Free tool Tier a model against OSFI E-23 in about five minutes — seven risk dimensions, the controls your tier requires, and a report you can print. Open it →
osfi-gate

Verify it yourself

Most products in this space ask to be believed. This one can be checked, and the check is the demonstration.

Everything below is real output from the demo estate — 65 runs for one repository — not an illustration of what it would look like.

The command

osfi-gate verify --evidence-dir ./evidence --repo owner/name \
    --policy-dir ./policy --pubkey ./key.pub

Offline. Nothing of ours runs, and nothing phones home.

A clean run

northwind__adjudication-service.jsonl: 65 entries, chain OK, head be34a161650392c4
  ? anchor NOT CHECKED (no anchor store was given): the chain and the
    signatures are checks this evidence makes about itself
  OK   000065-1b9bf83f9cfd
       - signature verified (key) with key 709d12adaafc185e
       - decision replayed with policy 453b48494395: fail
       - all 1 raw scanner findings present in gate input

Four separate statements, and it is worth reading them separately.

  1. The chain holds. 65 ledger entries, each linking to the one before it, ending at a head you can write down.
  2. The signature is good, and it says which key answered. After a rotation one ledger holds bundles signed by several keys, and "the key this bundle names" is a stronger statement than "some key we hold".
  3. The decision was replayed. The recorded inputs were re-run through the recorded policy — identified by digest, 453b48494395 — and reached the recorded result. Note that the result is fail. Verification confirms the record is intact; it does not care whether the answer was convenient.
  4. No finding went missing between the scanner and the policy. Every raw finding is present in what was evaluated, so nothing was dropped on the way in.

And one thing it refuses to claim. The second line reports the anchor as not checked rather than as passing, because no anchor store was given. That distinction is the point of this page: the chain and the signatures are checks this evidence makes about itself, and somebody who could rewrite the whole branch and re-sign it would produce a history that passes all of them. An anchor is an observation of the head recorded somewhere its owner cannot rewrite. Pass --anchor and it is checked; omit it and you are told you omitted it.

A tampered run

The same evidence, with one file edited — decision.json changed from fail with four violations to pass with none. Exactly the edit somebody would want to make.

  OK   000064-1b9bf83f9cfd
       - signature verified (key) with key 709d12adaafc185e
       - decision replayed with policy 453b48494395: fail
       - all 1 raw scanner findings present in gate input
  FAIL 000065-1b9bf83f9cfd
       ! decision.json: content changed
       - signature verified (key) with key 709d12adaafc185e

Exit code 1. The run before it still passes, which is what you want: the failure is located, not smeared across the history.

Look at what still succeeded. The signature verified on the tampered bundle. That is not a gap — only statement.json is signed, and it was not touched. The manifest is what carries a SHA-256 for every file in the bundle, and re-hashing caught the change. The layers are independent on purpose, and a report that collapsed them into one green tick would tell you less.

What it proves

What it does not

It does not prove the decision was right. If the policy was wrong, you get a perfectly verifiable record of a wrong decision. That is why the policy is versioned, reviewed, and identified by commit in every bundle — so the decision and the rule that produced it can be argued about separately.

A hash chain proves internal consistency, not authenticity. Somebody who can rewrite the whole evidence branch and re-sign it produces a history that verifies perfectly cleanly. This is the reason anchoring exists, the reason the evidence branch should be protected, and the reason the tool reports an unchecked anchor as unchecked. The audit trail documentation says so first, before it describes anything that works.

Download it and check for yourself

Four files. Nothing to install beyond osfi-gate itself, and nothing here talks to us.

File What it is
evidence.tar.gz (14 MB) One repository's ledger and all 65 bundles, as the gate wrote them
evidence-tampered.tar.gz (14 MB) The same, with run 65's decision edited from fail to pass
policy.tar.gz The policy checkout. Replay needs the policy the decision was made under, so it travels with the evidence
northwind.pub The public half of the signing key

Checksums are in SHA256SUMS, so you can check the download before you trust it — which is the whole point.

tar -xzf evidence.tar.gz && tar -xzf policy.tar.gz

osfi-gate verify --evidence-dir ./evidence \
    --repo northwind/adjudication-service \
    --policy-dir ./policy --pubkey ./northwind.pub

Then run it again against evidence-tampered.tar.gz and watch it fail. A verification that only ever passes teaches nobody anything.

Or without osfi-gate

The same checks, short of replaying the decisions, with nothing of ours installed: two small verifiers written from the published evidence format, one in Python using only its standard library and openssl, one in JavaScript using only Node. They are short enough to read before you run them, and they are licensed Apache-2.0 so you can keep them whatever becomes of this project.

File What it is
osfi_verify.py Python 3.8 or later. Also re-draws an audit sample
osfi-verify.mjs Node 18 or later. The same checks, written separately
evidence-format.md The format both check, written so it can be implemented a third time
LICENSE.txt Apache-2.0
python3 osfi_verify.py verify --evidence-dir ./evidence --pubkey ./northwind.pub
node osfi-verify.mjs verify --evidence-dir ./evidence --pubkey ./northwind.pub

Both fail the tampered set, on the decision file whose hash no longer matches its manifest. Neither re-runs the policy over the recorded input — that needs the policy engine — and both say so every time they run. Their checksums are in verifier/SHA256SUMS.

These files verify as published. The script that produces them runs the command above against both sets and refuses to publish anything unless the clean one exits 0 and the tampered one does not. If the published evidence did not verify, that would be a finding about the product rather than about the download, and it should not be possible to ship one by accident.

The estate is fictional — Northwind Trust does not exist, and neither do its people or repositories. What is real is everything the gate did with it: the scans, the policy decisions, the hashes, the chain and the signatures.