# Workflow — {{system_name}}

`/architect` reads this file to decide what runs next. The stage table below is parsed. Keep the
column order (`stage | agent | input | output | gate`), keep the stage ids, and keep one row per
stage. Add a stage by adding a row and a matching gate row; do not renumber or rename existing ids.

## Stage table

| stage | agent | input | output | gate |
|---|---|---|---|---|
| frame | discovery-analyst | the goal argument | `00-frame.md` | `gate.frame` |
| discover | discovery-analyst | `00-frame.md`, the repository | `01-discovery.md` | `gate.discovery` |
| flow | experience-designer | `00-frame.md`, `01-discovery.md` | `02-user-flows.md`, `03-experience-blueprint.md` | `gate.flow` |
| decide | principal-architect | `01-discovery.md`, `03-experience-blueprint.md`, the repository | `SYSTEM.md`, `adr/*`, `architecture.json` | `gate.decisions` |
| cost | economics-analyst | `SYSTEM.md`, `01-discovery.md` | `04-roi.md`, `prices.json` | `gate.economics` |
| secure | trust-reviewer | `SYSTEM.md`, the repository | `05-trust-boundary.md` | `gate.trust` |
| prove | eval-engineer | `05-trust-boundary.md`, `02-user-flows.md` | `06-evals/cases.jsonl`, `06-evals/rubric.md` | `gate.evals` |
| operate | delivery-engineer | `SYSTEM.md`, `06-evals/rubric.md` | `WORKFLOW.md`, `SOP.md`, `07-runbook.md` | `gate.operate` |
| verify | independent-verifier | every artifact above | `receipts/{{today}}-verify.md`, `review.md` | `gate.verify` |

## Gate table

A stage is complete when its gate is `PASS` in `architecture.json`, or `SKIPPED` with a written
reason. Anything else — `FAIL`, or absent from the gates map — means the stage has not run.

| gate | green when |
|---|---|
| `gate.frame` | a one-sentence outcome, named non-goals, and a falsifiable kill criterion |
| `gate.discovery` | at least one answered question per bank section; zero `[assumed]` tags (`[unknown]` is allowed) |
| `gate.flow` | every flow has a failure branch and a human step; the experience blueprint covers every stage |
| `gate.decisions` | four decisions each `MADE` or `OPEN` with an evidence pointer; every `OPEN` carries a dated deferral cost; every plane has an owner |
| `gate.economics` | every price row has `source_url` and `retrieved_at` within 90 days; a sensitivity band on three drivers; payback stated in months, no return-rate acronyms |
| `gate.trust` | retrieved text traced to the line where it becomes labelled data; every irreversible tool gated by a human step |
| `gate.evals` | at least 10 cases, at least one refusal case, at least one injection case; the harness runs; the injection case fails closed |
| `gate.operate` | every runbook command exists and dry-runs exit 0; rollback and owner named |
| `gate.verify` | a receipt written by the verifier in a fresh context; every evidence pointer re-derived |

## Fix-first order

When more than one gate is red, work the highest reversal cost first. Within `gate.decisions`,
when several decisions are `OPEN`, the order is `trust`, `run`, `loop`, `model` — trust rises in
cost fastest, because it scales with the number of tools already shipped.

Generated by AI Architect · https://www.frankx.ai/ai-architect