ReActive

Method

ReActive rebuilds a portfolio's structure from its own documents, independent of the reported figures, and tests whether that structure can all be true at once. Where reported and reconstructed values differ, it quantifies the delta and shows the documents behind it.

How a scan runs.

VDR, as delivered

  1. 01

    Ingestion

    Ingestion pipeline

    Documents classified and normalised. The raw data room enters as delivered.

  2. 02

    Extraction

    LLM-assisted

    Facts extracted under a fixed schema, with the source of every value recorded. Divergent values are preserved, not reconciled.

  3. 03

    Structure reconstruction

    Deterministic

    Entity, debt, temporal and income layers rebuilt from extracted facts in pure code, independent of the reported figures.

  4. 04

    Coherence testing

    Deterministic

    Deterministic tests check whether the reconstructed structures can all be true at once.

  5. 05

    Delta computation

    Deterministic

    Reconstructed values compared to reported ones, expressed as a quantified delta with its documentary derivation.

  6. 06

    Findings

    Deterministic

    Each finding carries a complete evidence chain to specific clauses in specific documents.

The AI boundary

Stage 02 only

LLM-assisted at extraction, deterministic thereafter. Every other stage is executable rule code: typed facts read from a fixed schema and evaluated by versioned rules in a fixed execution order, so identical inputs produce identical outputs.

Three levels, both directions

Consolidated, entity, clause

From a portfolio-level finding down to the clause that evidences it, and from a clause back up to everything that depends on it.

A finding that cannot be reproduced cannot be defended.

A language model asked the same question twice can answer differently. That is disqualifying where a finding has to be shown to an investment committee, a lender or a counterparty.

So language models are confined to extraction. Every check, consolidation and finding downstream is deterministic logic, and extracted facts are persisted, so a completed scan reproduces regardless of later model changes. Same corpus, same findings, every time.

Reproducible is not the same as checkable, so each finding carries its evidence chain: every fact leads back to the document and clause it was read from, quoted verbatim, and every computed figure leads back to the facts behind it.

A scan saves time. What matters more is what it finds.

When two documents cannot both be true, ReActive shows you where.

See a scan replayed.

The demo replays a recorded scan over a synthetic dataroom, from the documents to the findings.

Open the demo