diligence-reader

Reads an acquisition data room and writes a findings report a deal team can act on: the matter that connects across the documents under different names, weighed against what the seller disclosed and reserved, a recommendation with a number, and every sentence cited to the page it came from and checked against it by code.

Project page: muhanad-husn.github.io/diligence-reader

It is a fixed rebuild of the recursive language model (RLM) run that John Adeojo published on his Project Atlas data room (brainqub3/claude_code_RLM, brainqub3/synthetic-dataRoom). The RLM improvises its program per run; this tool keeps the shape of his winning run and replaces every improvised step with one that is the same on every run. Two of seven stages call a model. The rest is code that is byte-identical across runs.

100 / 100	Rubric score on Project Atlas, two runs, spread 0
82 / 84	The RLM skill as shipped, same corpus, same grader
about $0.30	One end-to-end run of the 100-document room
$18.95	The whole build, 5 to 9 September 2026, every dollar in LEDGER.md

Full results, method and what still stands against it: REPORT.md.

How it works

Stage	Kind	What it does
Ingest	code	PDF, XLSX, CSV, TXT, EML, MBOX and markdown into sections with anchors, plus an index of dates, amounts, names, identifiers, status words, version pairs and series
Notes	model, one call per document	A note under a fixed JSON schema. Every quote and figure is checked against the source by code; a failed note is re-asked once, then dropped and logged
Map	code	A document graph over shared identifiers, cross-references, dates and versions. Clusters the matter; consequence links catch a series that breaks at the matter's date and a model dated after it
Dossier	code	Per matter: timeline, every name each function gave it, figures with sources, draft against final, reserve against estimate, deadlines against actions, models blind to it, lesser matters ranked by money
Write	model, one call	The five-section report from the dossier alone. Every sentence cites; certainty words are copied from the source and never raised
Verify	code	Every citation resolves, every number exists in its source, no certainty word climbed, decoys ranked below the main matter. Failures go back to the writer once
Grade	code + subagent	Planted-fact recall by code against the sample's key; rubric score by a Claude subagent; spread between two runs

The stage modules live in src/rlm/, one per stage. The package keeps the name rlm after the run it rebuilds.

Samples

Four corpora ship with the repository or fetch from a public source. Each has a machine-readable key; nothing is "found" unless a test reads it out of an artefact against that key.

Sample	What it is	Result
atlas	Project Atlas: 100 documents, six formats, one planted matter across six functions, five decoys, a 100-point rubric	recall 100, rubric 100, spread 0
northwind	Eleven contracts and a board deck; a change-of-control cliff on a 30.1% customer, a decoy contract on the same template	recall 100
northstar-dental	A numeric contradiction: the CIM claims 18.0% growth, the workbook shows 11.7%	recall 100
yahoo	Real SEC filings, Verizon and Yahoo 2016 to 2017, no planted facts; key from the public record. Fetched at run time	recall 86.4

Five generated variants of atlas (misspelled names, an unnamed matter, a second matter, a doubled room, a control) are described in REPORT.md section 3 and PLAN.md section 10. Details of every sample and its key: samples/README.md.

Run it

Three ways to run it.

Docker

The web page on the user's machine.

docker run -p 8000:8000 -v <runs>:/app/runs ghcr.io/muhanad-husn/diligence-reader

Open http://localhost:8000, connect an OpenRouter key (sign in or paste one), pick the model of each task under Models, upload a room as a folder or zip, read the estimate, confirm, watch progress, read the report and export Word, PDF or the evidence CSV. Stop ends a run in progress and Retry resumes it from the stage that stopped; uploading a new room while a run is going asks first. Past runs lists every run on the machine with its status and dollars, and New room on the report clears the page back to the upload and keeps the key. The key stays in the browser and goes with each run; it is not written to disk. The models are saved on the machine, in the runs folder under Docker.

pipx

The command alone.

pipx install diligence-reader
diligence-reader run <room> --out <folder>

Set OPENROUTER_API_KEY. The command shows the price estimate and asks before the first model call; pass --yes to accept it. A rerun resumes from the stage that stopped.

Helm

For a firm that runs it on its own Kubernetes cluster.

helm install dr oci://ghcr.io/muhanad-husn/charts/diligence-reader
kubectl port-forward svc/dr-diligence-reader 8000:8000

Each run is one Kubernetes Job. The page is at http://localhost:8000.

From the repository

Python 3.13. Model calls go through OpenRouter; the tool refuses to call a model before it has counted the input tokens, printed the price and checked the ledger against the ceiling. A user's own runs spend the user's OpenRouter credits; the limit is the one the user sets on the key.

pip install -e ".[dev]"
export OPENROUTER_API_KEY=...

Run one stage at a time: python -m rlm.ingest, python -m rlm.notes, python -m rlm.map, python -m rlm.dossier, python -m rlm.write. Tests: pytest -q. Artefacts land under runs/ and are never committed.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗