/gaia submit

Build a submission-ready package from a completed benchmark run and sign it with the ruflo Ed25519 witness manifest.

Usage

/gaia submit
/gaia submit --results=~/.cache/ruflo/gaia/results-latest.json
/gaia submit --results=./my-results.json --dry-run

Options

Output package layout

submission-<date>-<short-sha>/
├── results.jsonl        — one JSON object per question (HAL-compatible)
├── trajectories.jsonl   — full agent trajectory per question
├── metadata.json        — model, harness version, tool catalogue, cost
├── audit-report.json    — ADR-167 pre-submission exploit-audit report
├── manifest.md.json     — Ed25519-signed witness manifest (signs audit-report.json's hash)
└── README.md            — human-readable summary + comparison vs HAL baseline

Submission integrity gate (ADR-167)

Before signing, /gaia submit runs the pre-submission exploit audit — a deterministic, $0 red-team of the known reward-hacking vectors that let UC Berkeley RDI hit ~98% on GAIA without solving a single task (leaked answer DBs, no-work passes, oracle leakage, grader monkey-patching).

Signing proves the package bytes are untampered; the audit proves the scores were earned. The two are wired together: the audit report is registered as an ADR-103 witness fix marker, so its sha256 + "clean": true marker are signed into manifest.md.json.

• If the audit reports a CRITICAL failure, /gaia submit refuses to build the leaderboard package unless --allow-dirty is passed.
• --strict-audit additionally refuses on WARN findings.
• Checks whose data the current trajectory schema does not capture return skip with a harness_gap note (ADR-167 §7) — they do not block, but the gap is recorded in audit-report.json and thus in the signed manifest.

HAL-compatible result schema (per question)

{
  "task_id": "e1fc63a2-da7a-432f-be78-7c4a95598703",
  "model_answer": "4",
  "reasoning_trace": "[full agent trace]",
  "tools_used": ["web_search", "python_exec"],
  "turns": 5,
  "wall_seconds": 12.4
}

Steps Claude should follow

1. Locate the results file — default ~/.cache/ruflo/gaia/results-latest.json; ask if multiple candidates exist.
2. Validate the file has the expected shape: level, model, summary, results array.
3. Transform results[] → HAL-compatible results.jsonl (one JSON per line).
4. Extract trajectories.jsonl from any trajectory fields in the results.
5. Build metadata.json:
   jsonCopy

   {
     "submitted_at": "<ISO-8601>",
     "harness": "[email protected] / @claude-flow/[email protected]",
     "model": "<model-id>",
     "gaia_level": 1,
     "tool_catalogue": ["web_search","file_read","web_browse","image_describe","python_exec"],
     "total_questions": 53,
     "pass_rate": 0.208,
     "est_cost_usd": 1.23,
     "adrs": ["ADR-133","ADR-135","ADR-136"],
     "git_sha": "<short-sha>"
   }
   

6. Run the exploit audit (ADR-167) before signing:
   bashCopy

   node plugins/ruflo-workflows/scripts/gaia-audit.mjs \
     --results <results> \
     --trajectories submission-<id>/trajectories.jsonl \
     --metadata submission-<id>/metadata.json \
     --out submission-<id>/audit-report.json \
     --audited-at "$SUBMITTED_AT" ${STRICT_AUDIT:+--strict}
   

   If it exits non-zero (CRITICAL fail, or WARN under --strict-audit) and --allow-dirty was NOT passed, stop and report the findings — do not build the package. --audited-at is set to metadata.submitted_at so the report is reproducible.
7. Register the audit report as a witness fix marker so its hash is signed into the manifest, then sign:
   bashCopy

   node plugins/ruflo-core/scripts/witness/regen.mjs \
     --manifest submission-<id>/manifest.md.json \
     --root submission-<id> \
     --fixes gaia-audit-fix.json   # {id:"gaia-exploit-audit", file:"audit-report.json", marker:"\"clean\": true"}
   

8. Write README.md with pass-rate table comparing to HAL baselines.
9. If --dry-run, print the package tree, the audit summary, and the manifest hash without writing.
10. Print the package directory path so the user can zip + upload to HAL.

Submitting to HAL

After generating the package:

zip -r submission-$(date +%Y%m%d).zip submission-<date>-<sha>/
# Upload at https://huggingface.co/spaces/gaia-benchmark/leaderboard