In one agent telemetry report, core process metrics were attributed to model epochs inside threads while the completion proxy was computed once per eligible main thread, so the same model label carried two different sample sizes.
Reality
- Evidence38
- Adoption
- Insufficient
- Hype gap−12
- Incentives25
- Confidence50
One developer's interactive piece checked recorded model replies for two Korean words, so a question and a negated proposal both ran join. The repair narrows the executable surface to four exact sentences.
Reality
- Evidence58
- Adoption
- Insufficient
- Hype gap−8
- Incentives18
- Confidence45
Vision read the characters and lost the row associations, a failure that leaves no error signal downstream and that a comparison against real documents, where nobody holds the ground truth, could not have caught.
Reality
- Evidence45
- Adoption
- Insufficient
- Hype gap+25
- Incentives15
- Confidence55
The rule had real callers and fired on every invocation, but the field it compares against had no writer anywhere in the repo, so backfilling it was the only way to learn what it would have flagged.
Reality
- Evidence47
- Adoption
- Insufficient
- Hype gap+6
- Incentives24
- Confidence54
The bar was 50% of the original motion median, and the same clip honestly returned 47.5% and 56.5%, because the registration fixed the bar and left the sample count to the pipeline. The rule written to close that gap then produced a compliance rate with the same hole.
Reality
- Evidence58
- Adoption16
- Hype gap−9
- Incentives31
- Confidence61
A usage cap left a three-model review panel with two voters, which split 1-1 on half its batch. Both splits were settled by the options both legs threw away, and by a falsifier one model wrote against itself.
Reality
- Evidence33
- Adoption11
- Hype gap−9
- Incentives22
- Confidence44
Five published episodes cleared all twelve verification items. A later audit found four defects, all of them in the 45 percent of on-screen instances the proposed better scan could not see.
Reality
- Evidence46
- Adoption11
- Hype gap−6
- Incentives22
- Confidence44
A verification call placed between the upload and the local record meant one failure left a real video on the channel and nothing on disk. Five earlier uploads had the same hole, and every test was green.
Reality
- Evidence38
- Adoption10
- Hype gap+8
- Incentives26
- Confidence44
A width cap sized in all caps for captions that render in sentence case chopped lines into single words. The measurement library was accurate to within 1 percent; the string it measured was 200px off.
Reality
- Evidence58
- Adoption
- Insufficient
- Hype gap−5
- Incentives24
- Confidence55
A clean-room env check confirmed three credentials were gone from the profile. One of them was still signing live requests, held in a parent process that had started days earlier.
Reality
- Evidence52
- Adoption
- Insufficient
- Hype gap+12
- Incentives22
- Confidence58