Evidence first · AI last

Diligence you can defend, line by line.

Every finding traces to a highlighted span on a source page, judged against a versioned industry baseline pinned before the work started. Nothing is asserted that cannot be shown.

Sign inHow it works
Deterministic
Parsers, rules and lookups produce the facts. Models never do arithmetic.
Separated
Evidence and industry baseline are two corpora that never mix.
Traceable
Page, span and baseline version recorded on every claim.
Reviewed
A person signs off. Nothing publishes itself.
The state of the craft

Due diligence is still a memory-bound manual trade.

A team reads thousands of documents, works through a checklist inherited from a partner's spreadsheet, and writes a memo under deadline. Four structural failures follow — and the real cost is not the fee. It is the missed finding discovered after close.

3–8 weeks
Slow

Typical transaction diligence. Vendor assessments queue for months behind it.

$50k–500k
Expensive

Per transaction in fees. Enterprises run thousands of vendor reviews a year at $2k–15k each.

Two analysts
Inconsistent

The same data room, materially different findings. No reproducibility, no measurable quality bar.

A few heads
Knowledge-bound

“What good looks like here” walks out at retirement, and does not scale to a new sector.

The principle

Deterministic first. AI last.

Anything a parser, a lookup, a calculation or a rule can do reliably is not given to a language model. The model's job is judgement, synthesis and narrative — the part that genuinely requires reasoning. Facts come from components that can be tested to 100%; only judgement is probabilistic.

  1. 01ScopeTarget, industry, jurisdictions, materiality threshold. The matching industry baseline is pinned at this moment and never silently moves.
  2. 02IngestData room, uploads, questionnaire responses. Every file hashed and written to an append-only ledger.
  3. 03ExtractOCR, classification, schema-constrained extraction. Every value carries its page and span.
  4. 04CheckCompleteness against the expected manifest, rules, benchmarks, contradictions — all deterministic.
  5. 05JudgeOnly now does a model read the structured facts against the baseline, to weigh materiality and draft the narrative.
  6. 06ReviewEvery finding accepted, edited or rejected by a person. Each edit becomes improvement signal.

Step 05 is the only step a model touches — and it reads structured facts, never raw documents.

The control that cannot be retrofitted

Two corpora that never mix.

What is true about this target, and what should be true in this industry, are different kinds of statement. They are stored apart, retrieved apart, and presented to the model in separately-labelled blocks.

Evidence corpus

What is true about this target

The target's own documents and the facts drawn from them, scoped to one engagement. It never defines a standard.

Knowledge corpus

What should be true in this industry

Practices, obligations, benchmarks and known failure patterns — versioned and cited. It never asserts anything about this target.

Accuracy control
A best-practice statement cannot be mistaken for a fact about the target.
Security control
Text injected into a target document cannot redefine the standard applied to it.
Compliance control
The report can always show which standard was applied, at which version.
Not all evidence is equal

Every document carries a tier.

The system encodes the weight of a source rather than leaving it to the reader. A finding resting only on tier 4–5 material cannot be rated material without an explicit reviewer override — and the report says so.

TierSourceWeight
1Audited financials, regulatory filings, court and registry recordsHighest
2Executed contracts, certificates, licences, insurance policiesHigh
3Management accounts, board minutes, internal policiesMedium
4Management presentations, forecasts, self-assessment questionnairesLow — corroboration required
5Verbal statements, undated drafts, unsigned documentsLowest — flagged unverified
Where it is used

One method, seven shapes of engagement.

01
Third-party & vendor risk

Many, shallow, recurring. Thousands a year — automation ratio matters more than depth.

02
M&A transaction diligence

Few, deep, deadline-driven. 500–5,000 documents against the clock.

03
Vendor self-assessment

Find the issues before the buyer does.

04
Investment diligence

A screening tier, then a full tier once it clears.

05
Compliance readiness

Gap assessment against a named standard.

06
Internal audit

Recurring self-review against industry practice.

07
Post-close monitoring

Re-run on a schedule, or on a trigger event.

Deliberate limits

What this will not do.

A diligence tool that claims everything is a tool you cannot rely on for anything. These exclusions are design decisions, not gaps in the roadmap.

Legal, tax or audit opinionsRegulated activity. It would change the liability and licensing posture entirely.
Automated go / no-go decisionsRemoves the human oversight that is both the accuracy control and the compliance control.
Valuation and financial modellingA different product for a different buyer.
Agents that contact the targetUnacceptable blast radius given prompt-injection exposure.

Start with one engagement.

Create it, drop in the data room, and watch the gap list build itself before a single model is called.

Sign inArchNova platform