Back to All Case Studies
venture-buildingFull Access Unlocked

A Venture Engine Where the Model Cannot Talk Its Way Past the Gate

Ask a language model to judge a business idea and it will argue itself into a yes. We built a screening engine where the verdict is derived by code, and the cross-examination runs on a different model.

European venture builder · corporate innovation

venture-building
7
pipeline Stages
3
agents
2
models Cross Examining

!The Need

A venture builder screens far more opportunities than it can fund, and the decision that matters most is the kill. Large language models are the wrong tool for that decision by default: they are persuasive, agreeable, and a model asked to evaluate a thesis it just wrote will reliably talk itself into a go. Same-model self-critique does not help — a model and its own critic share the same blind spots. The team needed a pipeline where an AI could do the thinking without being trusted with the verdict, and where a kill decision could be defended to a sponsor months later.

The Approach

The core design decision is that agents are stages and the method is the control flow: the model argues, the system decides. Three agents run across seven stages with typed, cross-validated hand-offs, each stage adding one key to a single accumulating artifact. Contrarian thinking is forced structurally — belief mining runs as a tournament of three to five rival theses, each of which must cite the specific assumption it negates, and code promotes the winner. The red team then runs on a deliberately different model, fact-checks load-bearing claims against live search, attacks the causal chain and classifies weakness severities — but emits no verdict at all. Deterministic code derives stop or go from those severity classifications, and a second coded gate forks the opportunity toward a corporate or venture-capital buyer. Every run persists, including abandoned ones, so the kill-memo archive is queryable rather than prose, and an evidence ledger lets a later refuting datapoint trip a pre-declared kill condition: the code proposes the kill, a human confirms it.

Technologies Used

ClaudeCross-model red teamZodTypeScriptSQLiteStructured generation

The Output

Seven pipeline stages across three agents, with schema-validated hand-offs that re-ask the model when output does not conform
A contrarian belief tournament in which each rival thesis must cite the assumption it negates
A red team running on a different model — it classifies severity and is structurally prevented from issuing a verdict
Stop/go and the corporate-versus-venture-capital fork derived by deterministic code, not by the model
Every run persisted, abandoned ones included, with a queryable kill-memo archive
An evidence ledger where a refuting check-in trips a pre-declared kill condition and flags the run for human confirmation

The Impact

The stop/go verdict is no longer the author's to give — it is derived from a different model's severity findings by code
A persuasive model cannot rescue a weak thesis: the methodology gates are code paths, not prompt instructions
Blind spots shrink because the critique runs on a model that does not share the author model's priors
Kill decisions became defensible months later, with the reasoning and the refuting evidence still attached to the run
Abandoned ventures stopped disappearing — the portfolio keeps what it rejected, and why

Ready to achieve similar results?

Contact Us
Supalabs AI solutions