Four classes of derived signal
Alpha is what the graph implies but no document states. The engine composes it from verified facts, scores the trajectory a company is on, and is willing to argue both sides of the same evidence.
What facts are individually mundane but collectively imply something the market has not priced?
“Experienced founders, $2M ARR, $10B TAM” is not intelligence — everyone can see it. Derived alpha is the conclusion that only appears when six unremarkable observations are held in the same frame.
No single row would survive a diligence call. The composition would change the valuation.
Founder alpha
A founder whose LinkedIn reads ordinary but who worked directly under the researcher who invented the underlying technology.
- Unusually strong prior collaborators
- Repeated success in hard environments
- Obscure research credentials
- Hidden network advantage
- Ability to recruit exceptional people
Technical alpha
Evidence that a moat is being built rather than described. Produces a technical-moat probability with citations attached.
- Patent filings and assignment records
- GitHub activity, package releases
- Benchmark deltas over time
- Unusual technical hiring
- Proprietary datasets, architecture shifts
Commercial alpha
Demand that shows up in other companies' behaviour before it shows up in a press release.
- Customer job posts naming the product
- Undocumented integrations
- Review velocity, traffic inflection
- Developer → enterprise expansion
- Procurement documents, reseller deals
Network alpha
Mediocre current numbers with extraordinary adjacency to the right people is a form of optionality that scoring models miss entirely.
- Founder and investor networks
- Advisor and early-customer networks
- University and former-employer graphs
- Syndicate and accelerator lineage
Classes light up as the observations above contribute to them.
Founders are scored on twelve axes, not one number.
A single founder score is a compression artifact. The engine reconstructs the history and reads the shape of it — because the slope of a career is more predictive than its starting point.
Evaluating the current product is the least interesting thing you can do.
The engine forecasts what a company is likely to build next, timestamps the forecast, and revisits it. Over enough companies that turns the system into something measurable — you can score the intelligence, not just the startups.
Founders build shift-scheduling tools inside a hospital system.
AI scheduling software sold to mid-size US health systems.
A large proprietary dataset of healthcare staffing behaviour across 40 states.
Labor optimization engine priced against overtime and agency spend.
A healthcare workforce operating system.
Worked trajectory for a healthcare staffing company. The same five-stage reconstruction runs for every tracked company.
If this succeeds spectacularly, what must become true?
Then: which of those things are already becoming true? A success-condition model is more defensible than a score, because each assumption can be tracked independently and revised when the world moves. Drag the conditions to see the expected value move with them.
The amber conditions are the investment. The others are macro bets many companies share; the amber ones are what this specific team either solves or does not — and they are what the pitch deck spends the least time on.
A single model pretending to be unbiased is worse than two models openly fighting.
The bear agent is told only to kill the investment. The bull agent is told only to build the strongest case. Both cite evidence ids. A judge then names the one fact that would move the decision most.
Kill thesis
- Patent ownership may overlap with a founder's former employer; assignment is not recorded with USPTO.
- 61% of revenueappears attributable to two customers, both in one vertical.
- Pricing may not support inference costs at scaleunder the observed usage mix.
- Competitor capability distance is asserted from public artifacts only; no private benchmark has been observed.
Bull thesis
- Hiring pattern suggests a transition into custom inference infrastructure, not model fine-tuning.
- unpublished but directly relevant prior IP held by the CTO, filed under a previous employer's assignment.
- independently hiring around the product — three early customers posting integration roles that name it explicitly.
- 12–18 months behind is where the nearest competitors sit on the public benchmark artifacts.
- The category is inevitable at scale regardless of whether this team wins it.
Both cases rest on the same evidence and reach opposite conclusions about what it funds. The conflict is resolvable from public record: patent ownership may overlap with a founder's former employer is checkable against filings today. Resolve that before the meeting — it moves the decision more than any other single fact.