Tinker RL · Project Defense Evidence
Defense safety: every inference response is labeled either LIVE HF ROUTER or DETERMINISTIC FALLBACK. Static evidence remains available even if providers are cold.
A hosted model chooses a tool; the Space executes it
The Router model is named in the status. The project tool-call LoRA is linked separately as an artifact.
Ordinary math reasoning — not a tool-call-only demo
This sends the question directly to the HF Router. The fixed fallback is only used when hosted inference is unavailable.
Frozen evidence snapshot
These figures are bundled in the Space, so they do not depend on W&B loading during the defense. The run links remain available for drill-down.
Audit layers: 983 account-level objects → 70+ curated runs with usable telemetry → 19 claim-critical gold rows.
Verify the fail-safe bundle locally
The verifier uses only Python's standard library. It checks the 983-vs-70+ reconciliation, the 19-row gold subset, the G=2/G=16 evidence pair, HTTPS run links, unique run IDs, image presence, and SHA-256 digests.
983 vs 70+: 983 is the account-level object count; 70+ is the curated set of runs with usable telemetry. Separately, 19 gold rows form the claim-critical evidence subset. These are different audit layers, so the counts are consistent.