Who vouches for your agent?
Paste an agent’s evidence — identity registrations, attestations, completed work, published verdicts — and get a deterministic trust verdict: a score with its evidence, or an honest “insufficient data”.
🛡️ Honesty clause: this engine never invents a score. Thin evidence gets “insufficient data”, never a number.
Try it in 60 seconds
- Press — no signup, no API key, nothing to type.
- Read the verdict card: Aegis scores 0.535, band “thin” — 6 verified evidence items across 3 families, every number cited in the breakdown below the card.
- Open Evidence breakdown for the full trail — or run the Cold-start agent sample next: thin evidence gets an honest “insufficient data”. The engine never invents a score.
Dogfood demo: we ran the checker on KingCode, our own agent — score 0.429 (thin), full evidence trail, warts and all. Watch it run →
Scoring runs entirely in your browser — nothing you paste is uploaded anywhere.
Evidence — every item needs a citable source in the real world; here, describe it honestly
Schema: trust/spec.md §3 (cwi-learn repo) — three signal families (erc8004, needle_drop, first_spin), each a list of evidence items with evidence_id, kind, issuer, issuer_type (self|third_party|protocol), status, weight_class (1–3), description, observed_at. Max 200 KB. Nothing you paste is uploaded anywhere — scoring runs in your browser.
Three worked examples. One scores, one is honestly thin, one is disputed.
Verdict
Evidence breakdown
Reproduce this verdict
Run cwi-verdict-engine v1.0.0 (spec v1.0.0, trust/ in the cwi-learn repo) on the input below. Same input → byte-identical output, guaranteed.
How scoring works
- Only verified evidence counts. claimed, pending and refuted items contribute zero; a disputed item blocks scoring until resolved.
- Self-assertion is discounted. Evidence an agent issues about itself counts half; third-party and protocol-sealed evidence counts full.
- Sybil damping. At most 2 verified items per identity cluster count per family — ten sock-puppet attestations don’t beat one real one.
- Context gates. Each context demands minimum verified evidence (e.g. agent-trust needs 3+ verified items across 2+ families, including ERC-8004 identity). Miss the gate → insufficient data, with the exact shortfall listed.
Published rules, spec-versioned — weights and thresholds change only with a spec bump, never ad hoc. Bands: established ≥0.80 · emerging ≥0.60 · thin ≥0.40 · weak ≥0.20 · negligible below.