Skip to main content

​
Summary

Since closing the M0-M6 baseline, we have focused on four connected outcomes: broadening the regulatory evidence Forge can work with, evaluating that evidence path against labels created from the source material, showing rule-backed issues directly on a representative vessel model, and carrying those issues into work-package readiness. These are not separate experiments. The regulatory corpus supplies the requirements; the evaluations test whether Forge reads and applies them faithfully; the model viewer shows where an issue exists on the vessel; and the readiness board shows which part of the build is responsible for resolving it. Together they move Forge from a completed technical baseline toward a workflow a yard can inspect, challenge, and use.

​
What we shipped

A broader, traceable regulatory corpus
  • The initial small-passenger-vessel fire-protection scope is connected end to end: the relevant sections of 46 CFR Parts 114, 116, 118, and 164 now join the acquired Coast Guard Plan Review Guides. A Plan Review Guide, or PRG, explains the evidence a reviewer expects to see for a particular plan or system. Forge preserves the guide’s relationship to the regulation rather than treating the guide itself as law.
  • More kinds of interpretive material now fit the same evidence graph: Marine Technical Note 01-13, Navigation and Vessel Inspection Circulars, a Coast Guard policy letter, and selected Marine Safety Manual chapters can be read with their own source-native structure. “Source-native” means a reviewer can return to the section, item, enclosure, or chapter as it appears in the original document.
  • Additional CFR sections activate previously recorded citations: Parts 72, 76, 95, 159, and 161 were added where the interpretive documents pointed to them. Forge can preserve a citation before its target is present, show it as unresolved, and connect it later when the governing text enters the corpus.
  • Scanned and historical material has a governed route into Forge: local optical-character recognition, or OCR, can make scanned documents searchable while preserving how the text was produced. Original editions and later changes remain distinct, so a 1997 circular is not silently replaced by its amended version.
  • Terminology begins connecting the corpus: defined phrases such as “Area of refuge” can link a CFR definition, Coast Guard guidance, applicable rules, and the vessel space using that term. The relationship is based on exact recorded language, not a guessed synonym.
Evaluations that improve the system before partner use
  • The first citation evaluation failed usefully: the initial run scored 62 out of 80 and exposed 18 genuine document-reading gaps, including nested bullets, table references, and citations distributed across a sentence. We corrected the source-grounded reading rules rather than weakening the labels, and the same set then passed completely.
  • Labels come from reading the source, not from asking Forge for its own answer: each example records what a human reviewer can verify in the document. The evaluation then joins that source-native label to the system output, making a missing or incorrect join visible.
  • The evaluation surface now reaches beyond citation extraction: Forge also checks rule-applicability precision and recall, whether it refuses for the correct reason when vessel information is incomplete, and whether it identifies required evidence that has not been supplied.
  • New document families are evaluated when they enter: representative examples are labeled during onboarding, so corpus breadth and quality evidence grow together. A new parser is not treated as complete merely because it produced output.
Rules and evidence become visible on the model
  • The CAD-aware viewer now has a real web application: the browser viewer and the Forge API share one published data contract, which means the model, measurements, citations, and issue rows are described consistently rather than being redefined by the interface.
  • Engineering measurements appear beside the geometry they describe: principal dimensions and per-component volumes are calculated from the engineering geometry and shown as cited overlays. The viewer helps people inspect the result; the underlying measurement record remains the authority.
  • A representative vessel model demonstrates the complete rule-to-part join: a synthetic small-passenger-vessel assembly provides named spaces and components without exposing partner designs. A real Part 116 requirement, carried through its PRG evidence chain, produces an open fire-protection gap on the Area of Refuge. Selecting the issue highlights the affected space and shows the citation chain.
  • The view responds to governed decisions: when the compliance gap is resolved through its signed decision path, it clears from the open-issue overlay. The model is therefore showing the current governed record, not maintaining a separate issue list.
  • The demonstration is shareable without partner infrastructure: the representative vessel and public-domain model can be loaded as static demonstrations, allowing prospective partners and reviewers to understand the workflow before connecting yard data.
Readiness reaches the work-package level
  • Engineering records can be connected to the work they affect: Forge now has a governed link between a compliance gap, reapproval, or other readiness record and the relevant work package. The link is explicit because the yard, not an algorithm, knows which module or production package owns the work.
  • The module-readiness board uses those recorded links: each work package shows its open blocker counts and readiness state. Vessel-level concerns remain visible, while package-specific issues appear where production teams can act on them.
  • Resolving the underlying record updates readiness: when a cited gap is resolved, the board reflects the governed resolution instead of requiring a second manual status update.
  • Operators have a controlled linking action: authorized users can connect a record to a work package without turning the read-only dashboard into an uncontrolled editing surface.

​
What we learned

  • Corpus depth and product usefulness reinforce each other. A viewer becomes more valuable when it can show a real requirement, and the corpus becomes more valuable when a reviewer can see where that requirement lands on the vessel.
  • An evaluation earns confidence by being allowed to fail. The first citation run exposed real weaknesses because its labels were independent of the system being tested. Correcting those weaknesses gave us stronger evidence than a first-run pass would have.
  • Labeling at onboarding is the scalable discipline. Evaluating each new document family when it enters prevents quality work from becoming a separate cleanup project.
  • An unresolved citation is useful information. Preserving the citation as printed gives a reviewer an honest research task. Guessing its intended destination would turn uncertainty into false authority.
  • Historical editions must remain visible. Regulatory and guidance documents change over time. Keeping editions distinct allows a future decision to cite the material that actually governed it.
  • Work-package readiness depends on recorded ownership. A blocker becomes useful to production when the project explicitly connects it to the package responsible for resolving it.

​
Blockers & open questions

  • The corpus is ready for partner and class review of its labels, applicability assumptions, and interpretation relationships. That feedback will expand coverage while preserving the distinction between binding authority, guidance, and yard practice.
  • The representative vessel proves the full workflow without exposing proprietary designs. Partner onboarding will map the same evidence flow onto the yard’s CAD structure, naming conventions, work packages, and approved deployment environment.
  • Regional interpretations and Officer in Charge, Marine Inspection (OCMI) practices are often learned through operational experience rather than a single published document. Forge will capture those qualifiers through partner walkthroughs and accountable review instead of inferring them from incomplete text.
  • Rich model selection and in-viewer authoring will be introduced when partner use identifies the interactions that create the most value. The current read-only foundation keeps the authority boundary clear while those needs are observed.
  • We will not fine-tune models ahead of the evidence. Production-scale labeled material, approved decisions, and confirmed outcomes will determine when a specialized model is justified. The governed corpus and evaluation record come first.
  • Some lower-quality scanned sources remain useful at the document level even when their structure cannot be extracted reliably. Forge keeps that limitation visible rather than creating false precision from noisy text.

​
Next week

  • Complete the remaining acquired Plan Review Guides, deepen the policy-letter structure where the source quality supports it, close the remaining document-segmentation questions, and clearly separate work Forge can complete independently from features that should be shaped by a partner walkthrough, real pilot data, or observed interaction needs.