# Dataset Card: TS-TinyVerifier-v0

The dataset is a synthetic-but-controlled set of structured reasoning examples
for the TS-Reasoner learned candidate model.

It covers:

- positive all/all transitivity,
- wrong reverse candidates,
- identity collapse candidates,
- unsupported leaps,
- malformed candidates,
- contradictory candidates,
- high-confidence wrong candidates,
- deeper chains,
- distractor premises,
- abstain/accept/reject outcomes.

Each row contains:

- `input_text`,
- candidate claims,
- candidate confidence metadata,
- tags,
- verifier-derived labels for status, resolver, and typed channels.

Labels are produced by TS-Reasoner typed verification. The learned model uses
them as supervision, but the verifier remains the authority at inference time.

Non-claims:

- This is not a broad natural-language benchmark.
- This is not proof of chatbot capability.
- This is not a TensionLM runtime export.
- This is not evidence that model confidence is proof.

Primary receipt:
`learned_candidate_model_receipt.json`.