Q Answer-format search page
Where do AI coding failures cluster — 7-bucket breakdown
✦ AI editor's answer
Of 11,147 measured failures: error 27.9% / timeout 22.9% / failure 21.3% / nonzero_exit 17.5% / permission 7.8% / exception 2.0% / fatal 0.5%. Permissions and timeouts are nearly as large as model errors themselves.
Source: 6 notes from this publication + operator work logs
Source notes (6)
- Failures & Cost2026-05-29
429 rate limit — the 6 minutes when the infrastructure died before the model did
11,147Cumulative failures (7,729 sessions / 132,293 events)› Automation dies first at outside infrastructure (quota, gateway, key cap) — not at the model. Same model, same prompt can still die in 6 minutes. Bake that in as a baseline.
- Failures & Cost2026-05-20
Remastering 178 cards — until the batch ran clean
11,147Failure events (7 buckets)› Enforce a No-Placeholder Policy on AI batch jobs — failures only become data if no fake file gets written on miss.
- Failures & Cost2026-05-22
Commands that should have been refused
874Permission denials inside 8,780 failures› Run every AI shell call through automated redaction and a gate before publish — human review is always too slow.
- In Practice2026-05-21
Five failures, one survivor at 1:40 a.m. — fixing the two-headed character
11,147Failure events across the batch› When an AI image pipeline breaks, try removing helpers (ControlNet, regional prompter, post-process segmentation) one at a time before adding more.
- Tooling2026-05-20
Mixing Claude, Codex and Gemini in one workspace — what 132K events revealed
81,764Claude events (61% of all)› Route 'hard-to-undo' work (design, review) to a stronger reasoning model; route bulk repetitive work to a fast cheap model.
- In Practice2026-05-20
One character took 13 phases to ship
13Phases to ship one character› Don't try to finish an AI training pipeline in one shot — split into cleanup → training → validation across 13 phases.
Related questions