# Automated pipeline

The search is a single automated chain with fail-closed evidence boundaries:

1. **Stage 1 — evolutionary generation.** GPT-5.6-sol/xhigh proposes mutation policies for the installed twisted-torus representation. Each round has 25 iterations. Structural validity and a low-weight SAT/MITM proof ladder are evaluated. Six candidates per sealed round are selected for trusted audit.
2. **Stage 2 — structural and sector audit.** Candidates are rebuilt from their construction, deduplicated against a pinned registry, checked for CSS commutation, positive `k`, row/column weight limits, and connectivity. Work is paginated 48 candidates at a time with 3 candidate workers × 4 solver workers.
3. **Stage 3 — exact SAT sectors.** Every unresolved candidate is audited by an exact sector backend (`sat-sectors`) with 2 candidate workers × 6 direction workers and a 3600-second direction budget.
4. **Stage 4 — certificate merge.** Complete X/Z evidence is merged into a candidate certificate. Incomplete directions remain unresolved; they are never upgraded by a heuristic score.
5. **Stage 5 — strict gate.** The certificate is independently replayed, known-answer baselines are checked, current-registry novelty is replayed, and the strict integer inequality `k*d*d > 12*n` is evaluated.

The public configuration is in `config/run_config.public.json`. It records the exact search and proof budgets used for this child campaign. Internal prompts, process state, logs, checkpoints, locks, and credentials are deliberately excluded.

## Representation

Each candidate is specified by lattice dimensions `(ell,m)`, an optional twisted-torus geometry descriptor, and finite monomial supports `A_terms` and `B_terms`. The exporter rebuilds the resulting CSS code and publishes the binary parity-check matrices directly, so matrix-level review does not require understanding the evolutionary renderer.

## Automation boundary

Reviewers can recommend exploration changes, but recommendations are not mathematical evidence. Only deterministic machine checks and pinned proof artifacts can reject or promote a code. The snapshot contains only committed rounds; a concurrently running round is excluded until sealed.