§Loading dossier
How ProjectX evaluates companies and startup ideas across multiple strategic frameworks — from research collection to scoring to cognitive profiling.
Developed by Dr. Manoj Thomas at Cornell University, CVA measures how well a company creates and captures customer value across six interconnected dimensions.
Each dimension captures a different facet of how a company creates, communicates, delivers, and sustains value for customers:
The first three (Product, Customer, Message) are value drivers — they directly create customer value. The latter three (Organization, Ecosystem, Finance) are value enablers — they sustain and scale it.
We use a deliberately coarse three-level scale to reduce noise and make scores directly comparable across companies:
Exceeds benchmarks. This dimension is a competitive advantage — it actively pulls the business forward.
Meets requirements. On par with competitors — not a differentiator, but not a drag either.
Underperforms. Creates friction or confusion — this dimension is actively holding the business back.
Why not a 1–100 scale? Fine-grained scores suggest false precision. Is a company’s product really a “73” vs. a “71”? The three-level system forces the model to make a clear call while keeping results reproducible across runs.
The overall score is not a simple average. The analysis follows a top-down methodology:
The binding constraint is the single dimension that most limits the company’s overall value creation. It’s not necessarily the lowest-scoring dimension — it’s the one whose weakness causes weakness in other dimensions. Think of it as the bottleneck in a system. The analysis also identifies a remedy: the most actionable lever for improvement, which may target a different dimension than the root cause.
When analyzing a startup or early-stage idea, the model switches to predictive mode:
stage_constraint when the score reflects stage limitations rather than fundamental flawsEach analysis runs a 27-query research sweep via Exa.ai, a neural search engine. Queries are organized across 8 groups:
The raw results are filtered and formatted into structured text, then passed to Claude as the primary data source. The model supplements with its own knowledge only where research is thin.
Analysis is performed by Claude (Anthropic), defaulting to claude-sonnet-4-6 at temperature 0 for maximum determinism. The model receives a structured system prompt defining the CVA framework, dimensions, scoring rubric, and methodology, then a user prompt containing the formatted research. The entire output is a single JSON object with scores, explanations, and strategic insights.
All analyses use temperature 0 with pinned model versions and locked prompts. The 0 / 50 / 100 scoring and structured research pipeline are designed so the same inputs produce the same outputs across runs. The admin workbench allows A/B testing of pipeline configurations to validate changes before promoting them to production.