Fact-checking all 202 cards from source — 45 real errors, no forced changes
After repeatedly claiming the study content was done when it wasn't fully verified, ran a pass where 41 agents each read cards from the actual source and fact-checked code, code explanations, English, and model answers. It corrected 23 more cards on top of an earlier 22 — 45 real errors across 202, about a fifth of the deck.
The honest backdrop: I'd said the deep-study content was done more than once before it was actually verified — the verify stage had been rate-limited and the draft shipped, and an answers-only re-check had caught 22 wrong answers. The owner, rightly, wanted every one of the 202 cards checked properly: code, the Korean code explanation, the English, and the model answers — not assumed.
So this pass had 41 agents, each reading its cards straight from the lib source (grep the id, read the block) and fact-checking every field against what's actually there, returning corrected values only for genuine errors — no style rewrites, no forced changes. It found 23 more cards to fix on top of the earlier 22, for 45 total: roughly a fifth of the deck had at least one real error. Examples: the availability card's runnable comment and nines table said 43.8 where it should be 43.2; the load-balancer L4-vs-L7, health-check, and sticky-session answers; the circuit-breaker state-machine walkthrough; the Kafka topic and consumer-group answers and code explanation.
Applied as full-value field replacements, html-unescaped and escaped for the template literals, zero entities left, tsc clean.
Two lessons worth keeping. First, verification that reads the integrated source — not the generation draft — catches integration-era and cross-field errors a generate-time check misses. Second, and this is the one I had to learn the hard way this session: "it generated, so it's done" is the wrong claim for anything the user memorizes. Say "generated," verify against source, and only then say "verified" — because about one in five cards was wrong.