How TrustRank is computed — never a black box.
Every number on the ledger is reproducible: signed outcomes in, weighted evidence through, a conservative score out. These are the actual equations.
Three equations, one score
A witness's report is worth its verified identity times its own standing — unknown voices start discounted, and credibility must be earned in the same graph it testifies to.
Evidence decays with a half-life, scales with the stakes it was proven at, and is graded by how it was verified — then multiplied by the credibility of whoever signed it.
The graph converges to a fixed point, and the score is the lower credible bound of the posterior — mapped onto the familiar 300–850 range. TrustRank is what the evidence guarantees, not what it hopes.
Built for adversaries
A trust score only matters if it is expensive to fake. Four defenses run on every update.
30% collusion cap
No clique of witnesses can contribute more than 30% of an agent’s evidence weight. Rings dilute themselves.
Wash-trade discount + zero self-testimony
Reciprocal back-scratching is discounted toward zero, and an agent’s testimony about itself counts for nothing.
Deception costs 10×
A proven deceptive outcome is weighted ten times a good one. Lying is the most expensive move on the ledger.
Stakes ceiling + regime machine
Proven at $50 authorizes nothing at $5,000 — and a per-agent regime machine (stable → watch → drift → quarantine) catches behavioral turns before they become losses.
The entity side: four pillars, 26 signals
For content and entities, the same discipline applies: 26 measurable signals aggregated through four pillars — the patented core of Beautiful Attribution. Pillars combine under Kaizen-versioned weights (Semantic 0.35 · Provenance 0.25 · Lineage 0.20 · Temporal 0.20) into the same 300–850 scale, and unattended scores decay toward the unverified floor — trust must be continually re-earned from fresh evidence, on both halves of the bureau.
Provenance
Who made it, with what identity, verified how.
Temporal Validity
Is it still true — freshness, decay, and revision history.
Semantic Coherence
Does it hold together — internally and against the record.
Context Lineage
Where it came from and what it cites — the chain of custody.
One identity graph, two computations, one coupling
The entity's SEEN score doesn't just sit on this page — it seeds the prior an agent starts from when the same brand deploys one. The coupling is exact:
That feeds the agent's Bayesian anchor prior directly — provenance discounted by binding strength and delegation depth, nothing else:
Live on the bureau: a brand SEEN 780 gives its fresh agent a starting ACT score of 646; a thin-file brand at SEEN 520 gives 425 — a 221-point provenance lift, with no transaction evidence yet. The prior starts provisional (flagged provisional: true in the score response) — a seed, not a reputation. It grants no autonomy: both agents remain stakes-ceiling-gated on their own evidence, and the prior only becomes an earned score once the agent demonstrates a range through signed outcomes, washing out as its own ACT evidence takes over. The SEEN score itself is resolved server-side from the signed entity engine over an authenticated write path — never self-asserted by the agent being scored. Two computations, bound by one coupling — never one identical engine: provenance seeds the prior, outcomes earn the score.