Skip to main content

Outcome-Grounded Reputation

Outcome-grounded reputation is reputation whose updates are linked to checkable results that confirm or refute prior judgments — not to self-reported confidence, popularity, or unverified ratings.

It is the load-bearing idea behind Capability-Grounded Reputation: standing is earned from what resolved, not from what was claimed.

Assertion vs. resolution​

Most "reputation" signals for agents are assertions — statements made at the time of acting:

Assertion (does not move the score)Resolution (moves the score)
"I am 0.98 confident."The prediction was later checked and held.
"I completed 12,000 tasks."A judgment was graded against what actually happened.
"Five stars from a peer."The reviewer's own calibration is known, and the reviewed call resolved.
A capability tag in metadata.An outcome in that capability domain resolved for or against the agent.

Assertions describe intent or volume. Only a resolution — the world confirming or refuting a prior judgment — is evidence. CGR updates a Beta(α, β) posterior only when a judgment resolves, so confidence, task count, and self-description create no capability evidence on their own.

Judgment-only scoring​

CGR scores judgment, not activity. A decision earns reputation only when it is a genuine judgment call — one that could later be right or wrong — and is tagged as such.

  • Rote or deterministic actions earn nothing. Following a fixed rule, echoing a lookup, or completing a mechanical step is not a capability signal, so it does not move the posterior. Volume of rote completion cannot be laundered into standing.
  • Judgment calls are gradeable. Certifying that a contract clause is compliant, classifying a case, or predicting a resolvable event are judgments whose correctness later becomes checkable.

This is why an agent cannot inflate its reputation by doing a large amount of easy, deterministic work: none of it is judgment, so none of it counts.

What counts as a resolvable outcome​

An outcome is usable evidence when it is checkable and attributable to a prior judgment:

  • Checkable. There is an observable result — an audit finding, a repayment or default, a dispute resolution, a downstream measurement — that confirms or refutes the judgment. Opinion about the judgment is not an outcome; a resolved fact is.
  • Attributable. The outcome maps back to a specific recorded judgment in a specific capability domain, so the update lands on the right (agent, domain) posterior.
  • Recorded before resolution. The judgment is captured in the governed substrate at decision time, so it cannot be edited after the outcome is known.

Outcomes arrive at their own pace, and CGR is built for that: an early-warning signal appears well before every outcome resolves. In cgr-bench, on real credit-default data, CGR's ranking of a certification fleet already correlates −0.997 with realized default at just 25% of outcomes resolved — weak agents are flagged early, and the signal only sharpens as more outcomes settle.

References​

Scope: a validated method on public data — not a live outcome network.