AI made alpha claims cheap. Veridict makes belief expensive again.
One claim enters. A frozen specification, hash-bound evidence, hostile tests, and theorem-backed thresholds return a typed verdict.
The court does not ask whether a claim looks attractive. It asks whether the claim deserves belief after search, leakage, costs, nulls, delay, multiplicity, and hindsight have been priced.
AI can generate thousands of plausible strategies, codebases, charts, and polished investment narratives. That abundance does not create trust — it creates an arms race in plausibility.
Future information masquerades as insight.
Best-of-many scans return lottery winners.
Drift, carry, trend, or beta gets relabeled as unique alpha.
Gross effects die when delay and costs arrive.
That moment is becoming more frequent, more automated, and more dangerous. The platform does not compete with managers for returns. It competes with false certainty for authority.
Do not argue with the story. Force it through a constraint system where the story cannot rewrite itself. Scroll to walk one claim through the court.
Strategy, scope, parameters, data, costs, and bars are written down — in full — before anything runs. The seller commits to the test they will be judged against.
Specification, data, parameters, lineage, and receipts are hash-bound before judgment. The story can no longer change after the number appears.
The claim is treated as a gambler betting against "no edge net of costs." Its bankroll is its evidence — and the bankroll is anytime-valid, so repeated looks earn no free pass.
A replayable certificate returns one of four typed outcomes — issued before capital, counsel, contracts, or reputation move.
Veridict refused the famous, louder evidence and recognized the quieter real premium — without inflating it into a claim. That separation is the product.
| evidence | 16.177 |
| claim bar | e ≥ 20 · not met |
| search price | best-of-12 · p=0.083 |
| verdict | CLEAN NEGATIVE |
| single evidence | 2.483 |
| delayed evidence | 2.159 · no collapse |
| multiplicity-taxed | 1.074 · deferred |
| replication | IR 0.448 vs 0.47 |
A normal validator would crown the taller bar. Veridict priced the search and refused it. The smaller bar was independently reconstructed through a provably uncontaminated route and recognized as real — while still held below certification. Neither verdict flatters the seller. Neither flatters the skeptic. That is what trust looks like.
A diligence product must be neither a pessimist nor a cheerleader. It must tell fraud, noise, dead effects, real-but-uncertified premia, and certified evidence apart.
| Failure mode a fake validator commits | Filed demonstration Veridict avoids it |
|---|---|
| Certifies a planted future-data leak | Poisoned explorer posted 6× the real edge, then collapsed 91% under the t+2 delay rung → rejected. |
| Certifies noise via multiplicity | Best-of-210 dead-world scan "discovers" alpha 96% of the time; priced mixture holds at 1–4%. |
| Certifies known-dead strategies | 13 graveyard corpses stay dead on redeployment; the weekend effect returns negative. |
| Launders hindsight into authority | The cheater promoted at chance and held zero authority; retroactive credit is unrepresentable in the algebra. |
TSMOM-12-1: IR 0.28, cost- and delay-robust, kept as the live trend organ. Carry: independently reconstructed through a clean-room route within 0.022 IR. ES overnight: real gross, correctly killed by friction. The engine can say "yes, structure" and "no, not a certified claim" in the same breath.
Veridict does not ask a hostile reviewer to believe the founders. It hands over the record: hashes, source, gates, receipts, proofs — and a command that checks what the package says it contains.
False-claim probability is bounded by the Ville threshold under the proved model.
The proof audit reports standard axioms and no proof holes in the checked obligations.
The system refuses to convert sub-bar evidence into performance claims.
That is a different category of asset: verifiable research. The formal theorem work covers the core evidence algebra over exact rational models; float execution and data/vendor constraints remain disclosed modeling layers, and source-bundle proof replay is skipped honestly when mathlib is not provisioned.
Most analytics companies are paid to make a buyer comfortable. Veridict is valuable because it can make the buyer uncomfortable — with receipts. That incentive difference compounds.
The platform's prestige comes from refusal. A rubber stamp destroys the brand.
The claim boundary is tied to machine-checked kernel obligations, not a slogan.
Every artifact, gate, mistake, repair, and receipt accrues into a reproducible record.
Dead strategies become standing decoys. Negative knowledge becomes infrastructure.
Once verdicts travel with claims, the market asks for the receipt by default.
The lineage records 26 typed errors — all human or spec errors — each converted into a structural defense. That is not a blemish. It is the asset: the system learned how its operator could be wrong and made the repeat offense harder.
Veridict does not need to own the trading stack to become valuable. It needs to own the trust layer in front of it.
SEC Form PF private-fund gross asset value, Q3 2025.
Preqin / BlackRock forecast for global alternatives AUM.
BlackRock estimate for private-markets data TAM by 2030.
Bank of America estimate for global family-office assets.
| Phase | Commercial target | Proof point |
|---|---|---|
| 1 · Refutation service | Paid external submissions from allocators, family offices, counsel, or honest vendors. | Bad claims die quickly; good claims receive a defensible next step. |
| 2 · Certification layer | Repeatable receipt package and buyer-recognized evidentiary grades. | Vendors bring claims to Veridict before buyers ask. |
| 3 · Standard | Portable verdicts, APIs, and recurring diligence infrastructure. | The market treats unreceipted claims as lower-trust claims. |
Veridict's narrow first market is quantitative finance — the pain is acute, the stakes are high, the claims are testable. The larger meaning is broader: as machines produce more claims than humans can inspect, markets will need institutions that decide what deserves belief.
A paid diligence product that kills weak claims, receipts strong-but-uncertified ones, and prevents premature authority.
A repeated standard for managers, vendors, allocators, and AI-alpha markets: bring the claim, leave with the receipt.
A portable discipline for claim generation, hostile testing, proof, and admissibility across high-value domains.
The world is about to drown in plausible machine-generated assertions.
Veridict is the machine that decides what can be believed.
No alpha claim, return projection, or allocation recommendation is made on this page. The commercial thesis is a process-certification and trust-infrastructure thesis. Market figures cite SEC Private Fund Statistics Q3 2025, Preqin Private Markets in 2030, BlackRock / Preqin materials, and the Bank of America Family Office Study.