The has-website quality gate

The asymmetry is the spec

A false RED skips one weak site out of hundreds. A false GREEN mails "I built you a better website" to someone whose site is fine — a credibility hit in a market where owners talk to each other. Every weight below leans one way: when uncertain, protect the prospect from the pitch. The machine judges neglect and era, which are objective; it deliberately refuses to judge beauty. A dated-but-working site walks free, and that is correct.

The failure classes

Verdict is mechanical: findings carry weights, WEAK at score ≥ 10. Forcing classes end the argument on their own; damning classes nearly so; corroborators must gang up. The one deliberate exception on absence signals: a local service site with no phone at all may cross on absences alone — a business that cannot be called is weak by definition.

The matrix, honestly

Nineteen synthetic fixtures, ground truth fixed before the first run. Three calibration iterations — every one a mechanism bug, not a weight change; the weights survived their corpus untouched (a corpus that shares authorship with the detectors, which is why five public pages were also run — see the wall notes).

Precision on WEAK 1.000 (12/12) · recall 0.923 (12/13) · false GREENs: zero — the cell that matters is the DECENT-row zero. The one miss is weak-borderline, a fixture designed to hover at the threshold (score 8 of 10): the accepted false RED, one prospect traded for no credibility risk. It is in the gallery below wearing its miss.

The gallery — every fixture, its verdict, no retouching

All businesses invented, all pages synthetic, photographed at 390 × 844 — the four-second view. Scores out of a threshold of 10.

Provably weak — the pitch goes out

Twelfth of the twelve: redirect-social — the domain 302s to a Facebook page. No screenshot exists because none was taken: the gate classified it from the redirect header alone and fetched nothing. WEAK · forced.

The hard middle — designed to test the seam

Fine — the people the gate protects

And the two nobody could render: unreachable-refused (dead server) and unreachable-hang (answered in 28s of silence — verdict inside the 30s budget). UNREACHABLE is a verdict, not a crash, and it is never judged weak.

Wall notes — what the honest record holds

• The wall had to be made mechanical. In iteration 1 a fixture's 302 sailed through the route guard to m.facebook.com — Playwright does not re-intercept server redirect hops — and the matrix still read green, because the stripped page scored WEAK by junk findings. One off-origin document was fetched, in-Lab, to a fictional URL. The fix classifies redirects from the Location header with zero bytes fetched. A green matrix must be read row by row.
• Transport states are simulated on localhost (marked "(simulated)" in evidence), so the live paths were proven against public test endpoints: expired.badssl.com (cert rejected → http fallback attempted → UNREACHABLE, correct) and http.badssl.com (upgrade probe found working https, so no flag — precision first).
• Fingerprints were validated on real markup, not just self-authored fixtures: a public Squarespace template demo (free-subdomain + platform both fired) and a public Wix template preview (platform fired). A wordpress.com theme demo answered 403 to headless Chrome — and the gate refused to judge it, which is the design working.
• Determinism: same input → byte-identical JSON across runs; staleness anchors to an explicit --today. No LLM calls anywhere; the run cost $0.

Graduation

Proven here ≠ wired there. The production memo (GRADUATION.md, in the Lab repo) specifies the input contract, the runtime, the human-review slice — every GREEN still gets the real four seconds before any pitch — and the one decision that belongs to the CEO, not the checker: what "batch-able" means for the bad-website segment. Lab code never copy-pastes into prod; the front door is the only door.