At a glance
- Records are Ed25519-signed and hash-chained, and anyone with the public key can verify them offline. You never have to trust us for the truth of a record.
- We never hold your signing keys. The hosted service stores only the public keys you pin, and a pin can never be replaced.
- Ten rounds of structured adversarial review; 124 findings, each fixed or published as a stated limit. No SOC 2 or third-party penetration test yet.
- Found a vulnerability? Email hans@decisionhorizon.dev with "Security report" in the subject.
The trust model
A record is tamper-evident: an alteration is detected, not made impossible. Here is what each kind of lie would take, and what the record states about itself.
| To… | An attacker needs | What's stated |
|---|---|---|
| Rewrite history | The signing key and k of n witness keys. With pre-action attestation, in real time. | With no witness, the signing key alone could rewrite a chain, and every record says so. |
| Write a false record | The signing key, in real time. Witnesses countersign the chain, never its content. | How the key was held, plaintext included, is in every record. |
| Fake an execution | The agent runner's key, when runner receipts are used at the gateway. | The runner is trusted for execution facts. |
| Hide an action | A path around the insertion point. | Every record names its insertion point and lists its known blind spots. |
| Blur what the agent used | Nothing, for use by value: it's decided from the bytes. | "Read and reasoned about without copying" is an estimate, labelled as one. |
| Make a hosted copy lie | Nothing the verifier accepts. | The operator can withhold, delete or roll back what it stores. |
The trust model post explains each row, and the docs list the stated limits in full.
Keys
- Your signing key stays with you.
dh keygencreates it on your machine. It can live in a separate signer process over a Unix socket, so the agent process never holds it, and it can be encrypted at rest with a passphrase. - The public key is pinned, never read from the records. The verifier only accepts records signed by a key you pinned. A record that carries its own key proves nothing.
- Witnesses are separate processes with their own keys. A quorum of them can be required, including before an action runs.
The hosted service
The hosted record store at api.decisionhorizon.dev and the Console are built so you don't have to trust them for the truth of your records:
- It stores only public keys. Pins are permanent; a different key under an existing id is refused.
- Every uploaded record is verified under your pinned key and linked to the stored chain before anything is written. Uploads are all or nothing.
- Records are append-only and returned byte for byte as received, so you can always download a chain and run
dh verifyyourself. - Verification uses the same code as the offline verifier. We test the hosted answer against a conformance pack of clean and tampered chains generated from the offline tool.
- Tenant API keys are stored only as hashes and compared in constant time. Every query is scoped to your tenant, and we test that one tenant can't read another's data.
- Every key pin, upload, verification and legal hold is written to a hash-chained audit log you can download.
- Traffic is served over HTTPS. The API accepts browser requests only from our own origins, limits request size, and rate-limits each key.
- The Console sets no cookies, loads no third-party scripts besides web fonts, and keeps your API key only in the browser tab's session storage.
How it's been tested
Before launch, the design went through ten rounds of structured adversarial review, with 124 findings. Each was closed by a fix or published as a stated limit. The verifier's detections each have a test: reordering, a forged record under the right key id, tail truncation, a dropped closing record, duplicates, swapped hash pointers, a tampered body, spliced chains, a record after the close, and an action with no outcome.
What we don't have yet: SOC 2, and a third-party penetration test. The reviews were structured adversarial reviews, not an accredited audit.
Reporting a vulnerability
If you think you've found a security issue in the packages, the verifier, the Console or the hosted API, please email hans@decisionhorizon.dev with "Security report" in the subject. Include what you found, how to reproduce it, and what impact you think it has.
- We'll acknowledge your report within three working days and keep you updated as we investigate.
- Please give us reasonable time to fix the issue before you disclose it publicly. We're happy to credit you when we publish the fix, if you'd like.
- When testing, use only your own data and tenants. Don't access other customers' data, degrade the service, or use social engineering. Good-faith research that follows these rules is welcome, and we won't pursue it.
Our security.txt has the same contact details in machine-readable form.