What is actually built
Every number on this page is counted from the product source.
Specialist sub-agents
A module is a team, not a prompt. The largest fields 21.
Modules
Each one pre-scoped to a discipline or a deliverable.
Dispatch waves
Ordered by dependency, not by preference. Wave 1 runs eleven modules at once.
Zero-retention gated
Tax, legal, financial and boedelonderzoek refuse to run off a retaining model.
Hard-block conditions
Conditions that stop a draft from being auto-approved, however many retries it has had.
Disciplines per engagement
Financial, commercial, legal, tax, HR, IT, ESG, vendor, operational, valuation, insurance.
All 31 modules, grouped by dispatch wave
A module that needs another module’s output cannot start before it. That dependency is the running order. The spread is the other half of the picture: the largest module fields 21 sub-agents, the smallest two.
Independent reads
113- Financial DDZDR11
- Commercial DD10
- HR DD10
- IT DD7
- ESG DD11
- Vigil monitoring21
- BoedelonderzoekZDR11
- Operational DD11
- Vendor DD8
- AI DD9
- IM screener4
Dependent reads
26- Tax DDZDR10
- Legal DDZDR10
- Insurance DD6
Synthesis
63- Independent Business Review11
- WHOA restructuring14
- Deal economics9
- Valuation6
- Deal structuring4
- Portfolio management6
- LBO model7
- Post-merger integration6
Deliverables
34- Vendor due diligence13
- IC memo5
- Teaser / CIM5
- Financing memorandum6
- Document factory5
Post-close
18- Exit readiness8
- Portfolio health2
- Monthly business review3
- IC report5
Counted from the dispatch graphs in the product source, not estimated. ZDR marks the modules whose provider routing is hard-gated to a zero-retention EU endpoint.
From upload to a finding you can act on
From upload to drafted findings there is no human in the loop. One mandatory human gate sits after it.
- 01
Grounded retrieval
The module dispatches its sub-agents. Each retrieves evidence from your documents.
Confidential financial and personal data is pseudonymized before it reaches this layer.
- 02
Drafting against the schema
A claim without a document and a passage behind it is not a finding.
- 03
Generate-critique-retry
A reviewer agent critiques the draft and sends it back. Up to two retries per specialist.
- 04
Four-layer grounding check
Index, quote, entailment, second opinion. Each layer catches what the one before it missed.
- 05
Cross-discipline reconciliation
Findings are read against each other and resolved into one narrative.
- 06
Human review and approval
A named person reviews the findings before release.
Pseudonymization is reversed only here, in the final delivered output.
Four checks, each narrower than the last
A citation is not evidence until something has read it. Four layers do, in order.
- 01
Citation calibration
Every citation index is checked against the set of passages that were actually retrieved.
CatchesCitations pointing at a document the run never opened. - 02
Lexical quote check
The quoted text has to appear in the cited passage, character for character.
CatchesParaphrase presented as a verbatim excerpt. - 03
Entailment judge
A separate model reads the claim and the passage and rules on whether one follows from the other.
CatchesReal quotes that do not support the claim built on them. - 04
Second-opinion verify
A local natural-language-inference model re-runs the judgment. This layer fails closed.
CatchesWhatever the first three let through.
Claims that fail get one repair attempt: re-quote from source, then re-judge. Claims that still fail are cut or marked unconfirmed.
7 conditions that stop auto-approval
Retries are capped at two per specialist. These conditions end the loop regardless and route the draft to a person.
Privacy leak
Data crossed a boundary it should not have.
Fabricated source
A citation that does not resolve to a real passage.
Scope refusal
The module cannot answer inside its brief, and says so.
Insufficient depth
An answer that does not meet the standard the schema demands.
Placeholder text
Filler left where evidence belongs.
Near-empty draft
Too little substance to review.
Low reviewer score
The critic agent scored the draft below threshold.
Every finding points back at a page
The trace is the product. A finding you cannot walk back to a passage is an opinion.
- Module
- Pre-scoped, one of the library
- Finding
- Withheld
- Verbatim excerpt
- Withheld
- Source document and passage
- Withheld
- Human review
- Required before release
Redacted because client documents are confidential, not because the trace is missing. Every bar is a line of a real page; the highlighted one is what the citation points at.
Three other ways to do this
Compared as categories of tool, not as named products. Marks answer one question each: does this category of tool do this thing at all.
| Virtual data room | AI diligence tools | General-purpose AI | Factum | |
|---|---|---|---|---|
| Pre-scoped module per discipline | No | Partly | No | Yes |
| Runs without being prompted | No | Partly | No | Yes |
| Citation required by the schema | Not applicable | Partly | No | Yes |
| Cross-discipline reconciliation | No | No | No | Yes |
| Produces the deliverable, not an answer | No | Partly | No | Yes |
| Zero retention at the model provider | Not applicable | Partly | No | Yes |
| Named person signs off | Not applicable | No | No | Yes |
| Published accuracy auditNobody in this table has one, including us. | No | No | No | No |
- Yes
- Partly
- No
- Not applicable
Categories, not vendors. Named products change quarterly and most public claims about them are vendor-published.
What we don’t claim about the platform
Everything above describes a mechanism. None of it is a certified metric.
- —No audited accuracy rate. Internal benchmarks are single-dataset and unaudited.
- —No claim that the human gate is optional. It is mandatory.
- —No completed public case studies yet.
- —No uniform track record. The financial module has run end to end on a live deal and been verified against source documents. The reconciliation layer is younger.
- —No general availability. Factum is pre-launch.
Worth answering before you ask
Isn’t this just an AI chatbot with extra steps?
A chat session is one context window that you steer. This is 254 sub-agents across 31 pre-scoped modules, dispatched in a fixed dependency order, with four grounding checks and a mandatory human gate. The difference is what happens when nobody is watching.
Why waves instead of running everything at once?
Tax cannot reason without the financial and legal reads. Valuation cannot price without tax. The dependency graph sets the order; wave 1 still runs eleven modules in parallel.
What happens when the model doesn’t know?
It has to say so. A finding is either backed by a citation that survived four checks or flagged as unconfirmed. Fabricated sources are one of the 7 hard blocks.
How do I know the platform can carry findings this material?
The financial module is the most numerically exposed part of any read. It has run end to end on a live deal and been independently verified against source documents. Newer layers have less history, and we tell you which is which.
How fast is it?
The pipeline pass runs in hours. The finished, reviewed output takes days. We don’t quote a delivery date before seeing the data room.