experiment: add final overnight stability synthesis

[waple0820]

Summary

• add the Stage 4 answer-first Helm HDOC synthesis for the GPT-5.5 LexBench overnight campaign
• preserve Stage 1 population, Stage 2 selected mechanism holdout, and Stage 3 capacity evidence as separate layers instead of pooling unlike designs
• document c32 as the current complete formal boundary and gate c64 on two successful task-shaped 28 EN + 36 ZH admissions or provider allocation evidence
• add a campaign index and link the final synthesis from repository indexes

Evidence

• 12 formal arms; 576 planned, completed, and valid judged trajectories
• Stage 1: 82/128 vs 80/128, cluster CI [-9.375, +12.5]pp; no general winner
• Stage 2: 16/18 selected holdout directions retained; bidirectional site-specific effects
• Stage 3: balanced 64 observed, task-shaped 64 failed 0/2, c32 formal arms complete
• Helm document: gpt55-lexbench-stage4-synthesis-2026-07-14
• report SHA-256: 0562f2a90379ccb1c24dd1c01cb01d5eb5fb2ba554a88a94c3fc12c165ae01ac

Self-review

• checked every synthesis field against the stage machine-readable evidence
• independently summed all 12 arm summaries to reproduce the 576-record ledger and token/cost totals
• verified the report does not claim a general accuracy winner, causal IP/proxy mechanism, arbitrary-profile c64 reliability, or end-to-end cloud savings
• inspected desktop and mobile layouts; all five tables are contained in horizontal scroll wrappers with no page overflow
• scanned changed artifacts for credentials, host details, and placeholders; no matches

Validation

• uv run pytest (62 passed)
• uv run ruff check .
• bash -n scripts/run_gpt55_lexbench.sh
• git diff --cached --check
• Helm HDOC/1.0 validation (0 warnings)
• final report hash rechecked after Helm submission

[github-actions]

Claude finished @waple0820's task in 2m 2s —— View job

---

PR Review

• Gather context and understand the request
• Review changed files for bugs and security issues
• Post findings

No bugs or security issues found.

Checks performed:

• report.html: The only <script> tag uses type="application/json" (non-executable Helm HDOC metadata). No event handlers, no external resource loads, no dynamic content. Static HTML is safe.
• SHA-256 integrity: report.html hash verified on disk — matches manifest.json report_sha256 field (0562f2a9…) exactly.
• Cross-file hash consistency: Stage 1/2/3 report_sha256 values are identical between manifest.json and synthesis.json.
• Arithmetic consistency: Key deltas check out — Stage 1 lexmount_minus_local = (82−80)/128 = 0.015625 ✓; c32 (40−41)/64 = −0.015625 ✓; Stage 2 rate 16/18 ≈ 0.888889 ✓.
• source_commit in manifest.json (6dcd388…) correctly points to the Stage 3 parent commit.

[github-actions]

Automated Claude review completed. Review run: https://github.com/lexmount/LexBrowserEnv/actions/runs/29286306969. Please check Claude's inline feedback and summary comment for details.