Source-backed state, not loose summary.
The method is simple: preserve sources, derive typed state, expose gaps and conflicts, then version the current view as new deal material arrives.
- Source intake
Normalise files, emails, notes, CRM records and generated outputs into a source ledger.
- Entity and metric extraction
Extract companies, people, dates, numbers, claims and obligations into typed state.
- Claim formation
Represent each material claim as structured data, not as a free-form paragraph.
- Source attachment
Attach every fact to a citation, or mark an inference or unsupported claim explicitly.
- Conflict detection
Keep conflicting values visible until a human or source-backed rule resolves them.
- Missing material tracking
Treat absent artefacts as part of the state, not as private analyst memory.
- Versioned updates
Record what changed since the last review so a deal team can catch up quickly.
- Human review boundary
Keep the system useful without pretending synthetic outputs are final judgement.
DealState does not treat summaries as truth.
Every material claim, number and issue should trace back to a source, confidence level and review state. Unsupported claims remain visible as unsupported. Conflicting sources remain visible as conflicts.
The state register is the durable object.
Chat can answer questions over the state, but the operating surface is where the team inspects what is known, questioned, missing and changed.
Scores are directional, not decorative.
Nine firm dimensions produce a rounded weighted mean. Higher is more favourable. Risk dimensions therefore score mitigation, not exposure.
Pass 1 is synthetic by design.
There is no real ingestion, retrieval or model call in this demo. The point is to show the product shape and the integrity spine before wiring live connectors.
Source coverage
received required artefacts + sourced key metrics divided by all required artefacts + all key metrics. Project Nova computes to 46%.