Digital Drosophila Archive
82 lines
4.9 KiB
Markdown
82 lines
4.9 KiB
Markdown
# Phase 4AK — frozen D18 validation, not learning
|
||
|
||
Prerequisite: the entire Phase 4AJ fixed-CAL experiment was completed, sealed
|
||
and published at cb6937666f96b00ae8d435a2cfb2f417ad25c040.
|
||
Its PASS is not evidence of usable learning. This stage tests the next necessary
|
||
representation-generalization gate without adjusting the successful branch.
|
||
|
||
## No new brain dynamics
|
||
|
||
Call the exact SHA-pinned AJ GPU, Brian oracle, comparison and timing code from
|
||
their original absolute paths. Do not copy or patch their numerical functions,
|
||
globals, signs, weights, thresholds, APL gain/tau, refractory or delay queue.
|
||
Keep dt=0.1 ms, anatomical delay=18 steps, and whole generated graded-APL h
|
||
applied 18 steps later. Preserve positive-zero initialization and pending tail;
|
||
never advance brain time to drain the tail. Shared APL arithmetic limits the
|
||
oracle independence claim. The 5845 fast-APL output mask, including 1212
|
||
unreplaced non-KC edges, is unchanged. No biological measurement claim.
|
||
|
||
## Input registration and data boundary
|
||
|
||
Read only the four SHA-pinned original static 4L encoder files (panel, vectors,
|
||
selection, mapping) and pinned encoder configuration/source. Extract only the
|
||
unchanged pn_drive function AST; do not import or execute unrelated learning
|
||
experiments. Validate the extraction and request generator against all 60 old
|
||
AG CAL request/PN-drive arrays before generating any new brain trajectory.
|
||
Keep the original 12-odor median 2.556107997894287 and original target/RNG column
|
||
order. Matched energy is expected input, not exactly equal realized events.
|
||
|
||
Predeclare both 6-odor x2-mode x5-seed blocks, every sparse request schedule,
|
||
source hash, gate, and stop rule before any new brain trajectory. Seeds are the
|
||
fixed first five integers above 2^30, subject to the enumerated static collision
|
||
audit. They are not chosen using simulation outcomes. Do not claim exhaustive
|
||
historical uniqueness, statistical independence of RNG streams, or biological
|
||
replication. Same seed across odors preserves the existing paired comparisons.
|
||
|
||
The second six odors are novel to AG/AJ calibration, not never simulated in
|
||
project history: old 4L traversed the full panel. Never open old heldout dynamic
|
||
results, mixed old result tables or conditioning/plasticity reports to choose
|
||
inputs. Full frozen-integrity hashing may read opaque bytes for checksums; that
|
||
is not permission to parse those outcomes. Static A/B identities are not used
|
||
to select a favorable subset. No new resource download.
|
||
|
||
## Ordered execution
|
||
|
||
1. Freshly run the two original AJ D18 fullgraph probe inputs. Verify every
|
||
scientific output against sealed AJ, allowing only output budget/path
|
||
differences in spec. Do not fake D0 to use a D0-only helper. Old eight-fixture
|
||
proofs are source-pinned inherited evidence, not newly rerun fixture claims.
|
||
2. For original CAL odors on new seeds, fixed first odor/first seed in each mode
|
||
is a Brian fullgraph probe, with independent LIF/own spikes/FIFO and the
|
||
existing shared-APL component audit. Any numerical failure stops.
|
||
3. Complete the 60 seed-validation trajectories plus 12 exact repeats. Reuse
|
||
the two probe GPU runs rather than duplicate them as independent samples.
|
||
4. Only if the entire seed block passes numerical, repeat and original
|
||
representation gates, unlock the second block. Run the first static novel
|
||
odor/first seed in each mode as two additional Brian probes, then complete
|
||
that block's 60 trajectories and 12 repeats. All planned conditions remain
|
||
visible; gated unexecuted cases are NOT PASS and are reported as such.
|
||
5. Recompute raw evidence, rerun the five old frozen reference trials into new
|
||
output, verify every frozen artifact, and seal a truthful result.
|
||
|
||
Each cohort/mode separately retains every original six-part 4L gate: mean KC
|
||
fraction in [3%,9%], all runs >0 and <50%, same-odor Jaccard > different-odor,
|
||
between-odor Euclidean > within-odor, and at least 3 of 15 odor pairs passing
|
||
their own two-odor within baseline. Every mode has 60 within comparisons and
|
||
75 between comparisons; each pair has 20 within and 5 between. No pooled
|
||
12-odor rescue, seed replacement, post-hoc metric/window/gate changes or scans.
|
||
|
||
## Outcome scope
|
||
|
||
PASS-FROZEN-D18-INDEPENDENT-VALIDATION requires all four cohort/mode gates,
|
||
complete data, fresh controls, oracle/component checks, exact repeats and old
|
||
frozen integrity/reference PASS. Scientific failure remains FAIL; incomplete
|
||
execution is not re-labelled scientific success. Passing does not generate a
|
||
candidate, establish learning, prove delay biology or repair upstream AL loops.
|
||
Next required gates remain credit specificity, real KC-to-MBON functional
|
||
transfer, whole-brain stimulus-specific response, and controlled conditioning.
|
||
|
||
All outputs are new under phase4ak-apl-timing-validation. Previous Phase 1–4AJ,
|
||
base CSR/W_syn/signs, eligibility-v2 and the 14232-edge plastic mask remain
|
||
read-only. No training, reward interface, game, vision, decoder, RL or Web.
|