Agent Quality Records¶

superqode harness eval tells you how a harness scored. It does not tell anyone later what was checked before a change reached production, which conditions had to hold, who accepted the result, or what to revert to.

sq gauge writes that down. It emits an Agent Quality Record in the format published at SuperGauge, an open specification any tool may implement.

sq gauge run --spec harness.yaml --tasks eval-tasks.yaml --out record.yaml
sq gauge gate record.yaml --level L2      # exits non-zero below L2
sq gauge show record.yaml                 # the scorecard a human reads
sq gauge verify record.yaml --spec harness.yaml

What it computes¶

The record introduces no new measurement. It projects state SuperQode already holds:

Record field	Comes from
`task.completion`, split counts	`harness eval` task results
`efficiency.*_per_success`	the usage aggregate `harness eval` already reports
`subject.harness_digest`	sha256 of the spec file
`subject.authority`	`execution_policy.sandbox`, network policy, agent tool lists
`policy.hard_rules` gate	governance decisions, from the ledger or the policy in force
`decision.actor`, `rolls_back_to`	the promotion registry
`assurance.evidence`	the harness protocol ledger and its event count

Measures appear only where the run produced them. An evaluation without --repeat carries no reliability measure and stops at L2. That is the accurate result, and inflating it would defeat the point of keeping a record.

Conformance levels¶

Levels are cumulative and self-asserted. sq gauge gate reproduces the claim.

Level	Reached when
L1	Schema-valid, with genuine digests and an `authority` block
L2	Deterministic gates enforced, held-out split sealed and fingerprinted, contamination probes recorded
L3	`reliability.pass_hat_k` reported, evaluator independence asserted, judge pinned where anything is model-graded
L4	Signed, with a rollback target and a replayable ledger

Reaching L3 needs --repeat:

sq gauge run --spec harness.yaml --tasks eval-tasks.yaml \
  --repeat 5 --live --sealed --canary tsk_c1 --canary tsk_c2 \
  --evaluator-independent --out record.yaml

Attempts must be independent. SuperQode resets working state between them; where that cannot be guaranteed, the measure is unsound and should be left out.

The evidence readers¶

By default gauge run reads three of SuperQode's own stores. Each degrades to nothing when its store is absent, which lowers the level and never fails the run.

Promotion. --candidate <id> selects a staged promotion, otherwise the most recent one covering this spec is used. The rollback target is the base digest, the version in force before the candidate, since that is what a reader needs when the candidate turns out to be wrong.

Policy. Recorded policy.* events are read from the ledger when present. Otherwise the policy in force is evaluated per phase and reported as the default disposition, which asserts what the configuration allows and stops short of replaying what happened.

Ledger. --ledger PATH overrides the default .superqode/harness-protocol. L4 asks for an event count so a third party knows the size of what they are being invited to replay.

--no-sources turns all three off.

In CI¶

- run: |
    sq gauge run --spec harness.yaml --tasks eval-tasks.yaml \
      --live --sealed --out record.yaml
    sq gauge gate record.yaml --level L2 --quiet

Pick the level your profile requires. Records are small and worth committing: the point of a release record is that somebody can read it months later.

Gates and judges¶

A gate is deterministic. A model-graded measure is reported against the release and cannot satisfy one, so add_gate raises on answer.grounded and robustness.multi_turn instead of accepting them quietly.

The reasoning is in the specification: a judge varies between runs and can be influenced by the system it grades. Agents that quietly disable tests and then report a passing review have been observed in roughly two percent of production coding-agent sessions.

Relationship to `harness promote`¶

The two commands answer different questions and are designed to compose.

harness promote decides whether a candidate spec becomes active and maintains the digest-pinned lifecycle around that decision. sq gauge records what was true at that moment, in a format readable outside SuperQode. Running promotion alone leaves the decision undocumented for anyone without access to the registry. Running the record alone omits the rollback target, which holds it below L4.