Dossier — take the suite with you
The dossier
Every export carries the six factor weights, the per-transcript grades, the sealed inputs, where the data came from, and the head of the audit chain — so a reader can check the number instead of trusting it.
Download
| Task | State | Markdown | JSON | CSV | Kaggle bundle |
|---|---|---|---|---|---|
| Retry storm under a flaky tool 3 transcripts · 3 assertions | reference | Markdown | JSON | CSV | files |
| Nested envelope collapse 3 transcripts · 3 assertions | reference | Markdown | JSON | CSV | files |
| CSV unit-of-measure drift 3 transcripts · 4 assertions | reference | Markdown | JSON | CSV | files |
Markdown
The readable one. Factor table, per-transcript grades, sealed inputs, provenance, chain head.
text/markdownJSON
The whole task plus verdict, grade and replay result. This is what a script should read.
application/jsonCSV
One row per recorded transcript, for a spreadsheet. Drops into a leaderboard notebook as-is.
text/csvWhat every export contains
- The six published weights and each factor’s real contribution, so the bars sum to the score.
- Every assertion outcome with the exact span or value that decided it.
- Which transcripts are recorded runs and which are bundled reference material, labelled individually.
- Where external data came from, when it was retrieved, and whether it was
liveor a sealed snapshot. - The chain head plus the command to recompute it:
curl https://crucibleforge.vercel.app/api/tasks/<id>/integrity
Caveat carried in every file: a grade is a statement about a task, not about a model. Where a grade would need a language-model judge, the weight is reported as undecided rather than counted as a pass.