Private Markets Software Has a Decision Integrity Problem

A $13 trillion industry has spent fifteen years optimizing the wrong layer of the workflow. The result is a stack that knows everything about today and nothing about whether yesterday's decision still works.

The premise

Every private markets investment is a structured argument. The deal team builds a thesis. They identify the conditions that thesis depends on. They quantify what has to be true for the entry price to make sense. The IC approves it, and the wire goes out.

What happens next is where the industry has a structural failure. The thesis becomes a slide deck on a shared drive. The conditions become a folder someone might re-read in year three. The assumptions become the team's institutional memory — held in the head of one principal who is now staffed on a new transaction.

And meanwhile, every quarter, the operator sends in updated KPIs. The accounting feed fires. The credit agreement covenants get tested. The model gets dusted off. None of it is connected to the original investment thesis in any structural way. By the time anyone notices the thesis is broken, the breach has already happened.

$3.6 trillion in unrealized value sits on GP balance sheets. The pressure on decision quality has never been higher.

The 6-layer stack

The dominant private markets software stack today has six layers. Each one is mature, well-funded, and excellent at its job. None of them holds the investment decision as a structured object the platform can re-test continuously against operator data.

Layer 1 — Market data

PitchBook, S&P Capital IQ, Preqin. Sourcing, comparable transactions, fund benchmarks. The outside view of any deal. Every team uses one. None of them is designed to maintain the inside view of a position you actually closed.

Layer 2 — Deal CRM

DealCloud, Affinity, Altvia. Pipeline, relationships, process workflow. Excellent at the workflow around a deal — the relationships, the stages, the IC scheduling. The IC memo is an attachment, not a data model. The deal closes and the structured representation of the decision ends with it.

Layer 3 — Diligence / VDR

Datasite, Intralinks, Hebbia. Hosting, securing, and increasingly synthesizing the contents of the diligence room. Mature inside the diligence window. Disconnects the moment the deal closes — the data room is a transactional artifact, not a living connection to the live position.

Layer 4 — Portfolio monitoring

iLEVEL (Solovis/SS&C — used by 700+ firms), Allvue, Chronograph, Cobalt LP. KPI ingestion, dashboard rendering, LP reporting. The dominant infrastructure for the post-close window. What none of them does is bind the live KPI to the IC assumption it tests.

Layer 5 — Accounting / fund admin

FundCount, eFront, AltExchange. The books, the waterfall, the LP statements. Mature, audited, regulated. Operates entirely below the decision layer. The numbers it produces feed the dashboards but never reach the original thesis.

Layer 6 — Generative AI overlays

The 2024–2025 wave: Hebbia, BlueFlame, AlphaSense, Eilla. Chat over your documents. Useful for research and diligence acceleration. Not designed to maintain a structured record of the IC decision or to test it continuously against the live operator data. Generative speed without a decision substrate.

The missing 7th layer: decision integrity

A decision integrity layer does three things the existing six do not. First, it captures the IC decision as a structured object — thesis, assumptions, and the conditions those assumptions depend on. Second, it binds every operator datum to the assumption it tests. Third, it scores decision validity continuously and surfaces the moment a prior decision stops being defensible.

None of those three is a UX layer over an existing tool. They require a particular extraction model (deterministic-first), a particular data model (the assumption-to-datum binding), and a particular feedback loop (closed, continuous, with time-to-consequence ranking). The reason no existing vendor has shipped this layer is that none of them was built around the investment decision as the unit of analysis.

What changes when the layer exists

• →The IC memo is not a slide deck headed for a shared drive. It is a structured object the platform tests against — every quarter, every operator update, every covenant report.
• →Monitoring stops being a visibility exercise and becomes a decision diagnostic. The KPI is interesting only because of what it tests about the entry assumption.
• →Time-to-consequence becomes a measurable metric. Teams sort the book by how much room is left before a structural break — not by quarter colour.
• →Institutional memory survives team turnover. The next principal inherits the structured reasoning, not a folder of memos written for a committee that no longer exists.
• →LPs see decision discipline as data. The same structured record can be exported into LP reporting without rebuilding it from scratch every quarter.

Why this category did not exist a year ago

Three structural shifts made it possible. The first is small, capable local LLMs that can run as a thin adjudication layer over deterministic extraction — without exposing client data to a third-party API. The second is the industry's growing acceptance that quarterly reconstruction is the wrong unit of analysis. The third is LP scrutiny: as exit windows extend and DPI becomes the metric of record, the question of whether prior decisions still hold has moved from "interesting" to "load-bearing."

Capital Refinery is the layer that closes the gap. It is not a replacement for any of the existing six — it is what they were never designed to be.