Security model

Built so you don't have to trust anyone, including us

An oversight system is only worth something if it can't be quietly bent: not by the AI agent it watches, not by the business running it, and not by Cosignary. Here's how each of those is ruled out, and how anyone can check.

In plain terms: quality control only counts if nobody can fudge it. So:

  • Rules decide, not an AI, so nobody can talk the system into approving something.
  • Every action goes into a record that can't be quietly edited. Change one entry and it shows.
  • Outside parties hold a fingerprint of that record, never your data, so deleting or rewriting history gets caught. One of them is the Bitcoin blockchain, which no one controls.
  • Which work gets checked is decided by a fair draw locked in ahead of time and mixed with public randomness, so nobody can steer the checks or know them in advance.
  • People sign their approvals, so a decision can't be forged later.
  • Anyone can re-run the whole audit from the record and confirm it was done honestly.

The technical detail is below.

Four principles

1

No AI in the decision path

Rules and code decide; people sign. An AI may flag an action for more review, but it can never clear one, so a fooled or manipulated model can only add work, never hide a problem.

2

Your data stays with you

The gate, ledger, sampler, review page and report run on your own servers or cloud account. Customer data, seeds and verdicts never leave.

3

The cloud is a witness, not a judge

The only outside piece stores fingerprints (hashes and counts) and signs receipts. It holds nothing that decides an outcome.

4

Verifiable by anyone

With the ledger and the witness logs, a third party can recompute the sample, the rates and the planted tests, and confirm nothing was edited or removed.

Where it runs

yours
Your servers or cloudGate, ledger, sampler, redaction, review page and report. Everything that touches data.
out
Fingerprints onlySeed commitments, ledger head hashes and entry counts go to the witnesses. Nothing an outsider could learn from.
witnesses
Several independent witnessesCosignary's own, which signs a receipt you keep, and the Bitcoin blockchain through OpenTimestamps, which no one runs and no one can backdate. A public timestamp authority (RFC 3161) is next.

What verification checks

Each audit period can be verified by anyone holding the ledger and the witness logs. These checks all have to pass:

  • The ledger chain is intact. Every entry is linked to the one before it; editing any past entry breaks the chain from that point on.
  • The witness logs are intact and signed. Each witness's own log is chained, and every entry carries its signature.
  • The ledger still matches every checkpoint. Deleting recent entries, or rebuilding the whole chain, no longer matches what the witnesses signed.
  • The seed matches the commitment made beforehand, and the witnesses recorded that commitment before any action in the period.
  • The Bitcoin proofs check out. Each anchored fingerprint leads, by a short chain of hashes, to a real Bitcoin block. The proof is a standard OpenTimestamps file, so it can be checked with no Cosignary code at all.
  • The beacon value is the public one. When the window closed, it named a round of the drand public randomness beacon that didn't exist yet. Verification fetches that round independently and confirms it's the value the sample used.
  • The sample recomputes exactly. Selection is HMAC(seed, action fingerprint) < audit rate; a cherry-picked or altered sample doesn't match.
  • Planted tests, rates and signatures check out. The planted-card key matches its commitment, audit rates follow the published rule, and every human decision is signed by a registered signer before its deadline.

Threat model

AttackWhat stops it
The agent behaves well only on actions that get checkedThe sample is picked by a secret seed committed in advance, from fingerprints that don't exist until the actions do; it can't be predicted
Someone edits a past actionThe hash chain breaks at that entry
Someone deletes an action, or rebuilds the whole chainThe ledger no longer matches the checkpoints the outside witnesses signed
The business cherry-picks a clean sampleThe sample is recomputed from the revealed seed; any difference fails verification
A convenient seed is chosen after the factThe seed's fingerprint was committed and witnessed before the period began
The business uses its own seed to foresee the sample and steer around itThe sample also depends on a public beacon round that wasn't published until after the window closed. Nobody, the business included, knows the sample until then
Everyone involved agrees to backdate the recordFingerprints anchored in Bitcoin can't be backdated by anyone, Cosignary included
Audit rates are quietly loweredThe policy is snapshotted at the start of the period and rates follow a published rule anyone can recompute
Many small actions slip past a per-action capSpeed limits count across a time window and trip a circuit breaker that pauses the category
A held action is approved by defaultNo answer by the deadline means deny, never approve
A decision is forged or altered laterEvery approval, denial and breaker reset is signed by a registered signer; flipping one breaks its signature
Reviewers rubber-stampPlanted test cards, indistinguishable from real ones, measure every reviewer's catch rate
Reviewers are pressured, or colludeAn outside reviewer re-checks a seed-picked share of cards; the stricter verdict counts and agreement is reported
The agent's own summary misleads a reviewerFacts from the system of record are shown first; any mismatch with the agent's claims becomes a warning on the card
Text in an action steers an AI flaggerAI can only add scrutiny, never clear an action; the random sample runs regardless
Someone breaks into Cosignary's cloudWitnesses hold no data, seeds or decision power; their logs are chained and signed; several independent witnesses means no single point of failure

Separation of duties

  • The approver isn't the actor. Signers are registered people with their own keys (passkeys in production); the agent can't sign for itself.
  • The outside reviewer isn't the internal reviewer. The same person can't give both verdicts on a card.
  • The flagging model isn't the acting model. Any AI used to flag actions must come from a different vendor than the agent it watches.
  • The witness isn't the business. Checkpoints are held by independent parties.

Privacy and redaction

Review cards mask emails, phone numbers, card numbers and Social Security numbers automatically, plus any fields or patterns the policy names (customer names, addresses, order numbers). Reviewers judge the action ("a $400 refund on a $40 order"), not the customer. The complete record stays in the ledger inside your environment, and outside reviewers, if you use them, work through your own secure access.

Frequently asked questions

What is a verifiable random audit?

A spot check that nobody can rig, and that anyone can confirm afterwards. Before any work happens, a secret random number is locked in by publishing its fingerprint. That number decides which actions get checked, so neither the AI nor the business can choose them. After the review, the number is revealed, and anyone can redo the draw to confirm the right actions were checked. Election audits use the same method.

What if someone breaks into Cosignary's cloud?

They gain nothing that changes an outcome. Witnesses hold no data, no seeds and no decision power; their logs are chained and signed; and you can anchor with several independent witnesses, including the Bitcoin blockchain, which no one runs.

Can anyone delete or rewrite the audit log?

Not without it showing. Editing an entry breaks the hash chain, and deleting entries or rebuilding the chain no longer matches the checkpoints the outside witnesses signed. Entries made since the most recent checkpoint are the exposure window, so checkpoints run frequently.

Does Cosignary use blockchain or crypto?

Only as a public notary and a public dice roll, and never as a token. Ledger fingerprints can be anchored in Bitcoin through OpenTimestamps, so no one can backdate the record. The audit sample mixes in the drand beacon, a random value published every few seconds by a group of independent organizations, so no one can foresee it. Your data never goes on a blockchain: only 32-byte fingerprints leave your systems, and there's no Cosignary coin.

Who can see our customer data?

Only people inside your environment. Review cards are redacted, and nothing but fingerprints leaves your systems. If you add outside reviewers, they log in through your own secure access to redacted cards.

Is the code open source?

The trust core (ledger, sampler, verifier, gate and the policy format) is planned to be open source under Apache-2.0 after an independent security review, so anyone can run the verifier themselves.

Test it yourself

Try five ways to cheat the audit

Edit a past refund, rebuild the chain, delete an action, cherry-pick the sample or swap the seed, and watch which check catches each one.