How VentureGPS Works
The actual production pipeline, stage by stage -- what's an AI judgment, what's a fixed calculation, and what neither one guarantees.
Probabilistic vs. deterministic, in plain English
A language model reads text and makes a judgment call -- what the evidence says, how strong it is, what a fair score looks like. The same input usually produces a very similar result, but not a mathematically guaranteed identical one, and the judgment itself can be wrong even when the process was followed correctly.
Plain code applies a fixed formula to numbers that already exist -- a weighted average, a confidence cap, a rule-based override. The exact same inputs always produce the exact same output, every time, with no judgment involved at this step.
This distinction has a real limit
A reproducible calculation only means the math is consistent -- it does not mean the AI-generated judgments or the underlying claims about the company are correct. A confidently-stated, perfectly-reproducible score can still be wrong if the evidence it was built from was incomplete, misleading, or simply mistaken. Always treat VentureGPS as one structured input to your own judgment, not a verified fact.
Startup information
Your inputYou provide a company website, a pitch deck, a plain-language description, or any combination.
This is the one part of the pipeline with no AI involved yet -- it's exactly what you submit, combined into one input for everything below.
AI-assisted research
Probabilistic (AI judgment)VentureGPS generates targeted search queries, retrieves public information, and synthesizes it into a research brief.
A language model writes the search queries and the brief; a real web search actually retrieves the pages. The brief is instructed to separate verified facts from assumptions and to say UNKNOWN rather than invent numbers -- but it is still an AI-written synthesis, not a verified fact-check.
AI-generated assessments
Probabilistic (AI judgment)Six Intelligence Pillars (Market, Team, Product, Execution, Traction, Financial Health) are each assessed in two steps: evidence extraction, then scoring.
For each pillar, one AI call reads your submission plus the research and decides, per dimension, whether it found direct evidence (Observed), an indirect signal (Inferred), or not enough to responsibly assess (Unavailable) -- never guessing a number for a dimension it can't support. A second AI call then judges a 0-10 score for whatever evidence was found.
Deterministic validation and scoring
Deterministic (fixed rules)Every score is then processed by fixed, non-AI rules -- the same rules, every time, for every analysis.
A confidence cap (a Low-confidence finding can't reach a 9 or 10, no matter what the AI judged). Rule-based overrides for a small set of dimensions where structured facts (like a two-point revenue series) exist -- those are computed in plain code, never an AI's number. A weighted average turns dimension scores into a pillar score, and pillar scores into the one overall VentureGPS Score, always excluding (never zeroing out) whatever couldn't be responsibly assessed.
VentureGPS Report
OutputOne VentureGPS Score, six pillar breakdowns, evidence, confidence, strengths, weaknesses, and recommendations.
Everything above is what you're actually reading in a report -- including which specific findings come from real evidence, which are AI inferences, and which dimensions simply couldn't be assessed. Nothing in the report is added after the fact that wasn't already computed here.