fstandhartinger/model-market-comparison/select-benchmarkheaven-workers

From the Benchmark Heaven repository (Sandy: /opt/model-market-comparison), run:

node ops/rebuild-2026-09/bin/pick-worker-models.mjs --json
bash ops/rebuild-2026-09/bin/worker.sh --help

The picker fetches the current OpenRouter catalog and joins data/dataset.json. Qualification requires AA Intelligence Index >=34, conservatively across matched variants. Explicit OpenRouter IDs are preferred; exact family/organization fallback is labelled. Dated model IDs are never guessed. A missing price is not a free model; missing AA scores and known-answer smoke success never qualify unattended work.

Choose the cheapest qualified model within the task's configured price ceiling. Record actual execution identities and .meta.json receipts, not just requested IDs. Use a critic from a different vendor family than every artifact producer:

bash ops/rebuild-2026-09/bin/worker.sh --file frozen-packet.json --out draft.json 'Collect the candidate artifact using the supplied primary evidence.'
bash ops/rebuild-2026-09/bin/worker.sh --critic --producer 'ACTUAL/PRODUCER-ID' --file frozen-review.json --out review.json 'Review the supplied artifact against its primary evidence and JSON acceptance contract.'

Replace the producer placeholder with actual receipt IDs. --file embeds contents; the completion transport cannot browse or read paths in a prompt. Supply source contents, locators, timestamps/hashes and explicit coverage IDs. Follow ops/rebuild-2026-09/GAUNTLET.md: malformed output, incomplete coverage and missing source evidence are failed reviews. At most three rounds; retain prior dated data or quarantine candidates that cannot be verified. Never treat the round limit as a pass.

The routine daily job uses the tracked ops/daily/ orchestrator. Astra is an explicit escalation path for repeated failures, not a daily worker. Any Codex escalation must unset API-key/base-URL overrides and verify ChatGPT subscription authentication.

For unattended work, set BH_WORKER_MAX_PRICE_PER_1M=4. The runner rechecks both input and output prices against the live catalog even when a model ID is pinned. If no qualified distinct-family critic fits the ceiling, keep the prior snapshot and record a failed review. Do not lower the AA threshold or raise the price cap implicitly.

The daily ceiling is $4/M per direction to allow a qualified fallback when the two cheapest families time out. Choose the cheapest first; unavailable-models.jsonl records failures and excludes those exact model IDs for the rest of that run. The next run starts with a new live catalog. Set BH_WORKER_REASONING_EFFORT=low; daily data work keeps reasoning enabled. The transport's optional reasoning-disable switch is not the daily setting. Every receipt records the effective choice and returned charges. A model that cannot complete the actual source task is not viable merely because it passed the AA eligibility gate.