Engineering brief
How Verdict learns from every memo.
A multi-step read produces the memo you see. The signals underneath it are what compound. Eight of them stack: the benchmark anchors every call, a per-fund profile personalizes without bleeding one fund's data into another, audit memory carries lessons from one memo to the next, imported context collapses switching costs, personal preferences tune voice without moving the call, an anti-bias guard treats your context as evidence and never lets it inflate the call, a sources whitelist constrains citations, and a fabrication detector catches publisher URL mismatches. Each is an independent signal. Together they compound.
01
Benchmark suite
A held-out set of real pitch decks with public outcomes anchors every update. Theranos. FTX. Fyre. WeWork. Buffer. Notion. DoorDash. Airbnb. Each deck has a known result, and Verdict has to land its read on those decks before any update goes out. The set re-runs on every meaningful change and grows as new known-outcome decks are added.
A new fund's first 50 memos benefit from the standard the set already enforces. Day-one output lands in the right range instead of learning from scratch.
02
Per-fund bias profile
Each fund accumulates a bias profile from its own activity: sector lean, stage concentration, recurring dealbreakers, and override patterns from feedback chips on prior memos. The profile silently shapes future memos for that fund, fund-isolated and never cross-contaminated across accounts.
After 8 portco memos, Verdict reads new decks through your fund's accumulated thesis instead of a generic, one-size read.
03
Audit memory
After every memo, a fast structural audit reviews the analysis for citation gaps, numeric inconsistencies, bull/bear asymmetry, and tone discipline lapses. Findings persist as lessons learned and silently inject into the next memo for the same user, so Verdict learns from its own past mistakes without anyone teaching it.
After 50 memos, Verdict writes in your reading style without you correcting it. The lessons compound; the chat session does not.
04
Imported context
On import, Verdict ingests the fund's own past memos, an existing system prompt, or a folder of prior reads. Imported context becomes orientation for the synthesis pass, so Verdict starts memo one already aware of the fund's structural reading patterns.
First memo reads as if Verdict has been reading for the fund for months. Switching costs collapse from week-one to day-one.
05
Personal preferences
A stable per-user preamble carries voice (sharp / measured / narrative), verdict bias (skeptical / balanced / optimistic), and priority focus (market / team / product / unit economics). Loaded as orientation at the start of synthesis for users with 10+ reports.
A managing partner who marks 'verdict too aggressive' three times gets calmer prose without the call itself softening.
06
Anti-bias guard
User-supplied context (founder relationship notes, fund preferences, sector mandates) is treated as evidence, never as a thumb on the scale. Verdict may use the context to weight what to investigate; it cannot use the context to inflate a verdict beyond what the deck and the sourced market evidence support.
Your edge does not corrupt your reading. The verdict on a portfolio company holds to the same standard as on a cold deck.
07
Sources strict whitelist
Citations are constrained to a published whitelist of academic publishers, regulated financial data sources, and reputable reporting outlets. Sources outside the whitelist are dropped at parse time, with one carve-out for Crunchbase and Pitchbook funding-history rows where the data is structured and the alternative is a fabricated funding round.
Every market-size figure traces to a citable source the partner can re-pull on Bloomberg or in Pitchbook. No mystery URL paths.
08
Citation fabrication detection
URL-shape heuristics flag known fabrication patterns across major publishers. A Bloomberg URL with a shape Bloomberg never publishes is flagged before the citation reaches the memo. Findings persist as audit lessons so Verdict learns the publisher's URL grammar over time.
A fabricated citation in a senior-partner-forwardable memo is the only error worse than a wrong verdict. The detector closes the gap.