Quality Ledger

Engine: raven-solana-live-scan@0.1.0 (vocabulary v4) · active key rvk_c2997e90215279c2 · latest public blackbox run: 2026-06-05 — 9 pass, 0 fail · 50 site tests.

Rules: a failing eval becomes a build order · new capabilities must ship with evals · no undocumented trust expansion. Sanctioned engine-work sources: measured feedback, beta users, explicit coverage gaps, eval failures.

Known limitations: holder beta key-gated · deployer history not yet live · liquidity gap without pool evidence · Raydium v4/CEX custody unadjusted · no price prediction or trading advice, by design.

Visible trust machinery

Raven's trust is machinery, not persuasion — all of it inspectable: public key (keyId rvk_c2997e90215279c2) · engine version + observed slot in every receipt · findings + coverage gaps · replay hash + official attestation hash · evals · quality ledger · evidence source registry · decision policy · abuse runbook.

Quality is not token consumption

Raven does not measure quality by model used, tokens consumed, agent steps, model calls, lines of generated code, PR count, prompt length, or UI badges. Quality is: signed receipts issued (especially EXTERNAL ones), signature-verification rate, verdict stability across re-checks, gap resolution, tamper-rejection rate, fail-closed behavior, and blackbox eval pass rate. Full good/bad metric lists: /quality-ledger.json. Model-neutral by construction: the verdict never depends on an LLM; provider swaps re-run the blackbox evals and nothing else changes.

Evidence per action, not tokens consumed

The verifier fails closed rather than hanging; agents apply their own timeout and retry policy. Deterministic checks need no LLM calls — do not spend a frontier reasoning model on deciding whether a signed receipt is valid; ed25519 verification is a deterministic function. Pre-spend actions reverify immediately; delayed execution reverifies at execution time. Machine-readable: /quality-ledger.json costLatency.