Skip to content

Audit structural search learning with five reruns and independent FX replay - #4

Draft
skishore23 wants to merge 3 commits into
mainfrom
research/structural-search
Draft

skishore23 wants to merge 3 commits into
mainfrom
research/structural-search

Conversation

@skishore23

Copy link
Copy Markdown
Owner

Adds an executable bounded decision-tree search space over six exact native join kernels, including forward constraints and contradiction/deduplication. All kernels are also fixed controls. Native cold timing includes actual lazy feature/index construction. Protocol and selected winners were published before both final tests.

The learning hypothesis fails: the primary winner is identical to the saved fixed baseline and about 3.60× slower than the declared expert dispatch on larger final problems. One secondary run shows 4.42× against the saved baseline but uses a branch never exercised during training in 12 final domains. The paired five-rerun improver study gives 0.813× [0.513, 1.171], providing no reliable improvement evidence. The manually supplied forward deduplicated kernel gives 3.11× [2.42, 4.03] against the previous inverse-fiber hybrid on this cold-cost suite; this is manual engineering, not learner discovery.

Independent FX replay checks all 24,808 measured native answers across 236 new tasks. An application-derived six-entry subset model calls actual getMemory code, synthesizes 12 nontrivial retention pipelines, and validates 768 actual API replays. This uses synthetic entries, not production traces.

Retains every candidate, gzip-compressed complete raw evidence, source/selection seals, sampled-goal shortfalls, branch coverage, analysis and manuscript. Two native conformance tests cover 59,136 independent oracle decisions. Verification passed locally, including prior artifact integrity. New CI verifies archive/source integrity, native conformance, independent FX replay and the memory application.

Limitations: authored grammar and kernels; one host; no unrestricted native-code invention or recursive self-improvement claim. Controls match evaluation and measured-kernel ceilings, not exactly total discovery CPU; full selection child CPU/wall costs are retained and the limitation is explicit.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant