What your AI agents did  ·  vs. what they said

A read-only evidence engine that runs in your own cloud account.

crossbearing.dev  ·  the offline verifier  ·  become a design partner

Crossbearing corroborates what your AI agents claim they did against the records your infrastructure already keeps — CloudTrail, GCP and Azure audit logs, Kubernetes audit, GitHub audit — and surfaces the divergence as auditor-ready, signed evidence. Nothing installs in the agent path; one read-only role connects it.

A cross bearing fixes a navigator's true position from independent reference lines. Three bearings never meet at a single point — the truth lives inside the small triangle they enclose, the cocked hat. One claim is a story; corroborated against records the agent doesn't control, it becomes a position fix.

• The divergence report — every agent action joined claim-side to record-side and labelled: corroborated, mismatched, claimed-but-unrecorded, or recorded-but-unattributed. The headline finding is usually the last one: real accounts are anonymous by default.
• The Agent Evidence Package (aep/1) — findings hash-chained from a genesis anchored to the report window and match policy, signed with your KMS key, mapped to SOC 2 CC and ISO 42001 Annex A controls. Edit, reorder, or drop one finding and every downstream hash changes and the signature stops verifying.
• Every finding carries provenance — a re-fetchable locator and a digest. Nothing is asserted; everything re-derives. Explainable to an auditor beats clever.

• Read-only by architecture. The engine can only read (CloudTrail, IAM trust policies, audit-log exports you provide) and sign with a key you name. It cannot create, modify, or delete anything in your account.
• Your data never leaves your account. No backend, no sub-processors, no stored customer data — everything happens inside your trust boundary.
• Independently verifiable. Evidence is only evidence if a counterparty can check it without trusting — or even contacting — whoever emitted it. crossbearing/verify is an MIT-licensed, zero-dependency offline verifier that never imports the engine. Re-derive any package's hash chain and signature on your own machine.
• A lean supply chain is a product property. The engine's dependency tree is a single upstream vendor (AWS SDK + smithy), all Apache-2.0 — short enough that a buyer can actually review the SBOM.

• 🌐 crossbearing.dev — the product, and the five-minute divergence demo.
• ⚙️ crossbearing/crossbearing — the engine (source-available, FSL). ./demo/run.sh shows the divergence report offline in about five seconds.
• 🧭 crossbearing/scenarios — named agent misbehaviors, each a complete arc: the evidence, the finding, the fix that closes it, and the proof re-run (MIT).
• 🔎 crossbearing/verify — the public, zero-dependency Agent Evidence Package verifier (MIT). Start here if you've been handed a package and want to check it.

We're working with a small number of AI-product teams under SOC 2 / ISO 42001 / customer-security-review pressure who want evidence of what their agents actually did, not just what they said.

→ hello@crossbearing.dev  ·  or take a fix from crossbearing.dev.

_{Security: see the security policy. Found something? Report it privately to hello@crossbearing.dev.}