DANIEL WIPERT
← DEPT. 02 · CHORUS AI SYSTEMS
PIPELINE LIVE CHORUS AI SYSTEMS

Dossier

CLOSED-LOOP DOCUMENT INTELLIGENCE

Lock the facts. Three models summarize on their own. Any contradiction fails the run.

VISIT THE FULL SITE → VIEW THE CODE
0.0
CONTRADICTION TOLERANCE

HOW IT WORKS

From a PDF to a dossier you can audit

  • IN / OUT
  • CODE
  • AI MODEL
  • CHECKPOINT
  1. IN Source PDF any document
  2. AI MODEL Extract facts atomic claims
  3. CODE Lock the fact list immutable from here
  4. THREE VENDORS, INDEPENDENTLY
    AI MODEL Summary A
    AI MODEL Summary B
    AI MODEL Summary C
  5. CHECKPOINT Contradiction check 0.0 tolerance

    One contradicted fact fails that summary

  6. CODE Assemble claims cite source spans
  7. OUT Dossier gaps labeled

    Not sure? It halts visibly

Seven stages, temperature zero, no agent autonomy. Control flow lives in code. Simplified. The full story is on the project site.

The problem

Ask a language model to summarize a document and you get something fluent. Whether it’s faithful is another question. For a compliance file, a contract, or a board pack, fluent and wrong is worse than no summary at all.

What I built

Dossier is a closed-loop, multi-model document intelligence pipeline. A PDF goes in. What comes out is a dossier in which every claim traces back to a span of the source, every gap is labeled as a gap, and the system stops visibly when it can’t be sure.

How it works

  1. Extract and lock. The pipeline pulls atomic facts out of the document and locks them into a fact list. Nothing downstream can change it.
  2. Summarize independently. Three models from different vendors each write their own summary. Different makers fail in different ways, so they’re harder to fool in the same direction.
  3. Check against the facts. Every summary is tested against the locked fact list. The tolerance for contradiction is zero: one contradicted fact and that summary fails.
  4. Assemble with receipts. The final dossier cites a source span for every claim and marks what the document doesn’t say.

The pipeline runs in seven stages, at temperature zero, with no agent autonomy. The control flow lives in code, not in a model’s judgment.

The result

A summary you can audit line by line, or an honest halt. Its sister project, CASDAM, applies the same closed-loop pattern to operating data.