Skip to content

sq gauge

Emit and check Agent Quality Records. Format: SuperGauge.

For the concepts behind these commands see Agent Quality Records.


gauge run

Evaluate a harness and emit a record.

sq gauge run --spec harness.yaml --tasks eval-tasks.yaml --out record.yaml
Option Default Purpose
--spec PATH required The HarnessSpec under evaluation
--tasks PATH required Eval task file
--provider ollama Model provider
--model spec's policy Model id
--repeat N 1 Independent attempts per task. Above 1 produces pass^k and pass@k, which L3 requires
--profile ID sg/coding-agent Profile the record claims
--tier {T0,T1,T2} T1 Risk tier
--sealed / --unsealed unsealed Assert the held-out split was closed to anything that tunes
--canary ID none Contamination probe task ids, repeatable
--evaluator-independent off Assert the grader saw only the artifact and resulting state
--ledger PATH .superqode/harness-protocol Event ledger to reference
--candidate ID latest Promotion supplying the actor and rollback target
--no-sources off Skip the promotion, policy and ledger readers
--out PATH stdout Where to write; .json selects JSON
--live / --dry-run dry run A dry run skips execution, so it reports no measures

--sealed and --evaluator-independent are assertions the operator makes. Neither can be established from a record alone, which is why L4 exists: a third party replays the ledger and checks for themselves.


gauge gate

Exit non-zero when a record does not reach a given level. This is the command intended for a pipeline step.

sq gauge gate record.yaml --level L2 --quiet
Option Default Purpose
--level {L1..L4} L2 Minimum level required
--quiet off Suppress the report, leave the exit code

Without --quiet it prints each level with its blockers:

L1  pass
L2  fail
      held-out split is not sealed
      no contamination probes recorded

highest level met: L1

gauge show

Print a record as a scorecard: subject, authority, task set, measures, gates, verdict and level.

sq gauge show record.yaml

gauge verify

Recompute what can be checked without re-running the agent: the harness digest against the spec on disk, and the presence of the referenced ledger.

sq gauge verify record.yaml --spec harness.yaml

A digest mismatch means the spec moved after the record was written, so the record describes a harness you no longer have. Replaying the ledger is the remaining step, and it belongs to whoever is checking the claim.