Finance lens

The premise: most of an analyst's hours on a private-company valuation aren't spent on judgment. They're spent reconciling what the client sent, a stale option pool, a SAFE that should have converted at the last round, a side letter with terms nobody transcribed, an anti-dilution column that says "see charter." Cleaning that file before any math runs is the work the tool does. Valuation, OPM allocation, DLOM, 409A, all explicitly out of scope. The tool is the input layer to the analyst's workflow, not a replacement for it.

Engineering lens

No LLM in the calculation pipeline. No probabilistic scoring of findings. Every checklist rule was authored by hand against a documented kind of real-world cap-table mess. Each finding carries a fields_referenced tuple pointing at the exact source cell, so any reviewer can verify the flag by opening the workbook. Side letters are stored verbatim; open Q-prefixed lines surface for in-session adjudication; no model is allowed to "interpret" charter terms.

Finance lens

Anti-dilution variant blank · full-ratchet documentation missing · unconverted SAFE past trigger · stale option pool · off-table warrant · side-letter scope unresolved · participating-with-cap missing the cap multiple · dual-class voting differential not recorded. Each was chosen because it is a documented practitioner gripe in NVCA Model Charter commentary, Singapore VIMA, or worked-example precedent. Findings sort blocker → warning → info; blockers prevent the waterfall from running until adjudicated.

Engineering lens

Each rule is a pure function over the parsed CapTable model (pydantic v2). Findings with a deterministic resolution shape (SAFE conversion share count, AD variant picker, warrant inclusion toggle, side-letter scope, pool date refresh) have inline HTMX-driven resolve forms. Submit, the cap table updates in memory, the waterfall re-runs, findings re-evaluate. No page round-trip, no model deciding what counts.

Finance lens

The waterfall produces the breakpoint set an OPM Backsolve consumes: LP-cleared thresholds, conversion thresholds, participation-cap thresholds, pure-conversion thresholds. Handles non-participating, participating-uncapped, and participating-with-cap including the senior-above-capped case (a Delaware double-cap fixture stress-tests this path). A cumulative-payout chart renders the per-class slope changes; the tranche allocation matrix shows the fraction of a marginal dollar going to each class within each tranche.

Engineering lens

A what-if scenario panel lets the analyst override share counts and LP multiples; the chart redraws and the breakpoint table populates with deltas within ~400ms. In-memory only, the saved cap table doesn't change. The same engine drives both the in-browser view and the exported live-formula .xlsx, where every breakpoint cell carries an IF/SUMPRODUCT formula against named input ranges. Edit a share count in Excel, the chart and breakpoints update natively without Python.

Finance lens

Real cap-table cleanup runs into side letters constantly, MFN clauses with undefined scope, redemption rights whose enforceability is in dispute, IPO vetos at specific valuation thresholds. The temptation is to "extract structured terms" via an LLM and apply them as overrides. That's exactly what an audit-defensible tool shouldn't do. Instead, the PDF is parsed text-only with pdfplumber; the analyst records counsel's adjudicated answer into a textarea that gets appended to the side-letter body. Image-only scanned PDFs return an empty-text warning rather than a fabricated transcription.

Engineering lens

pdfplumber-based extraction. Heuristic title detection (first non-blank line < 120 chars without a trailing period). Open-question detection scans for lines starting with Q:, Question:, or TODO:, these populate an unresolved_questions array that the checklist surfaces as a side-letter-scope-unresolved finding. The analyst's adjudication clears the finding and appends to the canonical side-letter body so the audit memo cites both the source PDF and the in-session resolution.

Finance lens

A sixth fixture (Sundar Foods, fictional Indian D2C Series B-1) is included as a stress test, not a curated demo. It's built the way Qapita's actual customer base sends cap tables: holder-level rows (one per shareholder, not per class), tab names that don't match the parser's whitelist ("Conv. Instruments", "Co. Info"), prose in numeric columns, dates like "14th Feb '25", and off-charter terms living only in side-letter PDFs. The engine intentionally degrades on it, only the ESOP rows survive validation. The result is a 72-warning parse report rather than a fabricated waterfall.

Engineering lens

That report is the deliverable on a real-world file. The audit-defensible record of what an analyst would email back to the CFO before any cleanup work can begin. The roadmap fix (holder-level rollup pass, broader tab-name whitelist, fuller class-type alias map covering "Equity Common" / "CCPS Series B-1") is real and known, and not shipped yet, on purpose, until I've had a conversation about which customer-file patterns to prioritize. Visible-honest beats invisible-broken.