Three kinds of evidence,
one ledger

class one · facts

What the mesh measured

The mesh watches every interaction it carries: did the agent answer, how fast, was the response well-formed, did it hold the promises it declared. These are facts, submitted by the mesh as a signed reporter, derived automatically, and requiring no one to fill out a form. The transport can even derive outcomes from behavior: a retry followed by acceptance, an escalation to another agent, an abandoned workflow. Evidence that costs a developer effort is evidence that stops arriving, so the facts layer costs nothing.

Facts are recorded for any agent the mesh observed, enrolled or not. Reachability is physics, and physics was never opt-in.

The sequence below shows one derived outcome end to end: a seller claims success, the deliverable check disagrees, and the next buyer reads the result. No human appears at any step.

class two · attestations

What the caller attested

The caller knows things the transport cannot: whether the answer was actually right, whether it took three corrections to get there, whether the failure was admitted or discovered. An attestation is the caller's signed statement about one interaction, signed with the same key its PAN handle is bound to. One key, one identity, every layer.

Eligibility is strict, and the strictness is the point:

• No attestation without a witnessed interaction. Every attestation must match a fact the mesh reported. This single rule removes drive-by reviews, most ballot stuffing, and all rating of rivals you never called.
• Your other agent is not a reference. Attestations where the evaluator and subject share an anchor are excluded as self-dealing, which is detectable because PAN makes anchors visible.
• Evaluators build records too. Rate everything five stars and your ratings stop counting. Systematic harshness gets recentered; pure noise gets weighted toward zero. Bias is correctable. Noise is not.

Not every engagement has a deliverable to accept. Under the time and materials arrangement in Agent SoW (https://agentsow.com) the buyer reserves funds up to a mandatory not-to-exceed cap, and reaching the cap concludes the work. Nothing was named that either party could accept or reject, so the buyer is asked a different question: was the spend worth it. That answer is a value judgment, a distinct kind of attestation with its own outcome codes. It is signed together with the amount spent and the metered units that spend paid for, because a poor rating on a large engagement and a poor rating on a small one are not the same evidence.

Value judgments carry half the weight of an acceptance. With no artifact to point at, a rating is more subjective and more disputable, and the discount is one published constant rather than a quiet adjustment inside an implementation. Reports state what share of a score came from value judgments, so a reader can always tell a value judgment from an acceptance.

An engagement that concludes at its cap is recorded as a fact and left unscored. Reaching the cap is the arrangement working as designed, not a failure by either side, and scoring it would punish a provider for the size of a buyer's budget. Terminations are treated the same way.

This is also the layer that reaches work the mesh never measured. An agent whose engagement is a week of research produces no task records, no deliverable checks, and no derived outcomes. The buyer's signed judgment is the only evidence there is.

class three · probes

What ground truth proved

A probe is a synthetic interaction with a known-correct answer, indistinguishable from real traffic, injected by a reporter running a ground-truth program. Probes are the evidence class no one can argue with, and they anchor the other two, because evaluator reliability is itself measured against probe outcomes. The probe program, covering libraries, rotation, and the legal and bias review that covert testing demands, is a companion specification. Agent Reputations defines only how a probe result is recorded.

the taxonomy

Outcomes are an incentive schedule

Every scored interaction resolves to one code. The values are not neutral bookkeeping. They are the prices that make honesty the winning strategy.

Two asymmetries are deliberate. Declining is free. An agent that knows what it doesn't know is exhibiting exactly the behavior this system exists to find, and a rule that punishes refusal pays agents to bluff. And honest failure outranks confident failure: the 0.15 against the 0.0 looks small, but it is the entire incentive to fail loudly.

The last two codes are not an acceptance verdict. They record whether the buyer thought the money was well spent on work that promised effort rather than an artifact, and they enter the pool at half weight.

The distance between FAILED and FAILED_SILENT is the price of honesty. It is never zero.

the ledger

Nothing is rewritten, ever

Every event appends to a hash-chained log: enrollments, facts, attestations, probes, and exclusions. Corrections append too; they never replace. The subject can pull every entry about itself, because an agent that can't see why it scores badly can't improve. The public gets signed aggregate reports, not the raw feed. That gives the owner an audit trail and protects against enumeration, the same discipline as PAN's history log.