diagnose-score

name	diagnose-score
description	Diagnose weak, missing, failed, or surprising miner scores from public MCP evidence. Use when a submitted artifact has a confusing score, missing result, timeout, crash, or weak answer.

Diagnose Score

Goal

Classify why an artifact scored poorly, failed, or looks missing, then choose the next workflow action.

Inputs

• artifact_id
• content_hash
• batch_id
• task examples when available
• completed-batch MCP evidence

Steps

1. If upload acceptance is uncertain, check submit response, local hash, signing wallet/hotkey, get_latest_submissions, and finalized non-terminal batch membership.
2. If the artifact is absent from a completed batch, compare submitted_at with cutoff_at.
3. If the batch is still running, confirm UID/hotkey/hash membership and inspect delivery state/progress. Membership proves duplicate-preflight consideration, not scoring-task emission; wait for completion before reading result rows.
4. If execution did not happen, inspect:
   • delivery state/progress from get_miner_task_batch(batch_id)
   • selected-artifact error_counts from get_miner_task_batch_artifact_comparison(batch_id, artifact_id)
5. If timeout, crash, or budget is suspected, inspect attempts, elapsed_ms, execution_log, specifics.error, and cost totals:
   • use get_miner_task_batch_results(batch_id, artifact_id, ...) to find the affected task_id
   • call get_task_results(batch_id, artifact_id, task_id) for full task result detail
6. If the answer is weak:
   • read reference_answer from get_miner_task_batch(batch_id).batch.tasks[]
   • read response, citations, and specifics.score_breakdown from artifact-scoped result rows
   • join by task_id
7. If judge rationale is surprising, read specifics.score_breakdown.reasoning.text when present.
8. If direct target-vs-champion answer comparison is needed, run local eval in vs-champion mode and compare the report's target and opponent fields.

Stop Conditions

• Stop if the batch is not completed and the question requires result rows.
• Stop if required identifiers are missing; preserve evidence first.

Output

• diagnosis category
• evidence used
• next action: config fix, script robustness fix, retrieval fix, synthesis fix, retry policy change, or wait for the next eligible batch