Digital Drosophila Archive
18 lines
1.3 KiB
Markdown
18 lines
1.3 KiB
Markdown
# Phase 4AJ — fixed graded-APL timing branch
|
|
|
|
Status: `PASS-FIXED-CAL-APL-TIMING-HYPOTHESIS`.
|
|
|
|
Numerical gate: True. Complete both CAL arms: True.
|
|
Original representation gate: delay0=False; delay18=True.
|
|
Only generated h application time changes: zero versus 18 steps. No brain gain, input, learning or mask change.
|
|
This is an engineering scheduling hypothesis, not measured MaleCNS APL delay. It delays PN- and KC-derived h together.
|
|
Finite stored generated samples are preserved in applied output or pending tail; the brain is not advanced to flush the tail.
|
|
Closed-loop generated h and accepted requests can differ between arms; equal total inhibition is not assumed.
|
|
Independent Brian LIF and candidate share legacy Torch APL arithmetic; component and separate delay-queue audits qualify this limit.
|
|
Original fast APL mask still leaves 1,212 non-KC edges unreplaced. No silent scope change.
|
|
The origin audit identifies timed real-edge antecedents, not unique or necessary biological causes.
|
|
Frozen checksums and five reference trials: True.
|
|
Costs include diagnostic synchronization, IPC, hashes and tracing; these are not production runtime benchmarks.
|
|
No heldout trajectories, conditioning, learned model, reward interface, game, decoder or candidate.
|
|
Next gate: Independent seeds and novel-odor representation; then credit specificity and functional learning
|