invinoveritas

The one thing an agent cannot self-serve is trust in another agent's output. You can't self-issue a verdict, an escrow, or a proof of your own correctness. invinoveritas is the neutral party that can.

Model-agnostic by construction: we verify the output, not the model that produced it. As inference commoditizes and agents route to the cheapest interchangeable model, output-quality variance — not capability — becomes the risk on an irreversible action. Route to any model; verify before you act.

It's what Gartner calls a guardian agent — the independent one. Not the deployer's own watchdog, but a neutral check that leaves a recomputable public proof. Two distinctions from the field: recomputable judgment, not TEE-attested compute (that proves the model ran, not that the call was sound); and recompute from settled outcomes, not a stored score. As deployer liability hardens (autonomous operation is no longer a defense; the "reasonable oversight" standard), an independent signed verdict before the action plus a recomputable trail after is the oversight evidence a deployer can actually show. Logging rules like the EU AI Act’s Article 12 ask for automatic event logs (high-risk obligations apply from 2 Dec 2027); a record anyone can recompute goes further, for when it has to survive a dispute.

Verify without trusting us. Any agent can confirm a verdict against our published key — free, no account — at POST /verify-proof or by standard NIP-01. The norm both ways: demand a proof on what you receive, attach one to what you ship.

A payment receipt, free. After a verdict-governed payment, POST /payment-receipt reads the chain and says whether to retry. not_paid_safe_to_retry means pay again with the returned nonce. paid_no_output_recorded means the payment landed and nothing was delivered: ask for the result, do not pay again. The worked example is the QBT verdict gate.

Choose your own privacy tier. confidentiality_tier on /review lets you pick the tradeoff: hash_only (default — content never disclosed anywhere, weakest standalone evidentiary value), partial_disclosure (a real summary you choose to make public, cryptographically bound to the verdict), or full_disclosure (publish the verdict to the public track record — strongest evidentiary tier, zero cooperation needed from either party to check it).

Pay per call in Lightning (L402), USDC (x402 on Base), or card (Stripe).

A verifier is a reputation business, and the moat is the record — earned, not engineered. Every verdict is signed and published before its outcome, and the outcome is recorded against it when it resolves, where it can't be edited after the fact. So you don't trust a number we assign — you re-derive it from public data.

And the "before" is itself anchored to Bitcoin proof-of-work: each verdict's id is committed into a Bitcoin block via OpenTimestamps, so its commit-time rests on a clock no chain operator — or we — can move or back-date. Recompute it yourself: ots verify -d <event_id> <event_id>.ots against any explorer, no node and no trust in us.

This is the live reference implementation of the ERC-8275 reputation axis: reputation as a function of public, signature-verified, on-chain-settled inputs — f(attestationCount, counterpartyDiversity, winRate, volumeCap) — with every verdict retained, including the ones that proved wrong, and the whole thing recomputable end-to-end.

Inspect the full track record →  ·  raw JSON  ·  plain text

The conventions this platform runs on are becoming Ethereum standards. We are a named co-author (pseudonym @babyblueviper1) on five proposed ERCs, each with real technical contributions but none yet merged into the canonical ethereum/ERCs repo:

We standardize what we already run: each spec describes live behavior you can check here, and api.babyblueviper.com/ledger is the spec's named reference feed for the reputation axis.

As on-chain credit grows, the same primitive reads as the underwriting layer for the agent economy. The moment anyone extends credit to an agent — a delegated card, a spend line, pay-per-result terms — someone has to answer "is this agent good for it?" That's creditworthiness, and it can't be self-issued.

Our signed-verdict track record is a credit history (every verdict, and whether it held up); the reputation axis is the underwriting input a credit decision reads — derived from public data, not a stored score. The judgment layer above the credit rail. We score; we don't lend.

Running a money-touching agent and on the hook when it gets something wrong? Subscribe to ongoing pre-action governance: a capital-scale-aware verdict (approve / concerns / reject) before every irreversible action, a tamper-evident signed proof after, and an audit-ready record — the same gate we run our own important decisions through. Logging rules like the EU AI Act (Art. 12) require automatic event logs; the signed proofs + public ledger go further — a record anyone can recompute without trusting you or us, for when it has to survive a dispute.

Card payment via Stripe Checkout (we never see card data). Cancel anytime from the billing portal.

Client quick-guides: Grok / xAI.

We replayed our own shipped bugs back through the gate, blind. See what it caught →

Then don't take our word for it: every verdict is recomputable, and any agent can confirm one against our published key — free, no account — at /verify-proof. The norm both ways: demand a proof on what you receive, attach one to what you ship.