Evaluation Framework
Evaluating AI Due Diligence Software for Private Capital
// THE 2026 STRUCTURAL FRAMEWORK
A structural guide for risk officers and general partners: how to evaluate diligence software across document custody, probabilistic text generation, and deterministic judgment infrastructure.
// THE MATURITY MODEL
Three tiers. One fiduciary question.
Every private-market firm is now being sold "AI diligence." Most of the comparison happens at the feature level — speed, integrations, supported file types. That is the wrong altitude. The question a Risk Officer should ask is not how fast a tool reads a data room, but whether its output can be verified, reproduced, and defended.
This is a structural framework for evaluating due diligence software, not a feature checklist. Diligence software sorts into three tiers of maturity. The differences between them are architectural, and they determine whether you end up with a faster reader or a defensible verdict.
Storage, generation, judgment. Each tier solves a strictly harder problem than the one before it.
TIER 1 · STORAGE
Traditional Data Rooms
Secure, but manual and dumb.
The virtual data room category — Datasite, Intralinks — solved one problem with discipline: secure document custody and permissioned access for transaction parties. That is real infrastructure, and it is not going away. But a data room stores documents; it does not evaluate claims. It will hold a 90-page CIM with bank-grade encryption and tell you nothing about whether the unit economics inside it survive contact with arithmetic. The analyst still reads every page by hand.
TIER 2 · GENERATION
Probabilistic AI Wrappers
Fast, but unverifiable.
The current wave layers a probabilistic LLM over the data room — generic SaaS wrappers that summarize a deck in seconds and draft the first pass of a memo. The speed is real. The problem is the substrate: these systems generate a plausible opinion, not a reconstructible one. They hallucinate figures, cannot show their work, and produce a different answer on a different day. LLMs optimize for persuasion. You cannot defend a probabilistic summary to an investment committee, and you certainly cannot defend it to an LP two years after the markdown.
TIER 3 · JUDGMENT
Deterministic Judgment Infrastructure
Stateless, compile-time, reconstructible.
The third tier does not summarize the documents — it compiles the math behind them. Extraction is the only step a language model performs; evaluation happens entirely outside the neural network, in a deterministic engine that routes typed claims against a benchmark universe of 110,000+ Clarity Scores scored on public deal data. The output is not a paragraph. It is a 0–100 Clarity Score across five scored pillars and three audit checks, a hash-anchored IC-ready memo, and a Defensible Audit Log™ that reproduces the same verdict, the same way, every time.
The Doctrine
A vault stores. A wrapper persuades.
Infrastructure compiles for physics.
// THE EVALUATION CRITERIA
Which five questions separate the tiers?
Run any diligence tool through these. The answers will tell you which tier you are actually buying — regardless of what the demo claims.
CLAIM 01
Data retention
Where does the deal data live, for how long, and does it train anything? A vault retains files for the life of the transaction. Ask a wrapper which vendor processes your documents and on what terms. Ask infrastructure for a stated retention ceiling and a no-training commitment, and which of the two is policy and which is enforced. askOdin holds raw documents no longer than 30 days under a documented ceiling and never uses them for training; the benchmark corpus is built on public deal data, not customer files.
CLAIM 02
Verifiability
Ask how a past result is checked. A probabilistic system can return two different answers and cannot tell you why. A deterministic engine returns the same Clarity Score for the same recorded variables under the same Framework version — and keeps those variables in the record, so a past result is read from its record rather than re-run and hoped for.
CLAIM 03
Output format
A summary is not a decision artifact. The fiduciary question is whether the system produces something an IC can act on and an LP can review — a scored, hash-anchored memo, not a paragraph of confident prose.
CLAIM 04
Security model
Every AI tier sends documents to a model somewhere, askOdin included: extraction runs on Google's Gemini API under a paid-tier agreement that prohibits training. The question is what the model is allowed to do. In a wrapper, the model writes the verdict. In deterministic infrastructure, the model only extracts; the verdict is computed outside it, and the source files are held no longer than 30 days.
CLAIM 05
Auditability
When the deal is questioned later — and it will be — can you reconstruct the reasoning? Access logs prove who opened a file. A black box proves nothing. A Defensible Audit Log proves the math.
// THE COMPARISON MATRIX
The same criteria, across all three tiers.
| Criterion | Tier 1 · Data Rooms | Tier 2 · AI Wrappers | Tier 3 · askOdin |
|---|---|---|---|
| Core function | Document custody | Summarization | Claim compilation |
| Data retention | Held for the transaction | Depends on the vendor | 30-day ceiling on raw files; never trained on |
| Verifiability | N/A — no evaluation | Non-reproducible | Deterministic & reproducible |
| Output format | Raw files | Prose summary | 0–100 score + IC memo |
| Role of the model | None | Writes the verdict | Extracts only; verdict computed outside it |
| Auditability | Access logs only | None — black box | Defensible Audit Log |
// TIER 3, CONCRETELY
U.S. PATENT PENDING 63/948,559What deterministic infrastructure produces.
The third tier is not a faster Tier 2. The architecture is different: a language model only extracts; evaluation is deterministic and routed outside the model. The RUNE Protocol™ compiles typed claims against the benchmark universe; cross-document contradictions are surfaced by the RAVEN Protocol™.
// deterministic compile — same recorded variables, same version, same verdict
+ Clarity Score: 0–100, reproducible across analysts and years
+ five scored pillars and three audit checks evaluated outside the neural network
+ Hash-anchored, IC-ready memo
+ Defensible Audit Log™ — reconstructible by an LP later
~ 30-day ceiling on raw documents, never used to train models
- Probabilistic summary: non-reproducible, unauditable
The architectural mechanics of RAVEN's triangulation engine are protected under U.S. Provisional Patent No. 63/994,876 and are not publicly disclosed.
Evaluate the architecture, not the demo.
Venture capital is the last unaudited asset class. askOdin provides the infrastructure to close the gap. Request a scoped pilot and run your own deal documents through it.