The Build Path

The milestones, from foundations to a CAD-aware IDE. What’s shipped and what’s next.

In plain terms: Forge is being built in stages. Each milestone lands a specific set of capabilities, and each one builds the clean data and the trust the next one needs. This page shows the roadmap and where we are on it.

The milestones

Milestone	Theme	What lands
M0	Foundations	Database schemas, per-yard isolation everywhere, source registry, event/audit logs, decision ledger schema, capability catalog schema, isolation tests
M1	Rules & Retrieval	ABS / USCG ingestion pipeline, citation manager, basic rules retrieval, prior-vessel memory ingestion, structured UI skeleton
M2	Governed Decisions	Decision ledger writes, rule applicability decisions, capability-gated agents, rule supersession events, re-evaluation queue
M3	Geometry Baseline	CAD metadata import, STEP / IGES geometry service, measurement records, simple clash and clearance checks
M4	Verification & Eval	Ordered verification gates, confidence filter, numeric geometry grounding, evaluation thresholds, isolation tests at scale
M5	Controlled DesignOS Pilot	Partner operating policy, build-readiness checklist, change-impact workflow, issue export, structured feedback
M6	CAD-Aware IDE	Visual model viewer, annotations, structured approval flows, deterministic review packages, CAD connectors where permitted

Where we are

The named M0-M6 build baselines are complete.

Forge is now in post-baseline work: widening the regulatory and interpretation corpus, deepening labeled evaluation, and preparing the controlled baseline for partner use and feedback. For each partner environment, production identity, key custody, deployment, and permitted CAD connections are configured to that yard’s requirements. The Research Log tracks this work week by week.

How the path is ordered

The rollout follows the same logic as How Forge Fits: start where the leverage is highest and the risk lowest, then expand.

• M0 to M2 establish the foundation: trustworthy sources, per-yard isolation, the data model, rules ingestion, the decision ledger, the capability catalog, and a basic UI.
• M3 to M5 add geometry, enforce that numeric claims are grounded, route rule changes, and stand up the build-readiness workflows for a first pilot.
• M6 and beyond bring evidence into the CAD-aware review experience, connect to existing yard tools where permitted, expand the governed corpus, and build the outcome-driven learning flywheel.

Throughout, the trust principle holds: the AI can retrieve, compare, draft, flag, and propose, but it never becomes the design authority. Structured human approval stays authoritative.