renderfact

A governed docs-as-code render framework: one full-candor markdown/YAML source becomes many governed, audience-specific projections — styled DOCX, accessible + archival PDF, diagrams, decks, posters — each carrying hidden provenance and passing deterministic QA gates.

Generic core, private skin. The engine is domain-agnostic and OSS; an organisation’s house style, terminology, and disclosure rules live in a private “skin” it consumes. The demo ships a fictional rail operator’s procurement dossier so you can taste the whole pipeline with zero real-world coupling.

Start here

• Getting started — install, then render the bundled demo.
• How-to recipes — render a document, gate it, round-trip a diagram, capture a decision.
• Reference — the render command surface, environment variables, the provenance schema.
• Explanation — the architecture doctrines: the D8 harness-optional contract, the D16 fuzzy-gate, projection-aware provenance, and the editable-diagram round-trip.

What makes it different

• Round-trip, not one-way. A rendered DOCX or an editable diagram (draw.io / Visio) can be hand-edited and re-ingested: the mechanical diff flows back to the source, and the human intent flows to a decision log.
• Harness-optional AI. Every LLM-touching step degrades across three modes — your own assistant (harness), a human pasting into any chat LLM (copy-paste), or, opt-in, a configured model — behind one contract, so no step is locked to a vendor.
• Deterministic first, LLM only past a gate. Steps run a deterministic result first and escalate to a model only when a confidence gate is missed. Most invocations cost zero tokens.
• Provenance + QA built in. Hidden source identity in every editable artifact (stripped for external projections), and a fail-closed gate chain (style, links, PDF/A conformance).

Repository: github.com/Wombat164/renderfact (MIT).