Product decision
Can the product recommend relevant experience for a target role while preserving factual integrity and user control?
Completed example · Vivid Resume
One grounded recommendation. One connected packet.
This example shows how the same decision ID moves through direction, implementation, evaluation, readiness, and release.
Connected trace
DEC-001 becomes evidence—not merely output.
The example remains intentionally narrow: help one user connect one target role to evidence already present in their experience without inventing facts.
Vertical workflow
Create target → read owned experience → generate grounded recommendation → show provenance → edit or reject → save.
Quality boundary
Every claim must map to user-provided evidence. Unsupported facts are an unacceptable failure, regardless of fluency.
Readiness evidence
Advance only after schema, policy, API, authorization, UI state, grounding, failure, and recovery evidence agree.
Controlled exposure
Release to a bounded closed-beta cohort with monitoring and reversal. Do not infer adoption or outcome proof.
Next rule
Any observed unsupported claim becomes a reproducible evaluation case, regression test, and revised constraint.
Example packet
What each artifact contributes.
No single document proves readiness. The identifiers connect intent to implementation and implementation to evidence.
Why this workflow?
User: a job seeker tailoring a résumé.
Human authority: accept, edit, or reject every recommendation.
Out: automatic invention, submission, or employer decisions.
Where must it work?
Product: input, loading, provenance, edit, rejection, save, error.
System: authorization, grounding service, model boundary, persistence, observability.
What counts as acceptable?
Representative roles and histories; factual-grounding rubric; zero tolerance for invented employers, achievements, credentials, or dates; latency and cost limits.
What is actually proven?
Layer tests do not establish integration. Integrated behavior does not establish production outcomes. Evidence moves one explicit state at a time.
Who may expose it?
The accountable owner approves a bounded cohort after evidence review, with visible feedback, monitoring, and a reversible deployment path.
What changes next?
User corrections, unsupported-claim incidents, abandonment, failure, latency, and cost update the evaluation set, rules, and next decision.
Create your packet
Start with one decision ID.
Complete the canonical brief, then reuse its decision and feature IDs through every artifact.