Class 88 taught you to connect workflows carefully and measure their value. Now volume rises — ten lessons a month, a hundred support replies a week — and quality becomes the thing that breaks silently. Nobody announces "standards slipped." They just slip, one unreviewed publish at a time. This lesson makes quality a mechanism, not a mood.

The vocabulary, in plain language

• Acceptance criteria: the checkable list a piece must pass before release — not "good," but "every claim sourced, every link live, voice match, image unique."
• Sampling: checking a subset deeply instead of everything shallowly — 3 of 10 publishes get full audit; all 10 get the quick gate. Statistics doing the heavy lifting.
• Audit: the deep check against the full standard, by someone other than the producer — the second pair of eyes with a checklist.
• Regression: re-breaking something you already fixed — the corrected error that returns two months later. Regression checks exist because memory fades.
• Corrective action: what happens after a failure — fix the piece, fix the cause (checklist, prompt, training), and verify the fix. A rejection without a cause-fix is just a delay.

The mental model: a print newsroom's copy desk. Writers write; the desk checks facts, sources, headlines, and libel — every edition, same checklist, no exceptions for star columnists or slow nights. The desk isn't the enemy of speed; it's why readers trust the paper.

The ByeBuy publication bar, concretely

Use the Classroom's own workflow (research → outline → draft → artwork → editorial review → publish → update) as the example. Every publish passes six gates:

1. Factual correctness — every claim a reader could act on is accurate as of the date; no invented capabilities, quotes, or statistics (per the editorial standard's "never invent" rule). 2. Source quality — claims that need evidence cite direct sources or docs; AI-unsupported guesses are marked unknown, never asserted. 3. Voice — clear, direct, professor-warm; no compliance-memo stiffness, no content-mill padding. 4. Unique image — one distinctive hero visual per lesson, readable at desktop and mobile size, with meaningful alt text. 5. Links — every tool path links to official docs (n8n, Make, Zapier, HubSpot, Stripe where used); no dead or guessed URLs. 6. Curriculum progression — the bridge from prior lessons holds, one outcome is taught, and the handoff to the next lesson is earned.

AI may prepare (structure, variants, metadata — Lesson 85.3) but the gates stay human until quality is visible over time (Lesson 84.4's autonomy ladder).

Exercise: write the quality standard

Create QUALITY-STANDARD.md:

# QUALITY-STANDARD.md — [Operation, e.g. ByeBuy publication]

## Acceptance criteria (checkable, no adjectives without tests)
- [ ] Factual: ___ (how verified — sources dated ___)
- [ ] Source: ___ / Voice: ___ / Image: ___ / Links: ___ / Progression: ___

## Sampling
- Quick gate: every item, by ___ (checklist ___ min)
- Deep audit: ___ of every ___ items, by ___ (not the producer)

## Pass / reject
- Pass: all gates green → publish + log reviewer + date
- Reject path: red item → return to ___ with note ___ → re-audit before publish
- Regression check: ___ (e.g. past corrections re-checked monthly)

## Corrective action
- Fix piece → fix cause (checklist/prompt/training change: ___) → verify in next audit
- Log: rejections + causes reviewed in OPERATING-REVIEW (Lesson 89.5)

Worked mini-example — Sonariq report: quick gate on every report (claims have citations, no unknown-stated-as-fact); deep audit on 1 in 4 by a second analyst; reject returns to author with the failing gate named; a repeated citation-format failure triggers a template fix, not another lecture. Sample numbers are small on purpose — a standard you actually run beats an ambitious one you abandon.

Finish line: a QUALITY-STANDARD.md with checkable gates, a sample size, named auditors, and a rejection path that fixes causes.

Verify quickly: take one recent piece through the checklist. If a gate can't be answered yes/no in five minutes, rewrite the gate. If the producer audits their own deep sample, reassign.

Common failure mode: the vibes review — "looks good to me" with no checklist, no sample, no log. It passes everything until the week it passes something embarrassing publicly.

Check your understanding

1. Why must the deep auditor differ from the producer? 2. What's the difference between fixing a piece and corrective action? 3. Which of the six ByeBuy gates would catch an invented product capability — and how?

Next

Quality gates hold the line on correctness. But scale has a second quiet killer: cost. Lesson 89.2 puts guardrails on model, API, infra, and per-workflow spend before margin melts.