Digital Drosophila Archive
1.7 KiB
Phase 4AG: author LIF + unchanged legacy graded APL
Status: FAIL-FIXED-CAL-REPRESENTATION.
Numerical closed-loop gate: True; original complete fixed six-CAL gate: False. The independent Brian LIF and candidate use independent state but share the APL arithmetic component. Original frozen APL AST replay and a float64 error audit are separately recorded. Author LIF uses float64; APL retains float32, tau20ms/gain0.01/normalization1, full original matrices and order. No learning/parameter search. The explicitly hypothesized unconditional end-APL write includes refractory KC. Original fast APL mask leaves1,212 non-KC edges unreplaced; this scope limitation was not silently fixed. Old checksums and five frozen reference trials: True. Successfully compared cases: [{'case': 'apl_only', 'pass_all': True}, {'case': 'arrival_collision', 'pass_all': True}, {'case': 'cal-matched_energy-00-s51101', 'pass_all': True}, {'case': 'cal-physiological-00-s51101', 'pass_all': True}, {'case': 'high_fanin', 'pass_all': True}, {'case': 'kc_ref_boundary', 'pass_all': True}, {'case': 'natural_request', 'pass_all': True}, {'case': 'pn_pulse', 'pass_all': True}, {'case': 'two_sources', 'pass_all': True}, {'case': 'zero', 'pass_all': True}]. Detailed CAL, exact comparisons, component errors, event/request traces and execution exceptions are linked from audit/final-status.json. Wall/VRAM here include per-tick CPU synchronization, IPC and hashes. They are diagnostic costs, not runtime-performance benchmarks. No heldout dynamic outputs, formal A/B, learning, game, reward, decoder or new candidate. Next gate: Audit first failure in this explicitly declared hypothesis; no gain search, no heldout tuning, no conditioning