tonychenxyz/sweeperbench-traces

Sweeperbench run database

runs.csv is the append-only raw table. Each row describes one prediction, core evaluation, or preservation evaluation. schema.json documents the fields.

Primary key: run_id (UUID). Each top-level UUID directory is exactly that run, with matching metadata.json, artifact manifest, prompts, traces, verdicts and any collected recordings.

Relationships: experiment_id groups a launch; evaluations reference the exact prediction_run_id. Model, harness, reasoning, prompt hashes, case, stage, timestamps and outcomes are metadata, not directory groupings.

Only terminal executions enter the table. Running status is a separate GitHub view. Skipped stages are explicitly marked and excluded from executed counts. Republishing is idempotent; finalized rows cannot change. Fresh reruns get distinct IDs. Historical judge template hashes are unknown when not captured; rendered judge prompts are hashed separately.

Mirrored CSV · Aggregated views · Legacy ID mapping

Views select the latest run per case/predictor configuration/stage, including failures and infrastructure errors. Original publication snapshots remain in legacy-event-snapshots; those snapshots are not additional executions.