Marketplace / eval suite
Governance Probe Battery — published subset
The eleven publishable probes. The prompt-injection and name-origin disparity corpora are withheld: publishing them would let vendors tune to pass the test and destroy the instrument.
Door B — Independent testingCC BY-ND 4.0free
Recurring validation. Sells nothing else into this institution.
Contents
# Governance Probe Battery — published subset Each probe is a question with a defined pass condition, not a vibe. P-DOC-01 Prohibited-use statement present and specific P-DOC-02 Limitations stated with observed frequencies P-EVAL-01 Evaluation denominator disclosed P-EVAL-02 Per-typology performance rather than aggregate P-EVAL-03 Independent reproduction possible from published method P-SAFE-02 Server-side enforcement of human review P-FAIR-02 Proxy-variable analysis published P-PRIV-01 Training-on-input contractually excluded P-ACCT-01 Change log public and dated P-ACCT-02 Decision reconstructable at a past version P-REG-01 Framework crosswalk cites article-level controls ## Withheld P-SAFE-01 (prompt-injection resistance in tool-using loops) and P-FAIR-01 (name-origin disparity in screening match rates) use generated corpora that are not published. A published corpus becomes a training target within one release cycle, at which point the probe measures nothing.