Workflows & credits

workflow on POST /v1/projects is a JSON array of stages. arbitr runs them in order. You do not assign vendors or sequence the later stages yourself.

Stages

AI only:

workflow=["AI_TRANSLATION"]

AI then human translation and edit:

workflow=["AI_TRANSLATION","TRANSLATION","EDIT"]

due_date (YYYY-MM-DD) is optional and informational. It never blocks or parks the project. GET /v1/projects/{id} echoes it on assessment and sets due_date_feasible to null if you sent no date.

Two wallets

Check the org balance (any valid key, no extra scope):

curl -s https://api.arbitr.ai/v1/credits/balance \
  -H "X-API-Key: abr_live_…"

intelligence_credits / trust_credits are remaining. Values are numbers (trust can be fractional). abr_test_… keys do not consume credits.

While a project runs

Poll GET /v1/projects/{id} (verify:read). assessment is only on this detail call, not on create or list.

• assessment.credits.intelligence — always present (required, available, sufficient)
• assessment.credits.trust — present when the workflow has a human stage; otherwise null
• review — human-stage lifecycle; fields stay empty until a review is requested

If credits run out mid-flight the project parks. An intelligence shortfall sets status to awaiting_payment; a trust shortfall sets review.status instead and leaves status unchanged (see Project lifecycle). Top up in the arbitr UI, then resume (verify:submit):

Still short → 402. Not waiting for payment → 409.

Both resume routes return the normal error envelope: 402 payment_required (with required / available / shortfall), or 409 not_awaiting_payment (intelligence resume may also include current_status). See Errors.