How We Know Our Claims Are True

/ 01 The problem

Marketing copy drifts from code truth. Always.

Nobody sets out to lie. A pricing page promises a perk before the feature ships. A directory headline quotes the size of a raw database table instead of the vetted subset actually served. A claim survives one copy review by moving to a different page, or by dropping two words that the reviewer searched for. Copy and code update through different channels, so they diverge — on every software team, everywhere.

We run four production platforms for Canadian professionals — accountants, lawyers, financial advisors, and business sellers. In July 2026 we stopped treating that drift as a copywriting problem and started treating it as an engineering problem.

/ 02 What the sweep found — and what we did about it

Every claim traced to its source of truth.

49 findings adjudicated individually. Every public claim on the estate was reviewed against the code and the production database. Each one was either evidenced, corrected, or removed — with the ruling recorded.
Published counts re-based to the vetted data, not the raw table. One platform advertised "225,000+" professionals; the honest, quality-filtered number was 143,391, so the page now says 140,000+. Another advertised "56,000+"; the raw count included 36,103 out-of-scope licensees, so the page now says 18,000+. The third platform's "100,000+" survived the check at 100,019 — a 19-row margin we logged rather than ignored.
A revenue calculator was retired outright. A calculator on a pricing page is a claim in math's clothing, and this one projected returns from a capacity that had never once been reached in production. It failed the standard on every prong, so it was removed — along with 131 links that pointed to it — and the page generator was patched so a future rebuild cannot re-emit it.
"Verified" now means a human verified it. Automatic status stamps were renamed to what they actually are — a registry check — and the word "verified" is reserved exclusively for the manual review a person performs.
Promise and behavior ship in the same deploy. Member-facing copy promised a paid perk that the code was delivering to everyone. The copy fix and the entitlement gate landed in one deploy, so the promise and the behavior cannot diverge in either direction.
Fabricated social proof deleted. Placeholder testimonials and a script that simulated "live" activity numbers were removed entirely. No replacement until real ones exist.

/ 03 The machinery

Honesty that survives the next deploy is engineered, not resolved.

A copy review fixes today's pages. The gates below run in continuous integration on every push, so the corrected claims cannot quietly return:

Copy-ban gate

Every phrase retired in the adjudication — unsupported guarantees, unenforced capacity caps, feature claims with no feature behind them — is a banned pattern. A pull request that reintroduces one fails CI with the ruling named in the failure message.

Count-claims test

Backend email templates are tested against the stale raw-count family. The test names the honest source number in its failure message — and it caught a second stale count before merge during the sweep itself.

Design-integrity and accessibility gates

Raw color values outside the design-token file fail the build, and six key pages must hold a Lighthouse accessibility score of 90 or better on every push.

Deploy-base guard

Every deploy is refused unless the code being shipped contains the current production history — born from a real incident, described below.

/ 04 The incident we disclose

The guard exists because we shipped a mistake.

During the same period, two deploys went out from stale local code and silently reverted already-shipped work on two platforms for roughly a day. We caught it through our own review discipline, disclosed it in the operations record, remediated it the same session, and encoded the fix: a pre-deploy guard that refuses any deploy whose code does not contain the production history, printing the exact divergence when it refuses. It has run on every deploy since.

That is the pattern behind everything on this page. The operating rule is written down: when the same class of mistake is corrected twice, it is not corrected a third time by hand — it is converted into a mechanism the system enforces, recorded in a ledger, and never silently deleted. Retired rules stay in the ledger, marked retired, with the reason.

"Prefer impossible over caught over complained-about. A constraint that makes the mistake unwritable beats a test that catches it, which beats an assertion that grumbles after the fact." Backpressure protocol · internal operations standard

/ 05 Why this matters to you

This is the discipline we install.

AI makes it cheap to generate claims, content, and code — which makes it cheap to generate wrong claims, content, and code. The competitive question is no longer whether your organisation can produce output; it is whether your organisation can prove its output is true. Registry-checked data, human-gated review of generated content, disclosure of how rankings weight paid tiers, consent-first outreach design: these are the same mechanisms documented above, and they are what a Phronisi engagement builds inside your operation.

We can show you this system live — the gates, the ledger, the failure messages — because we run it on our own platforms first.

Next step

See the system, then decide.

A 60-minute Diagnostic Call: your context, a live look at the machinery on our own estate, and a clear recommendation on scoping and sequencing — whether or not it leads to an engagement.

Request a Diagnostic →

Provenance: everything on this page is drawn from our internal operations record and backpressure ledger, July 2026. Dates are real, numbers are quoted from the record, and the incident in section 04 is disclosed, not curated. No client engagements or client data are referenced. The platforms are CanadaAccountants.app, CanadaLawyers.app, CanadaInvesting.app, and Canada Business Exits.