Skip to content

Evals

Evals grades outputs with versioned Evaluators, gates every Change before merge, and keeps each Worker’s track record.

Evals grades outputs with versioned Evaluators, gates every Change before it merges, and keeps each Worker’s track record. An Evaluator is code, a model, an Agent or a Human, and returns a verdict, never an Approval.

Summary

PartThis System
GroupMoves to Agents
StatusBeta: it works, and its Verbs may still change
ShelfFront: offered from a Startup’s first days

You can ask:

  • “How good are the support agent’s answers?”
  • “Grade this agent’s last 50 replies”
  • “Block changes that make answers worse”

What you get

PartThis System
StartsOn for every Startup at creation, with nothing to install.
Built fromFunctions, AI gateway, Event streams

Nouns it owns

NounDefinition
EvaluatorAn Evaluator is a versioned Function that grades an output against a rubric and returns a verdict.

Verbs

VerbObjectApprovalEvent
Evaluator.createEvaluatorNoneEvaluator.created
Action.evaluateActionNoneAction.evaluated

Events

DirectionEventHow
EmitsEvaluator.createdBy Evaluator.create
EmitsAction.evaluatedBy Action.evaluate

States

None.

Who operates it

PartThis System
AccountableQuality Lead
TeamQuality Team
SeatA role agent holds it until a Human is seated.
Autonomysupervised
Also calls its VerbsReviewer, Tester

Approvals and limits

None.

Processes that use it

ProcessHow it uses it
VerificationA step calls Action.evaluate
HiringA step calls Action.evaluate

Rail underneath

None.

Views and dashboards

ViewShowsMetrics
The Quality Lead’s viewThe Metrics its accountable Role answers for.Escaped defects

Access

What each Access level may call here.

Access levelMay call
ownerEvery Verb (6)
adminEvery Verb (6)
memberEvery Verb (6)
viewerRead only (2 of 6)

Import and export

None.

Fees and pricing

None.

Availability and limits

PartThis System
StatusBeta: it works, and its Verbs may still change

Compliance notes

Controls it carries:

None.

Example

An E2E managed service for customer support rewrites its refund-reply prompt. Evals replays 400 past Conversations through it; the new prompt scores no worse on accuracy or tone, so the Change merges. The Agent’s track record, which earns it wider Mandates, updates with the result.