Phase P7 — the publishable counter-intuitive.
pawbench --ablate <component> disables one pipeline component at a time and reports DQS delta. In scope:
- response gate
- AC re-verification (single-run only)
- DQS rewrite queue
- doctor pre-checks
- mid-trajectory pruning
pawbench --ablate response_gate,ac_verify,rewrite_queue,doctor,trajectory_pruning produces a 5-row delta table.
Policy: any component with negative delta on ≥ 2 consecutive weekly runs is filed as a removal candidate (issue auto-opened). Surprising negatives are published — that is the entire point of this phase.
Why: Fabian's "quality-focused instructions decreased performance" is the most counter-intuitive result in his write-up. We've added 5 quality gates over 6 months and never measured each in isolation. Either this validates the pipeline or kills dead weight.
Phase P7 — the publishable counter-intuitive.
pawbench --ablate <component>disables one pipeline component at a time and reports DQS delta. In scope:pawbench --ablate response_gate,ac_verify,rewrite_queue,doctor,trajectory_pruningproduces a 5-row delta table.Policy: any component with negative delta on ≥ 2 consecutive weekly runs is filed as a removal candidate (issue auto-opened). Surprising negatives are published — that is the entire point of this phase.
Why: Fabian's "quality-focused instructions decreased performance" is the most counter-intuitive result in his write-up. We've added 5 quality gates over 6 months and never measured each in isolation. Either this validates the pipeline or kills dead weight.