Inside the Four-Stage Technical Drafting Directed Graph
Vellum & Loom Systems structures technical writing as a directed acyclic graph rather than a single prompt pass. Each stage is versioned independently so teams can inspect inputs, intermediate artifacts, and downstream drafts in isolation. A single extracted claim graph feeds parallel synthesis branches for different document formats, while every branch converges on one shared verification gate before publication.
Stage One: Corpus Snapshot Ingestion
Ingestion converts raw technical inputs into immutable snapshots before any downstream analysis begins. Each file receives a document version identifier, exact section offsets, and a cryptographic content hash, while workspace access boundaries remain attached to the snapshot metadata. The parser preserves structured tables and code blocks together with their surrounding qualifiers instead of flattening cells into plain text strings. When an input contains unreadable optical character recognition segments or malformed markup, the stage logs an explicit parser reject rather than guessing at the missing content. Downstream stages only read from verified snapshot records, ensuring that every later reference points to a fixed byte range and a known source state.
Stage Two: Claim-Graph Extraction Pass
From the ingested snapshots, the second stage extracts a structured graph of atomic claims. Every node records a clear subject, predicate, scope, measurement units, operating conditions, and exact source spans back to the parent snapshot. The extractor separates directly quoted facts from derived assertions so reviewers can see where interpretation begins. If two source documents disagree on a parameter or condition, the graph records both claims and links them as an explicit contradiction or an unresolved node rather than picking a winner. All graph edges remain bound to specific snapshot versions, keeping the factual substrate stable across subsequent drafting runs.
Stage Three: Voice-Constrained Branch Synthesis
Multiple synthesis branches can run in parallel from a single versioned claim graph, producing an implementation guide, a release note, and an architectural whitepaper from the same underlying nodes. Each branch applies a structural template and cadence profile that governs sentence rhythm and organization without granting permission to introduce ungrounded claims. Whenever a draft requires a bridging explanation or transition that is not backed by an existing node, the synthesizer flags that span as an explicit review item. Because the system separates stochastic prose generation from deterministic artifact replay, editors can re-render unchanged sections or inspect prior branch outputs without re-running the entire pipeline.
Stage Four: Deterministic Citation Verification
Before any branch output can be exported, it passes through a shared verification gate. Mechanical checks confirm that every citation pointer resolves, snapshot hashes and section anchors match the active corpus, and numerical values and units in the prose match their linked claim nodes. Any mismatch triggers an immediate gate failure. Because mechanical checks cannot prove full semantic entailment, nuanced support judgments remain visible for human editorial review. When an upstream source or claim node is edited, the graph invalidates only the affected descendant spans rather than discarding unrelated branches. Completed runs conclude with a signed editorial approval record and a reproducible export manifest.
Operating the Human Editorial Review Boundary
Operational reliability depends on predictable failure handling across every stage of the graph. Each stage runs under bounded retry rules and strict execution budgets, and stale upstream dependencies block downstream output until refreshed. Named owners resolve flagged contradictions or failed gates directly, because the pipeline stops gracefully rather than silently skipping missing evidence.