The question is never
whether you monitor.
It is “show me the record for the last twelve months.” Scanning is the easy part — the product is a dated, verifiable record that survives being questioned.
Q3 2026 evidence pack
readychanges 41 · decisions 38 · notices 3
normaliser 1.3.0
ledger head 0f31c8ad9e42
entries 1,204
Every row verifiable against the public ledger, by anyone, without an account.
Assembling evidence costs more than gathering it.
The data usually exists somewhere. Proving it was true on a particular date is the expensive part.
Evidence nobody can verify
A PNG of a vendor page proves you looked. It does not prove when, or that it has not been edited since.
The same forty questions
Answered from memory each time, by whoever is free, with no shared source of truth behind the answers.
“What did you do about it?”
Detection without a recorded decision is half an answer. Auditors ask for the other half.
A record built to be questioned.
- Hash-chained from the moment it happens
Each row carries the hash of the row before it. An entry cannot be altered or back-dated without breaking every row after it.
- Decision log, not just detection
Accepted, objected, escalated or marked not applicable — with who decided and when, alongside the evidence they saw.
- Packs on a date range, not a dashboard screenshot
PDF, CSV and a manifest. Hand it over as-is; the auditor can verify it without asking us anything.
- Fits what you already run
Webhooks, an API, Slack and Teams. No agent, no browser extension, nothing to install on an endpoint.
A noisy detector gets muted, and then it protects nobody.
So we tuned for silence before sensitivity, and we publish the numbers.
- 0.6% false positives
Measured by mutating twenty real vendor pages with eight kinds of harmless edit. One in 160 produced a published change.
- Anything touching 40% of a list is held
A page rebuild is not eighteen new sub-processors, and publishing it as one would be the most damaging thing we could do.
- No model in the decision
Fetch, normalise, hash, compare. Models may suggest a page or draft a summary; none can create or suppress a record.