Methodology and data

Check a certificate yourself

Certificates AInsure issues are signed with Ed25519 at the moment they are created, and the public key is published. The check below runs in your own browser: nothing is sent to AInsure. Press “Change one character” and watch it fail.

It is a sample, signed with a demo key made for this page, so that no customer’s certificate is published. Real certificates verify the same way, against the key at /api/certificates/signing-key. Wallet proofs from The Vault have their own key at /api/vault/proof/key. How the whole thing works is on How proof works.

Automated checks run on every change

From the continuous-integration run that cleared the 30 Sep 2026 release (commit e2afdf45). The run log is private; these are the summary lines from it.

CORA is not in the total. Its continuous-integration run executes only the test files that need no database or other service, plus the routing benchmark below. The rest of its tests run on developers machines before a release and are not counted here.

CORA’s routing benchmark

Whether an insurance question is sent to the insurance specialist inside CORA, on a fixed set of test questions. On 75 test questions, 73 were routed correctly (97.3%). The release is blocked if accuracy falls below the floor. The floor is 97%.

This scores the router on its own. It is not an end-to-end result: the benchmark records that an earlier end-to-end measurement of the whole live graph routed 90% of the same kind of questions correctly.

JSON CSV

What a CORA answer costs, and how long it takes

What it cost, and how long it took, for CORA to answer a fixed set of test questions, grouped by the kind of question. Measured 11 Sep 2026 on 120 test questions. Across all of them an answer cost about $0.09 on average.

What this is not. A comparison. These are AInsure's own costs. They say nothing about what any other approach would cost, and they are not a claim about savings.

How it was measured

• The real CORA graph was run end to end on a fixed set of test questions (not customer questions).
• Cost is the model provider's list price applied to the tokens each answer used, summed over every model call in that answer.
• Reply time is wall-clock from question to finished answer, with CORA's tools switched on.
• One agency's test data, run one question at a time on one day.
• Measured before prompt caching shipped (2026-09-20), so today's cost per answer may differ. It has not been re-measured.

The rows add up to 122, two more than the 120 test questions. We have not reconciled that difference, so the per-row figures are the data and the average below is approximate.

JSON CSV

What we have not measured

Numbers you might reasonably want, that we do not have, and so do not publish.

• A saving against any other approach. We have no like-for-like comparison on record, so there is no “percent cheaper” here, and none anywhere on this site.
• How accurate answers are, end to end. The routing benchmark above scores one step. Our last measurements of whole answers predate fixes made since, so we are not quoting them.
• How often people rate an answer helpful. The ratings are recorded; the last count is from mid-September and mostly from one agency, so it is not quoted.
• Uptime. We keep internal health checks, and have no public status page or published figure yet.

All the files

• CORA cost and speed by question type

  Measured 11 Sep 2026

  JSON CSV

• Automated checks run on every change

  Measured 30 Sep 2026

  JSON CSV

• Public keys

  Certificate signing key · Vault proof key