# IVI Methodology Version 1.0 · Published 2026-09-05 This document is the long version of the IVI methodology — the trust surface behind every fund-statement figure we publish. The public `/methodology` page is the summary; this write-up expands each claim to the level of detail a technical buyer, a due-diligence reviewer, or an auditor needs: the guarantee in full, the mechanics by which a figure earns "proven", the taxonomy of holds, the benchmark method with every caveat, and why we publish the caveats rather than hide them. **Maintained against:** the repo canon — `dream.md` §1/§2/§4 and the measured deploy-gate records of 2026-08-19 and 2026-08-21. Every number in this document must stay traceable to those records; this document updates when the canon updates. **Change log:** Version 1.0 — initial publication; numbers identical to /methodology as of 2026-09-05. --- ## 1. The guarantee IVI extracts financial figures from private-market fund documents: capital account statements, capital call notices, quarterly fund reports, and feeder statements. Anyone can promise extraction. We make a guarantee that can be broken in public: > **Every figure we show you is either proven — backed by an evidence trail > you can hand to an auditor — or held for review. A wrong figure is never > shown with confidence.** There is no third state. A number is either accompanied by a proof artifact (verbatim source text, page, coordinates, and the checks it passed) or it is parked in the hold queue with an honest, specific reason attached. It is never quietly saved as if it were verified. That guarantee — not an accuracy percentage — is the product. Scores move between vendors and between test sets; a proof artifact does not. A vendor can always find a sample on which it scores well. What a vendor cannot fake is a per-figure evidence trail that survives being checked, figure by figure, against the documents it claims to describe. The guarantee is also what allows honest failure. When IVI cannot prove a figure, it says so, visibly, next to the figures it did prove. The alternative — guessing and displaying the guess with confidence — is the one failure mode this product is built to make impossible. ## 2. How a figure earns "proven" Five conditions must hold before a figure is shown as proven. They are enforced by deterministic code, not by model self-assessment. ### 2.1 Deterministic first, models last Language models read the document; deterministic code decides what the reading becomes. A model can propose a value; it can never approve one. The save gate is a mathematical stamp — the identity check and the validation rules — that no model can widen. If the proposal does not satisfy the stamp, the figure does not save as proven, whatever the model's confidence in it. This split is not a style choice. Every mechanism in which a model both proposes and validates its own proposal is a mechanism in which a persuasive hallucination becomes a persisted figure. We treat that as the worst-case failure and structurally exclude it. ### 2.2 Identity math on every transaction Each statement must satisfy the fund-accounting identity before it saves: ``` opening balance + contributions − distributions + income − fees ± carry = closing balance ``` Every transaction-bearing statement is tied against this identity at save time. A statement that does not tie is **not silently corrected** — the discrepancy is surfaced, not absorbed into the stored figures. If the numbers do not add up, you learn that they do not add up; IVI does not invent a closing balance that does. ### 2.3 Provenance to the page Every figure traces to its verbatim text, its page, and its coordinates in the source PDF. Click a number, see the document — the actual rendered page, with the source passage highlighted. Figures derived from the statement's own arithmetic rather than printed directly are labelled as imputed, never passed off as quoted. There is no figure on an IVI screen that cannot be traced, end to end, to the page it came from. ### 2.4 Auditor conventions, declared Documents are sometimes ambiguous: a label that could mean two things, a sign that could be presented either way, a total that could be read from one of two places. When that happens, IVI applies the convention a certified auditor would expect — and records, in the figure's provenance, that a convention was applied and which one. It never resolves the ambiguity by picking whichever value happens to match a ground-truth record. That would be optimising for the benchmark rather than the truth, and the convention would be invisible to the auditor who later checks the work. The convention is declared even when a different choice would have scored better. ### 2.5 No self-approval The extractor is never its own validator. Verification runs on independent signals, and a value back-solved from a check can never pass that same check. Concretely: if a figure is solved *from* an identity equation, then checking that figure *against* that same equation is tautological — it passes by construction and proves nothing. Such a value is marked invalid by that check and must be corroborated by genuinely independent evidence before it can earn proven. ## 3. We check the GP's math, not just our reading Most extraction tools ask one question: *did we read this correctly?* IVI asks a second: *is the document itself correct?* Issuers make errors, and a failed check has two possible authors — the reader and the issuer. When IVI reads a document correctly and the document still does not tie, that is not an extraction failure. It is a finding, shown to you with the page as evidence. Telling an LP that a GP's Q3 statement does not tie by €4,300 — and showing the page — is the most valuable thing we produce. No extraction-accuracy percentage, however high, would have surfaced that error; the error is in the source document, and only the second question finds it. This is why the identity check runs on every statement regardless of how clean the read was, and why issuer-side errors are reported as findings rather than absorbed to make the numbers tie. ## 4. Earned holds — exactly four reasons A hold is a claim that the entire capability chain was exhausted, and that claim has to be true. Only four reasons earn a hold: 1. The document contradicts itself beyond tolerance. 2. Two trusted sources disagree on the same fact. 3. A material figure is genuinely absent and underivable. 4. The fund/period/investor identity is genuinely undeterminable. Everything else — a missing figure that is derivable from the fund's own history, a convention question, an ambiguous label — is resolved and flagged, not parked. The hold queue is therefore rare by design: it exists for genuine exhaustion, not as a resting place for hard cases. Held documents sit next to verified ones, clearly marked, never hidden. A hold is not a failure to be buried; it is a statement about the document, made to the user, with the reason attached. ## 5. Measured, with caveats We publish our benchmark because a buyer needs it, and we publish its caveats because a buyer needs those too. ### 5.1 The benchmark corpus Our benchmark is not a hand-picked demo set. It is the complete historical portfolio of a real fund-of-funds: **268 unique documents across 49+ funds**, all live quarterly reporting, scored against the LP's own hand-maintained records. Every miss was adjudicated against document evidence — not averaged away into a single number. ### 5.2 The results, and the honesty row - **Median field accuracy 90.4%** on the canonical corpus at the last two production deployments (hold-out set: **100%**). - **The honesty row:** at this sample size, statistics cannot *prove* a ≥90% claim — the confidence interval is wide (roughly **81–95%**). That is exactly why we sell the per-figure guarantee rather than a score: the guarantee holds at any sample size, and the score hardens as new client corpora arrive. - **≤ €0.05 and under 60 seconds per document** on the production stack — measured per quarter, with a regression triggering a route review. ### 5.3 What the numbers do and do not mean The accuracy figure is a field-level median over the canonical corpus. It is not a promise about any individual document, fund, or client, and it is not smoothed over the corpus's hard cases — the misses are adjudicated one by one, which is also why the corpus is the corpus: re-running on it is how we detect regressions at deploy gates, not how we market. All figures above concern **statement figures** — the numbers on fund statements — not contract terms. --- Why publish a caveat against ourselves? Because a vendor quoting "97% accurate" without a sample, a method, or a proof artifact is asking you to trust the number. We are asking you to trust the evidence — and to check it. The methodology above is the standing offer: every figure proven, or held with a reason you can inspect.