Research: five tested structural-frontier extensions after PR #17

[CompleteDotTech]

Five executed structural-frontier extensions

Frozen protocol: ef8fc308d0a50cd37a6cdd3b7546e050093349c8, committed before implementation/execution against PR #17 baseline a62a3257645d8e35cd4e45be53bfa9511d27724b. Retain this ancestry with a normal merge.

Final integrated execution: https://github.com/CompleteDotTech/paper-package/actions/runs/35330895434

Final generated head: 433d5f03ef98e0f5dddb8d1a893790181ec8a6fe. This close/reopen triggers final-head PR checks after the generated-artifact bot commit.

Frozen tests and actual outcomes

Hypothesis	Experiment	Outcome
H1: Latent-source-conditioned lineage	96 independent joint-world oracle fixtures; 288 duplicate/order checks; 0.95 admission threshold	Target met. Zero oracle/invariance errors; nine false admissions prevented versus independent marginals. Wrong source metadata remains a failing control.
H2: Cost/noise-aware group review	Budget 40, heterogeneous proxy effort and supplied reviewer errors; required >=5% fewer expected contaminated groups than both comparators with bounded correct-edge loss	Target NOT met. All three selectors leave 13.043478 expected contaminated groups and 0.06 expected correct removals. 64 optimizer-oracle checks pass; modeled objective improvement is not demonstrated downstream improvement.
H3: Reusable state-bounded lineage diagrams	128 small formulas x five probability assignments; ten large analytic formulas; 64 reused-structure updates	Target met. Zero probability errors; all ten connected 17–256-atom formulas exact beyond the previous atom cap. Explicit state/input caps and fallback bounds remain.
H4: Certified bipartite conflict optimization	128 bipartite + 64 general finite-oracle graphs; nine large analytic structures; independent flow/cut certificate verification	Target met. Zero optimum/certificate/regression errors; nine previously staged large cases exact. Greedy also attains those nine optima. Certification concerns supplied priorities, not factual truth.
H5: Connected-path incremental utility	64 paths x 16 updates plus a 1,024-vertex path x 256 updates; utility and selected-witness verification	Target met. Zero mismatches. Summary transition candidates fall 789,504 -> 27,864 (96.47%), including cold build. Linear witness decoding adds 526,079 separately reported node visits and is excluded from that percentage.

Validation and deliverables

809/809 integrated graph-synthesis regression tests pass, with zero errors, failures or skips, including 75 new tests. Actual logs and test-file hashes are committed in graph_synthesis/frontier/all-regressions.json and .log.

Full deterministic numerical replay, recorded-response reconstruction, source-group separation, source/code SHA-256 checks and the original immutable 161-file inventory pass. Four concurrent study directories are verified unchanged against main 114450ef08736c21c97c8725de700d8d2c6ca94d.

Deliverables: full results.json, executed report, frozen protocol, novelty audit, control cases, review selections, flow certificates, update sequences, summary CSV, five SVG/PNG figure pairs, and the complete updated Markdown/HTML/PDF paper with source/renderer/PDF build hashes. The full manuscript and CURRENT_RESULTS.md retain all preceding studies.

Final-head PR checks must pass before merge: Windows/Linux replay and integrity, deterministic figures, complete sequential manuscript regeneration and all earlier research workflows.

Novelty and evidence boundaries

These are extensions relative to the frozen PR #17 baseline, not globally unprecedented algorithm claims. Established Bayesian source modeling, knowledge compilation, min-cut and dynamic programming are cited.

During execution main received PRs #18, #19, #24 and #21. All four are preserved with genuine merge ancestry and byte-verified study directories. In particular, H4 is a parallel implementation/additional frozen test suite because the concurrent source-structural study also implements certified bipartite optimization. Dependence bounds versus supplied latent-source models, frontier-width versus reusable decision diagrams, and tree-message versus path-segment-tree updates have overlapping goals and explicitly distinguished assumptions. No targets or fixtures were retuned after seeing these studies, and shared captured observations are not pooled as independent replications.

Fresh Jev calls: zero. H2 reuses 73 development and 263 evaluation candidates with assumed proxy effort and reviewer errors; it is not an observed human-review experiment. H1/H3/H4/H5 are controlled supplied-input algorithm tests. No new corpus, independent semantic-accuracy finding, KARMA superiority or production-readiness claim. Evaluation gold is excluded from selection. No production compiler policy changed.

Integration corrections are recorded openly: a CI-token workflow-write rejection was resolved through the authorized GitHub connection without escalating permissions, and an unchanged legacy regression caught an incorrect manuscript insertion point. The report placement was corrected; the test was not weakened. All 809 integrated tests then passed.

python -B -m graph_synthesis.frontier.run
python -B -m unittest discover -s graph_synthesis/frontier/tests -v
python -B -m graph_synthesis.frontier.run --check
python -B -m graph_synthesis.frontier.report --figures --update-paper
python -B -m graph_synthesis.render_current_paper --output manuscript/paper-current.pdf
python -B -m graph_synthesis.frontier.verify

Author-review research; experimental methods remain opt-in rather than silently replacing production policy.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-18T09:20:58.889396Z	c9c1d34	PR opened
🔒 Security Review	✅ Completed 2026-09-18T09:20:34.437899Z	c9c1d34	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.