Gateway draft assistant — 6.1
Essai 01 first version, built 2026-10-03. An evidence-linked drafting demonstration for two invented governance gates at Fellbridge Station Access. It has no approval, advance or operational-permission action.
Gate distinction
The concept review asks what supports further option development. Its original fixture has four requirements: scope/position, decision request/accountable role, cost baseline and temporary-access outline. Delivery preparation adds agreed-window evidence and disposition of decision conditions, and asks for route validation rather than an outline: six requirements in the original fixture.
Applicable clauses come from the shared invented playbook. Unknown context retains unresolved candidates. High-cost or novel profiles add independent-estimate or trial/fallback requirements. The fixed shelf supplies no dedicated evidence for those additions; a filled cost field does not establish an independent estimate.
Every requirement distinguishes missing, contested, superseded and current-source coverage. A current source can still say that an agreement is outstanding. No row becomes “satisfied”, “approved” or completed assurance through text presence or local wording review. Current guidance status is also exposed.
Working output
Select a gate, inspect the evidence matrix and assemble five editable sections: gate purpose, mandatory matrix, applicable guidance, decision/conditions and an unresolved checklist. Exact sources appear in the draft and in expandable snapshot excerpts. Source links lead to the root shell’s current shelf, while the draft retains the precise original packet for historical citations.
Check each saved section against evidence, then enter a reviewer label and note and confirm that gaps remain explicit. Copyable Markdown includes the gate, source snapshot, author labels, citations, matrix and unresolved checklist. The checklist stays open after local review. The shared shell saves/reopens full working JSON and handles explicit source packet transfer.
A gate or source change preserves old wording and its original snapshot, clears review and marks it stale. Editing, review and Markdown are unavailable until explicit reassembly. Assembly replaces the draft, so the interface advises saving a version first. User edits are labelled user-authored and require another section check.
Recorded AI comparison
The Codex implementation assistant genuinely authored a short AI synthesis during the 2026-10-03 build, using the original invented packet at concept review. Its text is recorded in this module, labelled as such and never generated live. It is unavailable after changing the packet or selecting delivery review. Imported provenance must retain the exact original packet/context; changed recorded text must be user-authored.
The comparison is honest: extraction exposes fields and preserves exact quotations; AI wording selects a gate-specific emphasis and suggests reviewer questions. It supplies no new evidence and needs factual and editorial review. Neither preparation time nor correction rate has been tested with human users.
Limits
This first version shares the editable drafting component with the meeting assistant but uses distinct gates, requirements, output and tests. It has two invented templates, one source shelf, no independent evidence authentication and no live grounded drafting service. Thus it does not claim the original method’s full functional completion. A complete-looking draft is not engineering acceptance, an investment decision, a return-to-service certificate or permission to work. Local review is an editable working record, not sign-off or a trusted audit trail.
Verification
Run node --test apps/gateway-draft-assistant/app.test.mjs from the repository root. Tests cover the 4-versus-6 requirement distinction, missing cost and contested route, review that cannot close assurance gaps, rejected approval/advance actions, unknown/high/novel profiles, source presence versus adequacy, stale gate/source changes, immutable recorded provenance, changed source conditions, user edits and escaping, malformed imports, unchanged failure states and JSON reopen.
Root-owned build, browser and publication checks are separate from these model tests. This module makes no deployment or human-adoption claim.