delivery
AI output stays a proposal until a human approves it
Structured plans are validated, reviewed, and only then materialized into delivery milestones and tasks.
Loading Workbench…
Technical evidence
This page links the main system claims to tests, source files, runbooks, and the interactive demo. It also marks which results came from fixtures and which ones were observed in a deployed environment.
Synthetic data only · isolated 60-minute demo · statuses below distinguish verified evidence from targets and planned work.
Build evidence
production / dbf6797d49c3
Current CI claims
—
Eval cases
15
Offline fixture gates
100%
CI provenance is not attached to this build.
No CI evidence commit is attached to this build.
Quick path
Architecture trail
Requirements and documents enter a tenant-scoped boundary.
Redaction and retrieval produce opaque source references.
Zod, citation, injection, and policy checks guard the output.
A human decision keeps valid AI output inert until reviewed.
Jobs, tasks, updates, audit, and recovery remain visible.
Pick a starting point
Each link creates an isolated demo workspace and opens near the relevant screen. You can change roles from the persona bar at any time.
Follow the shortest path from seeded requirements to a governed delivery artifact.
Inspect retrieval, repair, validation, citations, origin, and the human decision.
Start at the approval gate, then follow the work into tasks and customer updates.
Open the admin view for access controls, audit history, and failed-job recovery.
Claim registry
The manifest is generated from the same checked-in source that powers this page. A target is not presented as a measurement, and planned enterprise work is visible instead of hidden.
delivery
Structured plans are validated, reviewed, and only then materialized into delivery milestones and tasks.
ai
The grounded showcase connects retrieval, repair, normalized checks, citations, cost, latency, and the approval outcome in one evidence packet.
security
RBAC, organization-scoped lookups, transaction-local RLS context, private object storage, and synthetic demo isolation work together.
reliability
Atomic fenced claims, bounded backoff, partial-batch handling, dead-letter parking, and a visible retry path make recovery inspectable.
ai
Fifteen synthetic cases across three prompt variants exercise schema, coverage, citation, and injection gates without cloud credentials.
security
OIDC login and SCIM provisioning support controlled role mapping, hashed bearer tokens, and immediate deprovisioning.
platform
Versioned event envelopes travel through the durable job path with HMAC signatures, replay windows, SSRF guards, and bounded response capture.
security
Org admins can preview and tune retention windows while a bounded scheduled run records counts, outcome, and sanitized errors.
platform
CDK captures encrypted storage, SQS/DLQ, Lambda workers, scheduled cleanup, alarms, budgets, and deployment trust boundaries.
reliability
OpenTelemetry spans and safe trace context connect API requests, job dispatch, generation, and worker execution without exporting tenant identifiers or content.
reliability
Availability, API latency, queue start time, generation outcomes, and recovery objectives will be reported separately from aspirational targets.
platform
A public Vercel, Neon, AWS, and optional IdP deployment will be promoted only after health, checkpoint, retention, webhook, and capped Bedrock smoke checks pass.
Machine-readable
Safe build metadata and evidence links for automated review.
Narrative
Context, tradeoffs, failure modes, and evidence index.
Reproducibility
A commit-bound proof bundle is published only after the quality and browser jobs both pass.