diligence-reader

Reads the whole acquisition data room. Writes the findings report, every sentence cited and checked against its source.

A recommendation with a number, findings ranked by money at stake, the material matter quantified with a deal action, the lesser issues, the open items.

01What you get

The room in, the report out.

You hand over the room as the seller left it, in whatever formats: PDF, spreadsheets, email, exports.

You get back a five-section findings report. Every sentence carries a citation to the page it came from. Every citation, figure and certainty word has been checked against that page by code before the report reaches you.

Model cost for a hundred-document room is about thirty cents. The tool gives the same answer if you run it twice. What it does not do is replace your judgement; it reads the whole room so your people start from a map instead of a pile, and it refuses to hand you a sentence it could not verify.

Seven stages. Two call a model.

1Ingestcode

Files in seven formats into sections with anchors, plus an index of dates, amounts, names, identifiers, status words, versions.

2Notesmodel, one call per document

A note under a fixed schema. Every quote and figure is checked against the source by code.

3Mapcode

A document graph over shared identifiers, cross-references, dates and versions. The matter shows up as a cluster.

4Dossiercode

Per matter: the timeline, every name each function gave it, figures with sources, draft against final, reserve against estimate.

5Writemodel, one call

The five-section report from the dossier alone. Every sentence cites. Certainty words are copied, never raised.

6Verifycode

Every citation resolves, every number exists in its source, no certainty word climbed, decoys ranked below the matter. Failures go back to the writer once.

7Gradecode plus a subagent

Recall against the key, rubric score, spread between two runs.

Four corpora, one key each.

Sample	What it tests	Result
atlas	100 documents, one matter across six functions, five decoys, a rubric	recall 100, rubric 100, spread 0
northwind	Eleven contracts, a change-of-control cliff on a 30.1% customer, a decoy on the same template	recall 100
northstar-dental	A numeric contradiction: 18.0% claimed, 11.7% in the workbook	recall 100
yahoo	Real SEC filings, Verizon and Yahoo 2016 to 2017, no planted facts	recall 86.4

Five generated variants of atlas (misspelled names, an unnamed matter, a second matter, a doubled room, a control) score 94 to 100.

Five commands. The price prints first.

git clone https://github.com/Muhanad-husn/diligence-reader.git cd diligence-reader pip install -e ".[dev]" export OPENROUTER_API_KEY=... python -m rlm.ingest samples/atlas runs/atlas python -m rlm.notes samples/atlas runs/atlas python -m rlm.map samples/atlas runs/atlas python -m rlm.dossier samples/atlas runs/atlas python -m rlm.write samples/atlas runs/atlas --passes 2

The tool counts the input tokens and prints the price before any model call, and refuses past the ceiling. Every dollar is in LEDGER.md.

The rules the build kept.

The key is the only test.

A phase passes when its artefact carries every planted fact on all three gate samples, with no code change between them.

One artefact per phase, in order.

A fact dropped later is traced to the phase that dropped it and fixed there. Every later phase reruns.

Money is enforced by code.

LEDGER.md is written by the code that makes the call. Past a phase cap or the $50 total, the call refuses.

Cheapest model that passes.

Five candidates, a bake-off per model-calling phase, no preference, only the table.

Built against a published run.

It is a fixed rebuild of a recursive language model run that John Adeojo published on his Project Atlas data room. His RLM improvises its program on every run. This tool keeps the shape of his winning run and replaces every improvised step with one that is the same every time. Seven stages: two call a model, the other five are code that produces the same bytes on every run.

The task, the room, the key, the rubric and the winning shape are John Adeojo's (brainqub3). The rebuild would not exist without a published run to measure against.