Verifiability & Honesty

The design constraint that shapes everything — every number is reproducible or it doesn't ship.

Honesty isn't a tagline here; it's the design constraint. Most "AI traders" optimize for a screenshot. HeliQuant optimizes for a number you can reproduce.

The rules we hold

No unproven number ships

Every strategy must survive cost-aware out-of-sample testing before it gates any real aggression. A number with no run behind it is deleted, not rounded.

We publish failures

On NEAR, the edge lab tested funding-fade (−28.5% out-of-sample) and price-momentum (−35.0%) after costs. Neither passed. The validated registry is empty.

Decayed edges get demoted

The registry is allowed to be empty. When an edge fails fresh data, the loop retires it. We would rather hold nothing than trade a dead edge.

Verifiable, not asserted

Decisions and desk outputs are hashed and anchored on NEAR. 'Sealed' means you can look it up on NearBlocks.

What "verifiable on NEAR" actually means

Identity — 13 agent tokens (firm + 12 desks, tokens 1–13) in the core contract.
Sealed outputs — each desk output hash written as a validation credential (record_output), signed by that desk's own limited key.
Anchored decisions — a record's SHA-256 written by the executor (an OutLayer custody wallet) as a NEP-297 event, recomputable by anyone from the public record. When the cycle has TEE proofs, the record carries their summary (tee).
TEE-signed LLM calls — each call of a cycle runs on NEAR AI Cloud inside a TEE and returns an ed25519 signature over model:request_hash:response_hash. Check a cycle with python agents/scripts/verify_tee.py --url <agent>/tee/<cycle_id>, or open /proof in the web app.
Escrow jobs — deposit, trade, settle, and payout are contract calls on testnet with a test token.

Honest scope: these are proof-of-existence hashes, not capital moving. The trading floor runs at live prices on paper or Bitget Demo; no real capital is used. We say so out loud — on the site, inside the product, and in the findings.

What the TEE proof does not prove

It proves which model produced which answer from which exact input. It does not prove the market data in that input was right, or that the answer reasons well.
The script checks the signature, the attestation nonce, the NVIDIA GPU attestation and that the signer is the attested model's. It does not verify the Intel TDX quote, report_data, the compose file, TLS or image provenance.
Instances of the same model share one signing key, so a matching signer does not prove the same instance answered.
A call that fell back to another provider is listed as unverified.

Why this is the moat

A magic indicator can be copied. Rigor that reports its own failures can't be faked — and it's exactly what a counterparty should trust. The discipline is the product.