Verify work

Choose a verifier and submit a deployed URL or company Library references. Reviews run as company jobs and incur model usage charges. A previous verifier pass is not a prerequisite. You do not supply routes, roles or viewports.

Identity, voice and research references default to the company's saved Library documents. An explicit references entry overrides its default using key, title and library_path. Documents must be UTF-8 text. Ads manifests refer to PNG, JPEG, WebP or MP4 assets through Library paths. Missing required evidence is reported before starting paid judges.

The market-research judge compares the five frozen documents for material contradictions, citing both conflicting statements and their document locations. It does not verify sources, factual accuracy, research depth or market viability. A pass means no material contradiction was found. Existing jobs retain their frozen criteria; submit with a new idempotency key to use updated criteria.

Submit and poll

cofounder work verify --name copy --submission '{"url":"https://example.com","idempotency_key":"11111111-1111-4111-8111-111111111111"}'
cofounder work verify --name copy --job-id <job-id>

For a Library copy document, replace url with:

{"references":[{"key":"copy","title":"Launch copy","library_path":"marketing/launch/copy.md"}]}

Keep the idempotency key in the submission. Reuse it for identical retries; use a new key after revising work or references. Poll after poll_after_seconds.

Independent judges run in parallel, with at most three active judges per job. Brand and product reuse the Copy review step. Ads run independent copy review using the same rubric, a campaign coverage judge and per-asset creative judges. Creative judges use Inkling to inspect video frames and encoded audio together. All required reviews must pass. Steps return assessments by rubric, rubric revisions, findings and, for ads, target_keys identifying the reviewed variants.

Only passed means all required checks passed. changes_required reports defects; uncertain or error means inspection could not complete. Changed Library evidence produces stale. URL results describe the surface observed during review; keep the deployment stable while judges inspect it.

Verification uses the deployed app, including authenticated and paid screens. Repository submissions and local-app judges are retired. SEO uses browser and HTTP inspection; an OpenSEO integration is not included.

Test access

Save a dedicated, already-entitled test account through the encrypted env-file API (POST /env-files, using a user login and the selected company). Include LOGIN_URL, USERNAME and PASSWORD in that encrypted file. This store is separate from Library and does not require a managed hosting project. Use the returned file ID in the submission:

{"name":"product-dev","submission":{"url":"https://app.example.com","test_access":{"credentials_ref":"<encrypted-env-file-id>"},"idempotency_key":"<new-uuid>"}}

The submitted URL must have the same origin as the saved login URL. The runner fills credentials through a trusted browser tool; credential values travel separately from job inputs, Library evidence and model instructions, and the login tool never returns them. Each judge gets its own browser session. The workflow judge may create disposable test records and clean them up; other judges observe.

The account must already have access to paid features. Checkout itself requires payment test mode. Judges do not sign up, buy, publish or change real customer data. Missing access, expired credentials, MFA or missing entitlements block completion instead of silently excluding protected screens.

Jobs submitted under the previous contract return a retirement notice and require a new submission. They cannot restart the removed local inspection procedures. Existing stored reports are not deleted.

Operation