Show your work. Down to the filing.
Every figure in a memo, a letter or a questionnaire response traces to the document it came from — and every step the AI took is on an append-only record, with a person's name against anything that could not be undone.
Point-in-time fundamentals, the filing record back to 1993, and 13F and insider data — under one license.
get_company_fundamentals
Read 12 figures · 4 filings
Ran unattendedGREENverify_fact_lineage
Resolved every figure to its filing
Ran unattendedGREENsave_thesis
Wrote a private draft
Executed · logged · undo availableAMBERpublish_report
Would make the memo public
Held for a personRED
- Logged, with who approved it
- Every actionLogged, with who approved it
- Of every filing, delisted included
- Every vintageOf every filing, delisted included
- 13F and insider rows
- 78M+13F and insider rows
- The default, not a setting
- Human-on-the-loopThe default, not a setting
We did not make these up. People said them.
Roughly half the day is not deciding — it is explaining. To the committee, to the client, to the operational-due-diligence team. Every one of those conversations is a request for evidence.
The position has to be defensible, not just right
If you cannot explain why you made a trade, you should not have made it.
Ken Griffin's line is the standard, and evidence is what meets it. Every figure behind a position keeps the identifier that opens its filing, so the explanation is a document rather than a recollection.
You cannot audit a 40-tab model at 9:30am
Walk me through how you got there.
The numbers in a draft stay bound to their sources, so challenging one is opening it rather than rebuilding it. The committee reads the same receipt the analyst read.
Compliance is the real buying decision
My CCO will never approve this.
Then start there. Irreversible actions stage for a person by default, the record is append-only, and the due-diligence document is public — written for an ODD team rather than summarized for one.
A revision surfaces in the client call
That's not what they filed.
We keep every vintage since 1993, so a figure a later filing changed shows as a before-and-after with both documents linked — and it can reach you as an alert rather than as a question you cannot answer live.
The reasons people say no.
We already have a terminal and a research provider.
Most of our institutional users keep both. We are not a news or execution product. What we add is the record underneath — every vintage of every filing since 1993, with provenance, plus the governance layer over whatever AI your desk is already using.
Our LPs ask what happens if the AI is wrong.
The honest answer is that a model can still write a wrong sentence — what it cannot do here is invent a figure, because figures are returned by typed tools rather than generated as text, and an unsourced one has no identifier and stands out. The rest is oversight: reversible actions are undoable, irreversible ones wait for a person.
Is a small vendor a concentration risk?
Fair, and the mitigation is that your data is not trapped here. Institutional includes bulk Parquet export of everything you can read, so the archive remains usable if we are not in the picture. We would rather answer this than have you discover it during diligence.
What can you not do?
No news, no estimates, no consensus, no execution, no private-company data, and no score. Fundamentals, filings, insider and institutional ownership, and the provenance around them. If your question needs a sell-side estimate, this is the wrong product for that question.
The same conviction. With the evidence attached.
- The memo asserts a number.The memo carries the number's filing.
- “Do you use AI?” is an awkward question in a DDQ.It is a document you already publish.
- The AI acted; nobody knows precisely what it touched.The ledger names every figure, tool and approver.
- A restatement arrives during a client call.It arrives in the inbox, with both filings.
Numbers we are willing to show our working for.
70% of buy-side firms now use AI in the front office, up from roughly 10% a year earlier — but only about one in ten say it plays a primary or directing role.
Advisers Act Rule 204-2 requires records supporting an investment decision to be preserved and retrievable promptly. An audit trail is not a feature request; it is the books-and-records obligation.
Across the S&P 500, 497 of 500 companies have had at least one figure changed by a later filing — and 94.5% of those changes arrived inside a routine quarterly or annual report.
One token. Every surface.
A single Stripe-issued key unlocks all of it at your tier. There is no per-channel billing and nothing to integrate twice.
Workspace
Where the memo gets written. Figures arrive bound to their filings, anything irreversible waits for a person, and the whole chain exports as one document you can hand to a committee.
MCP Server
Your analysts keep the AI they already use; it reaches the same archive under the same governance, so the oversight does not depend on which chat window the work happened in.
Bulk Data API
Risk and reporting systems pull the same Parquet the research desk reads, so the number in the client letter and the number in the model have one origin.
Asked and answered.
What exactly does the audit trail record?
Every tool an agent called, every figure it touched with the identifier that resolves to the filing, every action staged for approval with who approved it and when, and the output. It is append-only. The export — the Auditable Research File — assembles that chain into one document; it cannot flatter the run, because it only reports what the ledger already holds, and it flags an empty ledger rather than rendering a clean-looking page.
Can the AI act on its own?
Only where the action is reversible, and every such action is logged and undoable. Anything irreversible — publishing, signing off a report — is staged and waits for a person. That posture is the default in code, not a configuration option, and an institution that contractually requires pre-approval on everything can have that instead.
We have to answer AI questions in our DDQ. Can you help?
There is a public compliance and due-diligence document covering data lineage, which models are invoked, retention, and how to produce records on request. It is written to be handed to an operational-due-diligence team rather than summarized for them.
Does our data or our research leave our control?
You can run on your own model key, in which case your prompts go to your provider under your agreement and never become training data for anyone. On our managed lane the key is sealed and zero-retention is available per connection. Either way we do not train on customer content.
What does Institutional include that Pro does not?
The full universe back to 1993 including foreign issuers, the 13F and insider tables, intraday filing acceptance timestamps, filing-event webhooks, the redistribution grant and an SLA. Pro is a single-analyst license on a 15-year rolling window.
Do you give a risk score or a recommendation?
No, deliberately. We show you what changed, when, and what document it came from. Turning that into a view is the job you are paid for, and a vendor score would be a conclusion you would have to defend without owning the reasoning behind it.
The fastest way to evaluate this is to bring a position you already have to defend, and see whether the evidence holds up.
Recommended plan for this desk: Institutional.