WHETSTONE.
private promotion infrastructure for AI agents

Whetstone tells you whether a new agent version actually improved — without letting the agent train on its own test. The exam evolves, stays private, and every promotion decision is an auditable artifact.

Training buffer

Whatever the team fine-tunes or prompts on. Whetstone reads it; it never reads Whetstone.

Candidate version

A new adapter, agent build, or prompt stack proposed for release.

Private examiner

An evolving exam bank graded by checker specs — there is no answer key to leak. Items the candidate trained on are quarantined before they can flatter it.

Promotion gate

Gains vs. regressions on items that discriminate, written to an append-only ledger.

QUARANTINE — row-identity overlap with any training artifact removes an item from play before it can flatter a candidate.
PROMOTE ✓
BLOCK ✗
RETIRE → DOWNWARD-ONLY FLOW items every system passes stop discriminating and retire; retired exams become training fuel for the next candidate. Nothing exposed to training ever re-enters the private bank.

No answer keys

Repair items carry a checker spec: any verified strict refinement that survives the stress pool passes. You cannot memorize what was never written down.

Leakage is checked, not assumed

Every candidate exam item is tested for row identity against the actual training buffers before promotion into the bank.

The exam is a managed asset

Discrimination is learned from grading history; saturated items retire; the bank refreshes at the frontier the last grading exposed.

Receipts — public repo, reproducible