How BearCase handles your documents.

Who can see your deal

A deal belongs to the account that created it, and every request for anything inside it, a document, a claim, a piece of evidence, a scenario, a chat thread, is checked against that owner first. Another account asking for your deal gets “not found”, not “forbidden”: the server does not confirm the deal exists. The owner can invite people by email as viewers or editors, and remove them; nobody else sees the deal, and there is no admin view. Never share your own sign-in.

The demo works the same way. Each visitor to /demo gets a private demo identity and their own copy of the fictional Northstar deal. Anything you upload, decide, or ask there is visible to your browser session only; the next visitor gets a fresh copy and cannot open yours. A signed-in account keeps its own identity when it opens the demo.

What leaves the server

Reading your files, checking the numbers, running the scenarios, and building the report all happen on the server with code and a rule-based reader. None of that sends anything anywhere.

The assistant is the exception, and only when a model is connected. The assistant provider you configure (Anthropic, OpenAI, Google Gemini, Groq, OpenRouter, a local Ollama, or your own server) receives what the chat tools return for your question: claim text, short evidence snippets, and the deal brief, a one-page summary of the deal's figures. It does not receive your files. If no key is configured, nothing is sent to anyone: the same chat runs on the offline rule-based composer, and the panel says so. If you switch the document reader itself from the rule-based reader to a model, the text of your documents goes to that provider too.

Read the provider's terms before you connect it. On Google's free tier, inputs may be used to improve Google's products, and every visitor to a deployment shares the operator's one key and its limits. Use a paid plan, or no model at all, for a confidential deal. The panel always names the provider that answered, and the audit history keeps every reply.

How long we keep documents

Documents in your own account are kept until you delete them. There is no automatic purge, and no backup either: this is a prototype, and the server that holds your files is the only copy.

Demo deals are temporary. A visitor's demo identity, deals, and files are deleted 14 days after their newest session expired, a few at a time when a new demo starts. Sign-in sessions last 14 days as well. Each account is capped at 50 documents per deal and 500 MB in total, and a single upload at 25 MB.

Deleting a document

Delete a document from the Deal Room and the server removes the file and every stored version of it, the evidence that was read from it, and the claims that came from it, together with their evidence links and any reviewer decisions on them. Citations that pointed at that document no longer open.

What the delete does not do is recompute. Findings, metrics, scenarios, the report, and the questions for the seller were built from the documents present at the time, so after a delete they may describe evidence that is gone. The Deal Room asks you to run analysis again, and keeps asking until you do. The audit history records that a document was deleted, with its name and time but not its contents.

What the labels mean

Every statement BearCase shows you is one of four kinds, and the mark beside it says which. Two of them carry a citation you can open; two do not. Keeping them apart is most of what makes the review checkable.

ESource fact

A page, a sheet row, or a CSV row in a document the seller gave you. The chip opens the document at that spot. The passage is quoted as it was written, so it can still be wrong; a source fact tells you where to look, not that the seller is right.

MCalculation

A number BearCase's engine computed from source facts with a stored formula and input snapshot. The model never produces one. It is only as good as its inputs: a calculation built on a mis-read cell is wrong in the same way.

AAssumption

A value a person chose rather than a document stated: the purchase price and loan terms you entered, a scenario input, a growth rate. Assumptions carry no citation. Change one and every calculation downstream changes with it.

AIAI interpretation

Prose the model drafted: why a claim got its status, what a finding means, an answer in the chat. It may cite source facts and calculations, and a citation that resolves shows where a figure came from, not that the sentence around it is right. Read the sources before you rely on it.

What a citation proves

A citation that resolves proves provenance: the statement came from this page and paragraph, this sheet and row, or this CSV row, in this version of this document. Code checks every citation before anything is saved, and a report fails if a material statement has none.

A citation does not prove the statement is correct. The seller's memo can be cited perfectly and still be wrong; a calculation can cite its inputs and still rest on a mis-read cell; the assistant can cite a real passage and misread it. What carries the judgement is the status rule that compared the claim with the evidence, the number recomputed from the statements, and you, opening the source and reading it. A citation is the shortest path to that source, not a substitute for it.

Limits

• The default reader is rule-based. It finds the claims it has rules for and can miss or misread others; a model can be connected for extraction, but every status is still meant to be reviewed by a person.
• The evaluation suite scores the curated Northstar fixtures. Passing it shows the pipeline is consistent on documents it was built with, not that it is accurate on yours. Reviewed claims can be exported as a dataset so that a release can be measured against real corrections.
• Uploads are checked by their content and never executed. Text inside a document that reads like an instruction is stored as inert content, labelled, and reported as a finding. Document text reaches a model only as quoted data, never as instructions, and no financial number comes from a model. How well a live model resists instruction-like text has not been measured here.
• A deployment for confidential deals would still need encryption at rest, backups, monitoring, an external penetration test, and a paid model plan. None of those exist here yet.
• BearCase is a first-pass check of a seller’s documents. It does not replace an accountant’s quality-of-earnings work or legal diligence, and it does not provide financial, legal, tax, or investment advice. The Northstar deal and everyone in it are fictional.

How the checks themselves work, status by status and formula by formula, is on How we check claims.