Good reputation is hard to share. Fraud reputation is easy to share.

This runs the real web/lib/reputation engine in your browser. Operator acme.com (enterprise, standing 80) owns 3 sibling agents. Inject fraud and watch the cluster react — the math is clamp(floor + earned − contamination, 0, cap).

Cluster — enterprise:acme.com

0 fraud events

Good reputation is "hard to share": fresh-03 inherits only a discounted, capped floor (0.5× the operator's standing), never full trust. Now inject fraud and watch it propagate.

scraper-01 · proven — 70 earned100

floor +40 · earned +70 · contamination −0 · cap 100

indexer-02 · mid — 30 earned70

floor +40 · earned +30 · contamination −0 · cap 100

fresh-03 · new sibling — 0 earned40

floor +40 · earned +0 · contamination −0 · cap 100

Why it can't be gamed

Upside is discounted (0.5×) + capped → you can't spin up a fresh agent under a star operator and inherit full trust.
Downside is fast + heavy (40 drag) → you can't dodge a fraud mark by switching to a sibling.
Fraud decays (half-life 90d) → one old mistake doesn't haunt a clean operator forever.
A pattern (≥3) escalates → the operator, not one agent, is the problem → cluster → 0.

Engine

Modelclamp(floor + earned − contamination, 0, cap)

Sourceweb/lib/reputation

Determinismclock passed in → replayable

Selfcheck7/7 invariants pass