Welford baselines for CPU/GPU/RAM/temps per workload. Voltage SPC with Nelson rules for real anomaly detection. Both were running. Both were tested. The chat just never called them.
When someone asked “82°C — is that bad?”, the AI had no idea if that's normal for your gaming laptop or a real problem. It gave the same generic answer every other monitor gives. I found it during a full audit of all 84 intents.
The two lines that killed it
I started grepping and found this in hybrid_engine.py:
except:
passTwo lines above it, the code tried to read a variable from another function. NameError every time. The bare except: pass just swallowed it. No crash, no log, no warning.
Fixed it in two places — removed the silent except, and actually wired the learning engines into the response builder. Now when hck_GPT answers, it finally reads what it learned about your machine. Same number, different meaning depending on whether you're gaming or idle.
The scary part isn't that the bug existed. It's that nothing ever told me it was there.
What's the longest a “working” feature in your code was quietly doing nothing?