KOkobaltlabs
risk-assessment

Start Training Run

Start a "Teach Kobalt" run from feedback on a rating the reviewer disagrees with.

When the suggested rating differs from the factor's current effective level, it is applied as a manual override here — not by the agent — so it stands whether the agent succeeds, gives up, or crashes. Otherwise the run records feedback only and leaves the displayed rating unchanged.

post/risk-assessment/factor-assessments/{factor_assessment_id}/training-runs

Request

  • Base URL
    https://api.kobaltlabs.com (the document declares no server; this is the origin it was published from)
  • URL
    https://api.kobaltlabs.com/risk-assessment/factor-assessments/{factor_assessment_id}/training-runs
  • Auth
    none declared

Path parameters

factor_assessment_idstring uuid required

Request body

guidance_textstring nullable
suggested_inherent_risk_labelstring uuid nullable
force_ratingboolean
reference_upload_idsstring[]
reference_urlsstring[]

Response

Successful Response

training_run_idstring uuid required
factor_assessment_idstring uuid required
status'pending' | 'running' | 'completed' | 'failed' required
outcome'promoted' | 'gave_up' required

How a RiskFactorTrainingAgentRun ended.

Both are terminal tool calls, not failures: an agent that cannot justify a rewrite is expected to GIVE_UP with questions rather than promote one.

proposal_status'active' | 'proposed' | 'hidden' | 'inactive' required

Where one risk_factor_versions row stands.

A factor's name and slug are permanent; its description, assessment guidance, rating rubric and applicability logic are versioned, so a rewrite never destroys the text that produced historical ratings.

ACTIVE is what the next assessment run uses, and a partial unique index (uq_risk_factor_versions_one_active) enforces at most one per factor. HIDDEN is a training agent's working draft — created mid-run and never shown to a reviewer. PROPOSED is a draft the agent promoted and the reviewer has not yet accepted or rejected. INACTIVE covers both "superseded" and "rejected"; the two are distinguishable from the training run's outcome.

createdstring date-time required
started_atstring date-time nullable required
finished_atstring date-time nullable required
experiment_countinteger required
errorstring nullable required

Changes

No recorded changes to this operation across all 8 revisions of this API.