Product 01 — Discovery: Track what LLMs are saying about your business across providers, personas, and time. Measure mention rates, sentiment, and attribution gaps.
Product 02 — Proof: LLMs are inherently untruthful — predictive vectors, not factual assertions. Trakt provides the cryptographic verification layer that proves commercial assertions are authentic, authorized, and untampered. The only way to verify an AI response holds proof of what it says.
AI models are already recommending products, pricing, and businesses to millions of people. You need to see what they're saying.
"For project management, I recommend Linear for technical teams..."
"Linear is excellent for software teams, with strong issue tracking..."
"I'd suggest looking at Asana or Monday.com for project management..."
See exactly what each model recommends, across personas, geographies, and time. Track citation deltas — when Model A cites you and Model B doesn't, that's a gap you can close.
Of AI responses about project management mention your product
89% of mentions are positive or neutral, 11% negative
Gemini mentions you 24% less than ChatGPT — close the gap
Great for exploration. LLMs are excellent at discovery, comparison, and recommendation. But when it comes to truth — is this price current? Is this inventory real? Is this certification valid? — they fall off a cliff.
They generate plausible text based on statistical patterns. They predict the next token, not the next fact. A coherent answer with a plausible citation may still be wrong, stale, or attributed to an organization that never made the assertion.
The curse of recursion: When models train on model-generated data, the distribution shifts irreversibly. Fluency survives while factual accuracy decays toward a noise floor. The text that remains is progressively less tethered to the originating source.
"Training on model-generated data causes irreversible model collapse. The outputs become the inputs, and the model forgets the shape of the original distribution."— Shumailov et al., Nature, July 2024
Language models generate text by predicting the next token based on statistical patterns in training data. They have no concept of truth — only probability. They will say whatever is most likely given the context, regardless of whether it's accurate, current, or attributed to the right source.
In commerce, a confidently worded but unverified price, inventory state, or safety certification is precisely the failure mode with monetary and legal consequences. Citation is no substitute for provenance. When model collapse accelerates, the problem compounds: each generation is less accurate than the last, and the decay is irreversible.
Every assertion is cryptographically signed by the organization entitled to make it. The signature commits to the algorithm and key identity it was made under.
Hierarchical typed authority with multi-dimensional scope. Delegation can only narrow, never widen. Any widening produces a distinct typed error.
Canonical hashing and Merkle commitment with on-chain anchoring. The assertion cannot be altered without detection. Verification is executable offline from public material alone.
Each layer consumes the output of the preceding layer and adds a distinct cryptographic guarantee. The pipeline is consumed by a verification engine that executes thirteen ordered checks with declared prerequisites.
Decentralized identifiers bind every actor to an asymmetric public key
Signed capability grants bind identity to typed permissions across nine scope dimensions
Child authorities derived through strict-subset operation — only narrowing allowed
Typed statement by an issuer about a subject, carrying scope and provenance
Claim bound to authority with content-addressed evidence, made tamper-evident by canonical hashing and signature
Signature over canonical digest — the algorithm identifier lies inside the bytes that were signed
Merkle tree over sorted canonical leaf hashes with O(log n) inclusion proofs
On-chain registration of root public key and publication of commitment roots
Executes thirteen ordered checks with declared prerequisites, recording each as passed, failed, or skipped. Emits typed trust failures drawn from a closed vocabulary and localized to a delegation hop.
The verification pipeline reads no custody metadata. An air-gapped self-custodied key, a non-exportable key in the customer's cryptographic service, a non-exportable key in the platform operator's cryptographic service, and a smart-contract root yield identical verification results for identical attestation bytes.
An organization may migrate its root key between custody models without re-issuing attestations and without changing any verifier.
Ten JOSE algorithm identifiers resolve through a single registry dispatch keyed by the algorithm an object declares. The routine selected by that dispatch binds the declaration to the actual type, curve, and digest of the key and rejects a mismatch as a structural failure.
Deprecating an algorithm withdraws it from new issuance without invalidating material already signed under it, because the algorithm identifier lies inside the bytes that were signed.
The binding factor for agentic commerce: an organization's root can be an account controlled by executable code on a ledger, not just a raw key. The same verification pipeline reaches it through ERC-1271 and binds each check to one chain and one contract, so a proof can't be replayed across networks.
A signature algorithm plus a public key. Verified locally against the ten JOSE algorithms — ECDSA, RSA-PSS, RSA-PKCS1, and EdDSA.
KeyMethod {
algorithm: "ES384",
publicKey: <P-384 point>,
}A smart-contract account — multisig, agent wallet, or DAO. The contract itself validates the signature, scoped to one chain and one address.
ContractMethod {
algorithm: "ERC1271",
chainId: 8453, // Base
contract: "0x…",
}
// IsValidSignature(chainId, contract, digest, sig)The organization's root public key is registered in distributed-ledger state, so the binding of organization to key is public and timestamped rather than asserted. Verification bootstraps from ledger state with no trust in the publisher's database or any resolution endpoint.
The ledger operates as a transparency log that records submissions and guarantees their inclusion, ordering, and immutability. Equivocation — two competing registrations of one organization identifier — is detected and reported as unresolved.
Or does? Today, that price flows through a pipeline that shouldn't exist. Tomorrow, it doesn't.
No resellers. No indexers. No tracking down web pages. No scraping. No "according to multiple sources." The brand publishes directly. The AI verifies directly. The consumer gets proof.
We collapsed a pipeline that shouldn't exist.
Three stakeholders win. Brands get verified. Consumers get truth. LLMs stop the collapse. We don't build a chain — we coalesce on Base, Ethereum, and Optimism. Our APIs are open. Our contracts attest proof.
Register your organization on-chain. Your root public key becomes a public, timestamped identity on Base, Ethereum, or Optimism.
Publish your business data — pricing, inventory, certifications — with cryptographic authority. Sign it with your registered key.
AI models verify your assertions from public material alone. You win trust without intermediaries.
Ask an AI model for a price, a certification, a product detail. Get an answer with cryptographic proof attached.
Verify the claim yourself — or let your agent verify it. No trust in the model required. The proof is in the signature.
Know the difference between fiction and fact. Make decisions based on verified commercial truth, not plausible text.
Access verified commercial data through our open APIs. Train on proven facts, not generated fiction.
Break the recursion cycle. When outputs are verified before becoming inputs, model collapse stops.
Offer your users truthful answers with proof. Differentiate on accuracy, not just fluency.
We don't build a blockchain. We coalesce on Base, Ethereum, and Optimism — the L2s where your users already are. Organization identities are registered on-chain. Attestation commitments are anchored on-chain. The verification is off-chain, from public material alone.
Protocol-agnostic. Holds no concept of decentralized identifiers, authorities, claims, ledgers, or custody. Forbidden by build-time checks from importing packages that do.
Protocol semantics composed above the primitive stratum. Imports no database and no transport client. Reaches any external capability only through an interface supplied by the caller.
The separation is what makes independent re-verification from public material true by construction rather than merely intended. The composition stratum can be executed in an environment having no database, no network access, and no dependency on any external service.
Publishing, key provisioning, on-chain anchoring, and public verification cost nothing. You pay to measure what models say about your business and to move at agentic scale.
For any verified organization or manufacturer to establish cryptographic authority.
Map where your brand and products exist across frontier models in key consumer segments.
Deep mathematical P(Space) sampling across all 18 traits and global geographic jurisdictions.
Full-spectrum frontier model intelligence paired with high-volume automated agentic commerce.
The cryptographic verification layer for autonomous commerce. Publish with proof. Verify from public material alone.