What changed, why it changed, and which document caused it.
Every KPI on a Capital Refinery deal carries a side-by-side reasoning trail. When a number moves between two readings, the platform names the document that drove the change, surfaces the adjudication math that picked the winner, and refuses to soften its own reasoning when a lower-scoring candidate wins on a plausibility adjustment.
Diligence partners spend half their time on a question the platform can already answer.
Two readings of the same deal, two months apart. A KPI has moved. The partner running second-round diligence opens last quarter's memo next to this quarter's workbook, lines them up in Excel, and starts asking the question every committee asks: did this number change because the data changed, or because the analyst's assumption changed?
The current industry answer is a file-version hunt — pull the prior workbook from the data room, compare cell by cell, chase the analyst for which tab the figure came from, and reconcile against a quarterly model that was already a snapshot the day it was built. The cost of getting it wrong is not theoretical. It is a re-papered memo, a second-round round of credit-committee questions, and deal velocity lost to a reconciliation that should have taken one click.
The platform already carries the answer. Every KPI on every deal records the reasoning that produced it, the candidates the adjudication considered, and the document that drove the most recent change — for every reading, every time.
Two mechanisms. One question answered end to end.
When a diligence partner asks “what changed since I last looked at this deal,” the platform answers in two parts: the reasoning trail behind the number, and the document that caused the most recent move. Both are live on every KPI on every deal in the tenant.
Point in time vs. point in time
Open any KPI. The popover shows the prior reasoning tree on the left, the current reasoning tree on the right. Method, confidence band, candidates considered, derivation formula, contested alternatives — for both readings, side by side. Numeric deltas are highlighted. Changed method codes or formulas are amber-bolded. The partner sees the reasoning drift, not just the number drift — which is the difference between a clean answer and an open question heading into IC.
Which file drove the change
When a number moves, the platform names the file that drove it: “Current state updated via new financial statement ingested (source: Q1_financials.xlsx).” Click the source chip and the evidence atlas opens at the right document with a banner pointing at the relevant cell. The full storage URI, the system-internal identifier, and any path component that could leak a firm's internal storage layout are stripped before the chip is ever rendered.
When the adjudication is contested, the platform names the contest.
Extraction adjudication picks a winner from the set of candidate numbers the platform observed across documents, tabs, statements, and operator entries. Most of the time the winner is also the highest-confidence candidate, and the reasoning trail says so plainly. Some of the time it is not. A plausibility floor, a symmetric outlier penalty, or a cross-KPI consistency sweep can promote a lower-confidence candidate over a higher-confidence one — because the higher-confidence one was implausible against the surrounding financials.
When that happens, the platform refuses to phrase the result as if the winner simply “exceeded” the alternative. The reasoning trail names the adjustment that promoted the lower-scoring candidate — the plausibility floor that the high-confidence number failed, the outlier penalty that the surrounding KPIs imposed, the consistency check that flagged it. A test catches any attempt to soften that explanation in the rendered output. The honest reason is the only reason that ships.
The LP-safe side of the same mechanism
The attribution mechanism that surfaces filenames to operators is sanitized for LP-facing views. The full storage URI, the system identifier, the directory tree, the dotfile prefix, the tenant-keyed path component — all stripped before the LP page renders. The LP sees the leaf filename (“Q1_financials.xlsx”), the document type (“Financial statement”), and the operator's date. The firm's internal storage layout never leaves the operator surface.
The sanitization is enforced in three places — the change event itself, the human-readable reason rendered into the LP view, and the source chip that opens the evidence atlas. A leak in any one of them would be a leak; the platform rules out the leak before the chip is composed.
Who reaches for this first
- Diligence partners running the second-round drill — "why is this number different from the last memo?" answered in one click, with both reasoning trees side by side
- Credit officers in covenant reviews where a 10bps move could mean a waiver, and the source of the move is the difference between a renegotiation and a standstill
- LP-facing partners who need to answer DDQ questions about evidence lineage with filename-level specificity — without exposing any of the firm's internal storage layout
- Audit committees and CCOs whose record-retention obligations require document-to-decision traceability across every KPI on every deal
What makes this work
Three pieces, all live in the platform today. Every KPI carries a reasoning trail with method, confidence, candidates, and derivation — recorded at the moment the reading was produced. Every change to a KPI carries the document that drove the change, with sanitization applied before the filename is surfaced. And every contested adjudication carries the rule that picked the winner, rendered honestly in the reasoning trail and protected by a test that fails the build if the language softens.
The diligence partner's question — “what changed, why did it change, which document caused it” — is answered in three clicks: open the KPI, read the side-by-side reasoning, click the source chip. The evidence atlas opens at the file, banner on the cell, and the trail closes.
Bring us one position with a contested KPI.
We'll show you the reasoning trail end to end — the file that drove the change, the adjudication math that picked the winner, and the version of the reasoning at the moment the prior committee committed.