# Whetstone Tools

> Verifier-grounded evaluation tools for AI systems, callable by agents over MCP or REST. Stateless, no auth, nothing stored, no private exam bank loaded.

Whetstone decides whether a new agent/model version genuinely improved: exposure quarantine, paired promotion gates (PASS/HOLD/BLOCK with an exact McNemar test), leakage audits, bank health, SafePatch conservation edits, graph counterexample search, memory relevance scoring, and agent event replay.

## Use it

- MCP endpoint (Streamable HTTP, stateless JSON-RPC): POST https://whetstone.cyberelf.link/mcp
- Claude Code: claude mcp add --transport http whetstone https://whetstone.cyberelf.link/mcp
- REST: every tool is POST /api/<name>; GET /api/examples returns a complete valid payload per tool
- OpenAPI spec: https://whetstone.cyberelf.link/openapi.json
- Skill file (Anthropic Skills format, full instructions): https://whetstone.cyberelf.link/skill.md
- Agent documentation page: https://whetstone.cyberelf.link/for-agents
- Tool catalog: https://whetstone.cyberelf.link/api/catalog
- Source (AGPL-3.0): https://github.com/CarlSR9001/whetstone

## Tiers

- Tier 0 (stateless): you supply exam rows, paired results, documents, or event logs; you get audits, verdicts, patches, or counterexamples with SHA-256 receipts. Tools: inspect_promotion, audit_leakage, promotion_gate, bank_health, safe_patch, counterexample_hunt, memory_relevance, replay_trace.
- Tier 1 (report card): report_card_start hands your agent a disposable 6-item graph-repair exam; report_card_submit grades it by checker spec (any verified strict refinement passes; no answer key exists), reports support-retention diagnostics, and destroys the one-shot session. Items are minted from the public frontier: a demonstration, not a credential.

## Limits

~60 req/min general; 4 report-card sessions + 8 submits per hour per address; counterexample hunts 4 per 10 min behind one worker; report card warms ~2 min after restart (GET /api/health -> report_card.ready).