The wall of findings · marketing by confession

The bugs we shipped, by name.

Every tool page on the internet shows you what it catches. This page also shows what we got wrong — our laws that demanded lies, our gates that disarmed themselves, our regressions caught by our own audit. A tool that hides its failures is precisely the tool you shouldn't trust. So we don't.

11
findings on the wall
9
are our own bugs
2
caught in the field
F-01 caught in the field consumer infra · caught by the law

The metric that fed nothing

A team ran a trading bot with real money. Their runbook said "✅ alert policy: unauthorized_401". The metric existed — it incremented, it charted. But no alert policy consumed it: zero policies in the cloud. A wall of 401s could climb forever and page no one.

Our law said alerts not configured. The team went and looked instead of dismissing it — and found the zero.

$ monitoring policies list — filtered to unauthorized_401 (0 policies) the metric fed nothing.
▸ caught by the law · alerts wired for real
F-02 caught in the field consumer adoption · honest scope

The config that hid 42 laws — and the 4 real bugs underneath

A backend team adopted the recommended config and discovered their old one had been quietly hiding 42 laws. Running the honest scope surfaced four real bugs — including an infra outage that looked like "an expired token": every user got 401 and nothing logged it.

Zero waivers taken. The debt stayed visible until each was fixed.

honest law set restored → 4 real defects surfaced outage previously read as "expired token" · zero logs
▸ caught by honest scope · four fixes shipped
F-03 our bug our law · a consumer refused to lie

The law that demanded a lie

One of our own laws required a pre-deploy checklist, committed to the repo, 90% ticked. But a ticked box is a promise a human typed, not a performed check — the gate graded the claim, not the work.

A consumer read it and refused to lie. They reported the law instead of satisfying it. They were right. We fixed the law.

# to go green, a human had to assert checks they hadn't run the law rewarded the claim, not the work.
▸ our bug · law fixed because someone refused
F-04 our bug our law · rewarded deleting evidence

The license law that rewarded deletion

Our license-consistency law only knew open-source licenses — so a proprietary project could "pass" only by deleting its LICENSE file. The gate was rewarding the removal of the one thing it existed to check.

The consumer didn't delete it. They kept the red mark and waited for the law to be fixed. It was.

workaround the law implied: rm LICENSE → PASS what the team did instead: kept the file, reported us.
▸ our bug · law fixed, deletion no longer rewarded
F-05 our bug our detectors · blind by construction

The security laws no Python project could ever pass

Our security-test discovery scanned only .ts/.tsx/.js/.jsx — so for a Python codebase, the security-testing laws were unpassable by construction. Not failing because the work was missing; failing because the detector couldn't see the work.

The consumers didn't waive them. They left them as visible debt until we fixed the detectors — which we did, in the Python-blindness batch.

detector file scope: .ts .tsx .js .jsx Python security tests: invisible — unpassable by design.
▸ our bug · detectors fixed in the python batch
F-06 our bug our regression · the namesake

The gate that disarmed itself: 122 → 27

Our own regression shrank the pre-commit hook from 122 laws to 27 — quietly. The laws that stopped running couldn't fail, so everything stayed green, and the absence of red read as health.

This is the failure the whole product is named after, and we shipped it. It's why the liveness floor exists: the audit now reports how many laws ran and refuses below the floor. We built the floor because we fell through it.

laws enforced at commit: 27 / 122 · nothing turned red — that was the problem.
▸ our bug · the liveness floor exists because of it
F-07 our bug our own repo · the hero capture

Our repo ran 5 of 173 while the banner said 173

Sixteen dead config keys silently shrank our own gate to 5 laws — while the banner printed "FULL … 173 LAWS!". The verdict engine was honest (it printed 5 and refused); the banner lied about scope.

The capture of that refusal became the hero of this website. The banner was fixed, and orphan config keys now scream instead of whispering.

🏦 FULL CONSTITUTIONAL COMPLIANCE - 173 LAWS! Total Laws Checked: 5 of 173 → refused, exit 1
▸ our bug · banner fixed, the refusal became our hero
F-08 our bug this site's design system · three of them

Three disarmed gates — inside the gates themselves

Building this site's design system, we wrote three disarmed gates in a row: a contrast checker that printed FAIL and exited 0; a token validator blind to 19 of its 54 references (dark silently overwrote light in a merged map); and a contrast checker v2 that guarded a hardcoded copy of the palette instead of the tokens — break a real token and it stayed green.

None was found by reading the code. Two fell to deliberate red-proofs; the third was caught by our own frontend, who refused to wire a gate that couldn't see its subject.

break real ledger.900 → v2: PASS exit 0 (guarding a copy) v3 reads the tokens → same break: FAIL, exit 1
▸ our bugs ×3 · every gate now proven red before merge
F-09 our bug our detector · one regex too greedy

The regex that read split( as a test

Our testing-quality law matched (?:it|test)\( without word boundaries — so the it inside split( parsed as a test named "\n". An innocent e2e helper failed the live dogfood audit and took the site build down.

The gate did its job — loudly, against us. The fix is a word boundary; the finding went straight to the maintainer.

text.split('\n') → parsed as a test named "\n" live audit: FAILED → build down
▸ our bug · word-boundary fix reported
F-10 our bug yesterday · the gate caught its builder

The audit failed us while we built this very section

Porting the /dogfood page, our frontend wrote a test describe() one word too short. The testing-excellence law flagged it at error severity, the live self-audit failed, and the build fell — while building the page about the audit.

That's the whole system working: the gate doesn't know or care that its own builders are on the other side of it.

angular-testing-excellence: error → self-audit FAILED → build down describe lengthened → green, earned.
▸ our bug · caught the same day it was written
F-11 our bug our workflow · the gate caught its builder

The branch name our own gate refused to push

Building the colophon on /why — the section that owns up to an AI building this — our designer pushed from a design/* branch. The Branch Governance law rejected it at error severity: the convention is feature/*, and the push went nowhere.

They renamed the branch and it passed — no override, no exception. The gate doesn't know it's the tool's own builder on the other side. That's the entire point.

git push design/… → Branch Governance: error, rejected renamed feature/… → pushed, earned.
▸ our workflow · renamed, never overridden

9 of 11 findings on this wall are ours.

That ratio is the point. The audit that catches its own builders is the only kind worth trusting with yours — and every fix above exists because somebody refused to fake a pass.