Document intelligence · evidence-to-manuscript workflow

Research product prototypeReviewed September 2026Auditable AI workflows

Turn a life archive into a reviewable manuscript.

Bibliography is built around a simple product promise: preserve the evidence, make the proposed change inspectable, protect the author’s voice, and export a document that can be reviewed outside the system.

PythonTyped evidence recordsProvenanceQuality gatesDOCX export

The product problem

Personal archives are rich but difficult to use: email, messages and documents arrive in different shapes; chronology and attribution drift; a fluent generated paragraph can still be unsupported; and a “final” document can hide how it was assembled.

Bibliography treats this as an evidence and workflow problem, not a prompt-writing problem. It gives each stage a boundary, a machine-readable contract and a reviewable artifact.

For an author or editor

Move from raw correspondence to candidate material without losing source context, unresolved questions or the ability to reject a proposed change.

For a product team

See how retrieval, generation, human review, provenance and export can be composed into an operational system with explicit failure modes.

A controlled evidence-to-document pipeline

  1. 01 / INGESTNormalize source evidenceAdapters turn email and other channel exports into typed records with stable IDs, timestamps, provenance and attachment metadata.
  2. 02 / ROUTEFind the right chapter or questionSummaries, entity context and routing services guide evidence toward the part of the manuscript where it can be reviewed.
  3. 03 / PROPOSEGenerate a candidate changeProposal CLIs emit JSON and Markdown patches instead of silently mutating the manuscript.
  4. 04 / SAFEGUARDCheck anchors, agency and provenanceConfidence, structural anchors, narrative safeguards and fact packs keep unsupported or unsafe changes visible.
  5. 05 / APPLY AND EXPORTCommit a traceable resultChapter files, reports, run manifests, metrics and DOCX exports remain inspectable as a complete run.

What the build says about my skills

Systems thinking

The architecture separates evidence, services, NLP, manuscript, storage, utilities and CLI wiring so the workflow can evolve without collapsing into one opaque script.

AI with operating controls

OpenAI access is lazy and boundary-controlled; generation is surrounded by schemas, confidence thresholds, source snapshots and human review rather than treated as an authority.

Reproducibility

Project-relative paths, deterministic helpers, idempotent commands, run IDs, metrics and artifact indexes make a result explainable after the original terminal session is gone.

Editorial product design

The system understands that “good output” includes chronology, attribution, agency, tension, unresolved questions, export quality and the ability to roll back or defer.

Why it could become a commercial product

The product opportunity is a trustworthy workspace for evidence-heavy writing: memoir, investigative narrative, family history, case records or other long-form documents where provenance matters as much as fluency.

  • Source coverage and missing channels are explicit instead of silently treated as complete.
  • Every proposal can carry a source trail and a confidence or review state.
  • Quality reports turn editorial risk into a visible operating queue.
  • Exports are deliverables with validation and archival, not just a download button.
  • The architecture is mature enough to support product discovery without pretending the commercial product is already launched.

Current status: a research workflow and product direction. It does not claim publishing-scale adoption, autonomous authorship, completed citation activation or customer revenue.