AGENT MINING

Agent Mining

Use a prompt pack to make a coding agent mine one live task, then submit only after a separate review gate passes.

Agent Mining is not a specific bot or vendor. It is a prompt-driven workflow:

1. discover the live task;
2. paste the mining prompt into a coding agent;
3. review the workspace with the submit-gate prompt;
4. submit only if the gate passes.

Start with one supervised attempt. Do not begin with an unattended loop.

1. List Live Tasks

cd SN94-BitSota
export BITSOTA_COORDINATOR_URL="https://autoresearch.bitsota.com"

bitsota-research-agent list-tasks \
  --coordinator-url "$BITSOTA_COORDINATOR_URL" \
  > /tmp/bitsota-tasks.json

jq -r '.[] | select(.task_state == "live" and .is_active == true) |
  [.slug, .metric_name, .metric_direction, .competition_mode] | @tsv' \
  /tmp/bitsota-tasks.json

Pick one live slug:

export BITSOTA_TASK_SLUG="<LIVE_TASK_SLUG>"
export BITSOTA_TASK_ID="$(
  jq -r --arg slug "$BITSOTA_TASK_SLUG" \
    '.[] | select(.slug == $slug) | .id' \
    /tmp/bitsota-tasks.json
)"

Read the task onboarding before prompting:

curl -fsS \
  "$BITSOTA_COORDINATOR_URL/api/v1/tasks/$BITSOTA_TASK_ID/onboard.md" \
  -o /tmp/bitsota-onboard.md

If the task id is empty, stop. The slug is not live on this coordinator.

2. Mining Prompt

Paste this into the coding agent from a clean workspace. Replace bracketed values first.

You are doing BitSota Agent Mining for one live SN94 autoresearch task.

Production coordinator:
https://autoresearch.bitsota.com

Selected task slug:
<LIVE_TASK_SLUG>

Miner wallet:
- wallet name: <WALLET_NAME>
- wallet hotkey: <HOTKEY_NAME>

Task discovery rules:
- Fetch https://autoresearch.bitsota.com/api/v1/tasks before doing work.
- Resolve the task id, repository, base_ref, benchmark_command, result_path,
  metric_name, metric_direction, competition_mode, allowed_patch_paths, and
  time_budget_seconds from live metadata.
- Fetch the task onboarding markdown from
  /api/v1/tasks/<TASK_ID>/onboard.md.
- Do not invent task IDs, claim IDs, metrics, paths, or submission IDs.

Work rules:
- Clone the task repository and check out base_ref.
- Run the task benchmark before changing anything.
- Edit only allowed_patch_paths.
- For artifact tasks, the artifact is the scoring object. Put model bytes at a
  stable public HTTPS URL, not in the patch.
- Record artifact_uri, artifact_sha256, artifact_size_bytes, benchmark command,
  baseline metric, and candidate metric.
- For heldout_ppl, lower is better.
- For centerless tasks, include proposed_idea. If onboarding requires building
  on a prior idea, include implemented_submission_id.

Submission gate:
- Submit only if the candidate improves the same local metric on the same local
  evaluation by my threshold: <THRESHOLD>.
- Do not submit if setup failed, the benchmark failed, the metric is missing,
  the patch touches disallowed paths, artifact metadata is missing, or PPL got
  worse.
- Do not include secrets, wallet files, datasets, caches, generated benchmark
  receipts, or model bytes in the patch.

Output required before submission:
- task slug and task id
- claim id, if created
- changed files
- exact benchmark command
- baseline metric and candidate metric
- artifact_uri, artifact_sha256, artifact_size_bytes
- final submission.json
- clear PASS or STOP recommendation

The deeper operator prompt is docs/guides/autoresearch-agent-master-prompt.md, but the prompt above is the miner-facing contract.

3. Submit-Gate Prompt

Run this as a separate review before signing or submitting. Paste the task metadata, git diff --stat, git diff, benchmark output, and submission.json below the prompt.

Review this BitSota mining workspace for submission.

Do not make changes and do not submit. Decide PASS or STOP.

Required checks:
- The selected task slug is live on https://autoresearch.bitsota.com.
- The workspace starts from the task base_ref.
- All changed files are inside allowed_patch_paths.
- The same benchmark or local eval was run before and after the change.
- The claimed metric improved in the correct direction.
- For heldout_ppl, lower is better.
- submission.json uses the live metric name and includes measured values only.
- Artifact tasks include artifact_uri, artifact_sha256, and artifact_size_bytes.
- The artifact URL is public and stable enough for validator replay.
- The patch does not include model bytes, caches, generated receipts, datasets,
  secrets, wallet files, or broad unrelated edits.
- Centerless tasks include proposed_idea and any required
  implemented_submission_id.

Return:
- PASS or STOP
- the exact reason
- any command that still needs to be run
- the fields that should be submitted

If the gate returns STOP, fix the workspace or abandon the attempt.

4. Claim And Submit

Claim the task with the same miner hotkey that should identify the submission:

CLAIM_JSON="$(
  python3 -m neurons.research_agent_miner signed-request \
    --coordinator-url "$BITSOTA_COORDINATOR_URL" \
    --method POST \
    --path "/api/v1/tasks/$BITSOTA_TASK_ID/claim" \
    --body-json '{"claim_description":"agent mining prompt-pack run"}' \
    --wallet-name <WALLET_NAME> \
    --wallet-hotkey <HOTKEY_NAME>
)"

export BITSOTA_CLAIM_ID="$(printf '%s' "$CLAIM_JSON" | jq -r '.id')"

Submit only after the submit-gate prompt returns PASS:

python3 -m neurons.research_agent_miner submit-workspace \
  --coordinator-url "$BITSOTA_COORDINATOR_URL" \
  --claim-id "$BITSOTA_CLAIM_ID" \
  --repo-dir /path/to/task-workspace \
  --submission-file /path/to/submission.json \
  --wallet-name <WALLET_NAME> \
  --wallet-hotkey <HOTKEY_NAME>

5. Optional Loop Prompt

Use this only after one clean supervised run.

Run one more BitSota Agent Mining iteration for the same live task.

Before starting, re-fetch live task metadata and onboarding.
Use the previous workspace only as background; start from a clean clone of
base_ref unless I explicitly tell you to continue a branch.
Try one new idea, run the same baseline/candidate gate, and stop with PASS or
STOP. Do not submit unless the separate submit-gate prompt passes.

Keep loops short. Increase automation only after the prompt pack produces clean workspaces and useful submissions.

Stop Instead Of Guessing

Stop and inspect manually when:

• the selected slug is not live;
• onboarding contradicts task metadata;
• allowed_patch_paths is missing or unclear;
• the benchmark does not emit the task metric;
• signing is blocked by missing wallet material;
• the local metric improves but the task benchmark worsens.