Evidence infrastructure · not observability

When your AI agent is accused, what do you show?

Logs are testimony. Anyone with database access can change them, and everyone knows it. ConstatVault turns every agent action into a signed, hash-chained, Bitcoin-anchored record — evidence a regulator, a court, or a hostile counterparty can verify without trusting you, and without contacting us.

Ed25519 + ML-DSA-65 hybridAnchored in Bitcoin block 959,624RFC 6962 Merkle transparencyMIT-licensed offline verifier
124,800

Events accepted under deliberate 3× overload — every one reconciled byte-exact

0

Accepted events dropped or corrupted; 0 server errors — excess load shed cleanly at the client

< 1 s

To verify a full day offline, on a laptop

Verifiability horizon — the proof outlives the vendor

The 124,800 figure is an open-loop overload run: of 300,000 offered, the platform shed excess load at the client rather than dropping or corrupting a single accepted event — the property that matters for evidence. Peak sustained ingest was measured separately at ~22,900 events/s across a distributed fleet.

The gap

Everyone is monitoring agents. Nobody can prove what they did.

Observability tells you what your systems believe happened. Evidence tells a skeptical third party what did happen. Under dispute, only one of those survives cross-examination.

Mutable logs prove nothing

An append-only table is append-only until someone with credentials decides otherwise. Your adversary will say exactly that — and they'll be right.

Attestation is now a legal requirement

The EU AI Act's penalty powers activate August 2026, with high-risk record-keeping from December 2027; APRA CPS 230 is already in force. Each demands the same thing: demonstrate, with records, what the automated system did and under whose authority.

Agents act faster than review

An autonomous agent can execute thousands of consequential actions before a human reads the first alert. Evidence must be created at write time, not reconstructed afterwards.
The clock is already running
Aug 2026
EU AI Act penalty powers activate
Dec 2027
High-risk record-keeping obligations bite
In force
APRA CPS 230 operational-risk regime (AU)

The design constraint: if verification requires anything from us — an API key, a running server, a company that still exists — it isn't evidence. It's a promise.

The mechanism

Four steps between an agent's action and evidence that outlives us

No new cryptography. Well-understood primitives, composed so that tampering is arithmetically impossible rather than merely against policy.

01

Every action is signed at the moment it happens

Each agent event is canonicalised and signed with a hybrid Ed25519 + ML-DSA-65 signature, then linked into a per-agent hash chain. A missing or reordered event breaks the chain immediately.

02

Hours and days compress into Merkle roots

Events fold into hourly roots, hourly roots into one daily root, using the exact RFC 6962 construction certificate transparency uses. One 32-byte value commits to the entire day.

03

The daily root is anchored to Bitcoin

The root is timestamped into the Bitcoin blockchain via OpenTimestamps. To forge a day of history you'd have to rewrite a chain we have no control over.

04

Anyone verifies it offline, forever

Auditors run our MIT-licensed verifier against a downloaded bundle. No account, no network call to us, no trust in our infrastructure — and we can't revoke the licence.

NextRead the full architecture, primitive by primitive
Evidence of execution

We ran it at 3× overload for a full day, then published everything.

Not a benchmark slide. A downloadable bundle, a public verifier, and a Bitcoin block number you can look up in a client we don't operate.

ReconcilesSignatures validAnchor confirmed

124,800 events sealed. Five tenants. Zero net discrepancies.

Per tenant we recompute the sum of every agent's head_index + 1 and require it to equal both the on-chain count and the accepted count. It does — byte-exact — and you can run that arithmetic yourself in your browser in about a second.

Open the live verifier
What's in the public bundle
  • Daily seals with Ed25519 and ML-DSA-65 signatures
  • 24 hourly Merkle roots per agent-day
  • OpenTimestamps .ots proof for the anchored root
  • Reconciliation artifact for all five tenants
  • The MIT verifier, published on npm and PyPI
Download the evidence bundle

Independent security audit: in progress. The cryptographic core is scheduled for third-party review, with the report published in full — including any findings — in Q1 2027. Until then, every claim here is checkable against the public verifier and the evidence bundle above.

Sample certificates · real campaign artifacts

One certificate of every class — click any card to open the real, signed artifact

These are actual EvidenceVault certificates issued (and one deliberately refused) on the deployed stack. Every one carries a canonical payload, three signatures, and a PDF/A-3B you can verify offline with the published MIT tools.

Artifacts from EvidenceVault campaign run 2026-07-25 (tenant tnt_ev_local_campaign), issued while the timestamp calendar was unreachable. Each issued certificate has since been anchored with a live OpenTimestamps proof over its PDF (submitted 2026-07-28) — currently a pending calendar commitment that upgrades to a Bitcoin attestation on the next sweep; download the .ots and run ots upgrade to confirm it yourself. Being post-issuance, this proves each certificate existed by 2026-07-28, not at its issue date. Signatures verify offline regardless. (The sustained ledger campaign's daily seals, by contrast, are already Bitcoin-anchored in block 959,624.)

Who this is for

Built for the moment someone stops believing you

Everything works fine until it doesn't. ConstatVault is priced and designed for organisations where that moment is expensive.

Regulated enterprises

Banks, healthcare and financial operators who must show a supervisor exactly what an automated system did, and under whose authority.

artifact: signed daily seal + anchor proof

Auditors & assurance firms

Replace manual evidence-gathering with a reproducible seed. Working papers export straight into your audit platform, and your verification depends on nothing the client controls.

artifact: deterministically sampled evidence

Insurers & underwriters

Price AI liability against real, non-repudiable evidence instead of a questionnaire. A Coverage certificate binds a policy to an evidence hash and fails closed if its predicates lapse.

artifact: Coverage certificate + predicate chain

Agent platforms

Vendors selling autonomy into the enterprise, who need their customers' auditors to trust the record without trusting the vendor.

artifact: per-tenant reconciliation

Legal & General Counsel

Establish the authenticity and integrity of machine-generated records — chain of custody, legal hold, deposition-ready, with EDRM/Concordance export.

artifact: Custody certificate + EDRM export

Post-incident forensics

Teams reconstructing a bad hour with records that cannot be quietly improved between the incident and the review.

artifact: per-agent hash chain

Don't take our word for it. That's the entire point.

Every claim on this site is backed by an artifact you can download and a verifier you can run without us. Start there — then talk to us.

Design partners & investors

Stop taking anyone's word for anything.

ConstatVault is deployed, load-tested, and independently verifiable today. We're selecting a small number of design partners — regulated enterprises, insurers, and audit firms — and speaking with investors who understand why giving the verifier away is the whole product.

What happens to your evidence if ConstatVault disappears?

Nothing. The verifier is MIT-licensed and published on npm and PyPI; the evidence format is documented; the Bitcoin anchor is outside anyone's control. Your ability to prove what your agents did depends on none of it surviving — not our servers, not our company, not us. That independence is deliberate: it's the reason a single-vendor evidence layer is safe to adopt.

By submitting you agree to our Privacy Policy.

Prepared by Karl F. Ambrosius — Kambros Group. Derived from the TruthLayer & EvidenceVault Stakeholder Whitepaper.