Private-market early warning, organized around the investment decision.
Most of the private-capital stack — CRM, VDR, monitoring, accounting, AI assistants — is organized around objects other than the investment decision. Capital Refinery makes the decision the first-class record: evidence at approval, pressure detection after close, and a permanent trail of what changed, who acted, and why.
Each tool is excellent at one slice. None of them holds the position.
The category is not a CRM. It is not a dashboard. It is not a chat window. It is the layer underneath all of them — the structural record of the decision, alive across time.
- 01Market data & screeningPitchBook · S&P Capital IQ · Preqin · CoStar
- 02Deal CRM & relationshipDealCloud · Affinity · Altvia · Dealpath
- 03Diligence / VDRDatasite · Intralinks · Hebbia · BlueFlame
- 04Generative AI for PEHebbia · BlueFlame · AlphaSense · Endex · Eilla
- 05Portfolio monitoringiLEVEL · Allvue · Chronograph · Cobalt LP
- 06Accounting & operationsFundCount · eFront · Standard Metrics · QBO
- 07Decision integrityMissing layerCapital Refinery — early warning + decision record
A note on terms: we use “operating system” internally for this architecture. Externally the claim is decision integrity and private-market early warning — not a category coronation.
Three properties no incumbent ships.
The investment decision as the first-class object
Not the document. Not the dashboard. The decision.
The platform schema has a node for the IC decision — the thesis as a structured field, the assumptions as separate testable rows, the conditions as logical clauses with quantitative thresholds. From IC approval forward, that node is the reference point for every operator update. This is the architectural choice the incumbents did not make, because their schemas were built around different primary objects.
Deterministic-first extraction with provenance
The brain is deterministic. The LLM is the adjudicator.
Generative-first stacks cannot guarantee provenance. Capital Refinery uses a regex-and-constructor sweep to produce candidate figures from CIMs, financials, credit agreements, and operator updates. A small local LLM adjudicates conflicts and produces summaries. Every figure has a candidate ID, a method, and a source-page provenance trail. The numbers are reproducible — the LLM is not asked to do the part of the work it cannot do reliably.
A continuous re-test loop across the life of the position
The decision stays alive after IC.
Each accounting feed, KPI report, and operator narrative is mapped to the entry assumption it tests. The connection is structural, not generated. When conditions change enough that the original IC decision is no longer defensible, the platform surfaces it — with the assumption that broke and the time-to-consequence ranking against the rest of the portfolio. Decision validity stops being a memory exercise and starts being a property the platform can compute.
Where the lines fall.
| Capability | The existing stack | Capital Refinery |
|---|---|---|
| Primary object in the data model | Document, dashboard, or relationship | The IC decision |
| Structured investment record at IC | No | Yes — bound at approval |
| Deterministic extraction with provenance | No — LLM-first or manual | Yes — every figure has a candidate ID |
| Operator data → entry assumption mapping | No | Yes — the data model |
| Decision validity across the life of the position | No | Yes |
| Covenant + breach probability for credit positions | No | Yes |
| Time-to-consequence ranking | No | Yes |
| Survives team turnover as institutional memory | No — relies on analyst memory | Yes — the structured record is the memory |
The view that shows the entire book sorted by time-to-consequence. The positions running out of room are at the top. The IC review reorders around runway, not colour code.
The research behind the category.
- The Decision Integrity GapThe flagship research on why the existing stack does not produce a system of record for the investment decision — and what would have to be true for one to exist.
- Your Data Room Dies at CloseWhy the most analytically dense work the firm does on an asset becomes inaccessible the moment it matters most.
- Why Your Monitoring Dashboard Isn't a Decision ToolThe structural distinction between knowing and deciding, applied to the post-close stack.
- Time-to-Consequence as the Missing MetricWhy ranking the portfolio by time-to-consequence — not by IRR — is the operating model the IC needs.
The decision-integrity layer your stack was never built to produce.
Bring us a deal from your portfolio. We will parse the entry pack into a structured investment record, bind the live operator data to the assumptions, and show you the operating system the existing stack is not designed to produce.