research: preserve reviewed three-direction checkpoint #191
Loading…
Reference in a new issue
No description provided.
Delete branch "research/session-checkpoint-20261025"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Records the other three SMC directions without claiming completed model studies. Links PR190 unchanged. D2 passed a two-seed initial-acquisition gate (population accuracy0.85, learned emissionNLL0.688–1.030), but counts had lower NLL and fewer sampled wrong repairs both seeds; no16-seed or plasticity/relearning panel ran. D3 is an analytically justified park, not a general refutation. NCA passed6fixtures and3fresh zero-W2 kernel profiles, with no training/pool update or trained repair result.
Independent saved-evidence review passed49732assertions plus6selected NCA reference checks. It reconstructed896saved D2 actions, causal labels, counts, endpoint MLP inference and costs without production/scorer imports; no SGD replay or new panel/profile. Review processCPU15s plus20s reserve within60s; source/evidence hashes and failures/cleanup ledgers retained. D2's129.4375s is conservative reservation/debits, NOT measured consumption; full end-to-end actual CPU unmeasured. Historical misleading measured120 field remains immutable with explicit correction. Broad process inventories omitted in favor of marked aggregate derivatives, originals referenced by private Library ID/hash.
Compact index gives exact next/revisit conditions and source/review/ledger links. Only additive reviewed synthetic research under research/session_checkpoint_20261025; PR189/190, dirty checkout and a2a unchanged. No access, dependency, training or deployment changes. Local31-document integrity check passed. Exact-head required CI and complete blob readback must pass before authorized merge.
WIP: research: preserve reviewed three-direction checkpointto research: preserve reviewed three-direction checkpoint