Skip to content

Add exact finite synthesis and measured limits of planner learning - #3

Merged
skishore23 merged 4 commits into
mainfrom
research/exact-join-search
Oct 31, 2026
Merged

skishore23 merged 4 commits into
mainfrom
research/exact-join-search

Conversation

@skishore23

@skishore23 skishore23 commented Oct 30, 2026 •

Copy link
Copy Markdown
Owner

FX can now synthesize bounded operation sequences from complete finite transition tables, including irreversible operations, and independently replay immutable result certificates on every modeled state. Resource exhaustion is distinct from bounded impossibility. Stacked on #2.

Measured results:

  • 396/396 retained primary and deeper tasks independently reconstructed through the actual FX API.
  • Separately sealed C++ scaling study: 14.97× geometric warm-query speedup over compiled pairwise search on 76 length-at-most-ten tasks (95% domain-bootstrap interval 10.43–21.87×). A conservative estimate including common bank construction and full shared index cost is 2.75× (1.73–4.11×).
  • Known five-, six- and seven-state recovery examples reproduce reset lengths 16, 25 and 36 with preceding-length bounded impossibility checks.
  • Negative learning result: all three 96-candidate searches select the existing strongest fixed hybrid. No demonstrated learned improvement or finite discovery break-even.
  • Benchmark audit: the random sampler found no eligible targets in one scaling domain, but exact enumeration exposes 405 reachable novel tables. Historical saturated_domains fields mean sampler exhaustion, not semantic saturation. Frozen timing and selection data are unchanged; exact witnesses are retained separately.

Includes research/exact_join/paper.md, raw timings, every candidate, task specifications, replay certificates, source/data hashes, plots and reproduction instructions. Primary and scaling source/protocol/winner seals precede their respective tests. Algorithmic novelty, recursive self-improvement and production Beetle recovery remain unproven.

Validation: all GitHub CI jobs passed on c49f75d (Node 20/Python 3.12), including type check, lint, build, 198 Jest tests, seven existing experiment checks, three Python conformance tests, artifact integrity and independent FX verification of all 396 tasks. Two additional rare-target audit witnesses were independently solved and verified locally.

@skishore23
skishore23 changed the base branch from feat/program-discovery to main October 31, 2026 04:46
@skishore23
skishore23 marked this pull request as ready for review October 31, 2026 04:46
@skishore23
skishore23 merged commit bf93654 into main Oct 31, 2026
4 checks passed
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