60-82% accuracy swing on 4B model classification task: the only variable was harness design

I ran a pre-registered ablation on a classification task (Kubernetes issue → SIG triage) using a 4B model on a 6GB laptop GPU. Same frozen weights, same 250-issue gold corpus, same scorer across every run. The variable under test was harness design: rule placement, evidence order, turn structure, what survives between turns. Result: 22 points of accuracy, same model, same task. 60% at the worst harness, 82% at the best. "This model is bad at X" is often actually "my harness is bad at X." What moved accuracy: Explicit rules in the prompt: +13 Task before reference material (not after): +6.5 One extra reasoning turn: −5 Clearing context each turn, carrying a summary forward instead of raw evidence: −12 Fresh-session handoff between stages: −15 The worst-designed harness paid for an extra stage and 250 tool calls and got nothing for it - landed right back at bare-model accuracy. Everything's public and archived - corpus, scorer, pre-registration, every run manifest. You can re-score the results without a GPU; you only need one to generate new predictions. Eval harness: https://github.com/TGPSKI/leather/blob/main/examples/14-sig-triage/eval/README.md Example overview: https://github.com/TGPSKI/leather/tree/main/examples/14-sig-triage Matrix results: https://github.com/TGPSKI/leather/blob/main/examples/14-sig-triage/eval/results/MATRIX.md