Reading the evidence

What a decision means. And what it does not.

An Etalon decision is a precise statement with a narrow scope. This page is the plain-language version of that scope, written so your reviewers can quote it.

Scope

A decision is scoped to four things.

Change any one of them and the decision does not carry over. It has to be re-run.

This endpoint

As fingerprinted at run time: model identity, serving details, hash. A different model, quantisation or prompt template is a different run.

This pack, at this version

The cases, evaluators and thresholds that were applied. A new pack version means a new run.

The cases that ran

Every input and output is in the bundle. The decision covers those cases and nothing else.

The coverage recorded

Which parts of the taxonomy were exercised and how heavily. Incomplete coverage cannot produce a QUALIFIED result.

Three outcomes

Exactly one per run.

There is no score to interpret and no partial result. If a run cannot support a decision, it says so instead of guessing.

QUALIFIED

Every critical dimension cleared its threshold and coverage is complete. This endpoint, at this fingerprint, may do this job under this pack version.

NOT_QUALIFIED

At least one critical dimension fell short of its threshold. The bundle shows which cases, with the raw outputs.

INDETERMINATE

The run could not support a decision: incomplete coverage, an evaluator that could not be applied, or an endpoint that failed mid-run. It is a first-class outcome, not an error to hide.

What a decision is not

Five things a QUALIFIED result does not say.

A decision travels: into a slide, a ticket, a contract discussion. It should mean the same thing when it arrives.

  1. A statement about other endpoints

    Another model, another quantisation, another prompt template or another serving stack is another run.

  2. A statement that a regulation, contract or control is satisfied

    Etalon evaluates an endpoint against a pack. Whether that meets an obligation your organisation holds is a judgement for you and your advisers, using the bundle as one input.

  3. Permanent

    Models drift, taxonomies change, packs get new versions. A decision carries a date and a fingerprint for a reason.

  4. A substitute for your own review

    The raw outputs are in the bundle precisely so a reviewer can read them and disagree with an evaluator.

  5. Issued by an external authority

    The runner ran on your machine, on your endpoint. Coyos wrote the pack; it did not sign off on your deployment. The decision is yours to act on.

How to cite a result

Three sentences that survive being quoted.

Each one names the artefact it rests on. Each value comes from the bundle.

"Endpoint support-router-8b (fingerprint e41b…2d77) was qualified under pack contact-routing@1.2.0."

decision.json · fingerprint.json · manifest.json

"The endpoint was evaluated against methodology contact-routing, version 1.2.0, with coverage 38 of 38 intents."

manifest.json · coverage.json

"Evidence generated by runner etalon 0.x on 2026-09-24, bundle 4a1e…9c03, signature verified offline."

manifest.json · manifest.sig

Values shown are illustrative.

Language

Evidence, not adjectives.

The words around a result do as much work as the result. So Etalon's vocabulary is small on purpose: evidence, qualification, pack, runner, bundle, decision, fingerprint, coverage. Each names something you can open.

We do not use words that imply an external authority has signed off, that a standard has been satisfied, or that a result is stronger than its scope. Not on this site, and not in a bundle.

If you find language on this site or in a bundle that oversells, tell us. We will treat it as a bug.

Bring your reviewer's questions.

If the person who has to approve your deployment has a question this page does not answer, send it to us. Those questions shape the packs.