineering branch repairs process ownership, not neural
|
|
mathematics. Phase 4AO remains sealed `FAIL-POSTPROCESS-MODULE-IDENTITY`.
|
|
It is not resumed or reclassified. No source in its 51-file inventory changes.
|
|
|
|
## Registered scope
|
|
|
|
The new registration covers 12 trial slots: **one explicitly imported, complete
|
|
AO neural checkpoint and 11 new neural executions**, 14,400 covered ticks and
|
|
13,200 new ticks. The imported slot is D1/reference/epoch-000. Its original case,
|
|
requests, spec and checkpoint paths remain unchanged. Incomplete AO process
|
|
evidence is never replaced by a successful AP process claim.
|
|
|
|
Reuse both sealed AO request arrays byte-for-byte. Zero new sampler calls;
|
|
no seed selection, resampling, rate/parameter rescue or data-selected retry.
|
|
The two A epochs, original full 63-bit seeds, fixed PAM15 compartment, timing,
|
|
D1/D0/frozen arms and reference/light recording modes are inherited exactly.
|
|
New case specs differ only in output budget root/cap, with per-field derivations.
|
|
Each epoch starts a fresh neural state. Only its own arm/mode's preceding delta
|
|
is carried. This does not implement retention-state transition.
|
|
|
|
## Process ownership
|
|
|
|
Each new brain trial runs in a new `python -B` exec process using the frozen
|
|
ordinary runtime. It seals its checkpoint before returning. The separate CPU
|
|
update process loads only the frozen epoch adapter E and its verified dependencies,
|
|
with CUDA hidden before scientific imports. The coordinator never calls E.
|
|
No module binding is deleted, no private cache is cleared, no verifier is relaxed.
|
|
|
|
The imported trial also receives fresh CPU postprocessing into AP, with no brain
|
|
rerun. Its delta raw float32 bits and update receipt must match the separately
|
|
sealed AO isolation diagnostic. NPZ container byte equality is not required.
|
|
|
|
## Integrity and admission
|
|
|
|
Commit all five new source files before preregistration. Bind actual input/source
|
|
and prepared-case SHA-256, prior seals, source commit, actual saved requests,
|
|
and the imported authority. Pin external registration SHA before executing.
|
|
Job SHA is externally passed to each child. Verify input/source bytes before
|
|
and after execution. Bind the complete output hash inventory, checkpoint/chain,
|
|
authority, parent delta, actual child exit status and logs. Exclusive attempt,
|
|
output and receipt paths prevent retry after partial failure.
|
|
|
|
Single physical RTX 4090 UUID GPU-79f3a85d-adce-7833-7003-6488140ca596 only.
|
|
Check all three existing ComfyUI queues before every brain trial; never stop or
|
|
evict their services. Preserve precision configuration. Finite AP output cap
|
|
4 GiB, reserve 10 GiB, child timeout 1,800 seconds. Stop on first failure and
|
|
preserve all partial evidence. Never restart because the observing SSH command
|
|
or tool times out; inspect actual process/log state first.
|
|
|
|
## Pass gate and interpretation
|
|
|
|
- Six reference/light pairs: exact observed spikes, counters, final state,
|
|
delay state and delta; inherited Q gate with independent expected effective
|
|
weight hash computed from frozen base, frozen fast-APL mask and actual delta.
|
|
- Six zero-weight cross-arm comparisons: D0/frozen in each epoch and
|
|
D1/D0 before any update, separately for each recording mode.
|
|
- Original U receipt recomputation, correct compartment/mask/floor, exact
|
|
D0/frozen positive-zero delta and actual D1 nonzero update coverage.
|
|
- Second D1 epoch actually installs its own preceding delta file/raw bits.
|
|
- All previous sealed AO/diagnostic/AN/AM and 28 protected roots remain intact.
|
|
|
|
If all checks hold, status is
|
|
`PASS-EXACT-PROCESS-ISOLATED-DELTA-INTERFACE-NOT-LEARNING`.
|
|
If D1 has no actual nonzero update coverage, report
|
|
`INCOMPLETE-NO-NONZERO-UPDATE-COVERAGE`, not a synthetic success.
|
|
Failures remain failures; do not rerun with alternative inputs.
|
|
|
|
Scheduled D is an artificial unconditioned-signal scalar, not a measured dopamine
|
|
concentration and not multiplied by accepted DAN events. Report requested,
|
|
accepted/rejected US events and actual DAN spikes separately. Unknown anatomical
|
|
mapping stays unknown. The frozen APL approximation and its non-KC limitations
|
|
remain disclosed. No intermediate light-mode state is claimed observed.
|
|
|
|
This finite gate is **not learning success**, an A/B behavior evaluation,
|
|
retention, full acquisition or a learned-model candidate. Full conditioning
|
|
still needs its bounded registered gate/resource plan and validated post-update
|
|
same-state transition. Phase 4C's five trial evidence is preserved via AM seals;
|
|
this gate does not claim it reruns those trials. No game, RL, decoder, visual or
|
|
Web work is included.
|