Counting redirects changes which rows on the sheet look bad. Q01 passed three attempts out of three and needed nine guard redirects to get there [9], an average of three per successful run [2]. Under pass counting that row is green and indistinguishable from a clean solve.
Redirects are only countable if the guard refuses in a fixed, parseable string. Theirs does: `STOP: use scaffold, not ng generate` [5]. That is the first precondition for any of this transferring to your stack. The second is a stable grader. The 25/36, or 69 percent, is scored by the team's own detectors [8], and those detectors were being edited in the same period: synonyms, corrected precedence between task "kinds", scaffold guards for reports took the third holdout set from 22/36 to 36/36 [14], which is 61 percent to 100 percent [3]. Fourteen attempts flipped without the weights changing. Cross-week comparisons mean nothing unless the detector set is frozen between them.
The environment bug is the part worth copying into your own postmortem template. The between-runs reset ran `DELETE WHERE id > baseline`, which reverses inserts and nothing else, so `UPDATE`s like `hide_in_list` survived every reset until columns on the `customers` table were hidden and `route_columns` started answering "no columns found" [10]. The fix was to restore the full column set from the clean baseline on every reset, because a reset has to cover what you mutate, not just what you add [12]. Separately, a shared file, `gen-report-from-template.mjs`, had been overwritten by a model run and shrank from 246 lines to 60, in a directory the reset never touched [13]. An agent with write access to its own harness will eventually exercise it.
Then the compounding failure: some test prompts had been reworded over time to compensate for the drifting database, so the team was fixing symptoms of an environment bug and reading them as backend behaviour [11]. Slow drift produces phantom regressions that look exactly like model bugs [18]. The prompts had absorbed the defect.
The technique choice follows from the metric. Supervised fine-tuning masked the rejected turns, so the model never received the signal that a first move was wrong [15]. DPO scores the pair instead: same context, prefer the accepted action over the rejected one [16]. That maps directly onto the redirect log, because a redirect is a rejected action with an accepted one two turns later.
Note what the write-up does not yet claim. It frames the result as a 30B model that did learn, though not exactly what was asked [17]. Take the method, not the score. The scoring conditions are specific to one framework's scaffolds, one guard vocabulary, and one model served locally: qwen3-coder:30b, a Mixture-of-Experts with roughly 3B active parameters, on a single RTX 4090 via Ollama with no data leaving the machine [2]. That last detail is why the loop is affordable enough to rerun after every detector change, which is the actual enabling condition here.
