Comparison — pm-morning-briefing-agent

Baseline run (40 runs, gate FAILED) → candidate run_approved (40 runs, gate PASSED)

Verdict: NO REGRESSION

Dimension	Baseline	Candidate	Δ	Direction
task completion	2.5%	2.5%	+0.0%	— unchanged
groundedness	5.0%	0.0%	-5.0%	▼ improved
completeness	2.5%	0.0%	-2.5%	▼ improved
constraint adherence	0.0%	0.0%	+0.0%	— unchanged
tool use quality	0.0%	0.0%	+0.0%	— unchanged
efficiency	5.0%	2.5%	-2.5%	▼ improved
recovery escalation	2.5%	2.5%	+0.0%	— unchanged

Cross-run signals

Informational — never affects the regression verdict; small samples make these too noisy to gate on.

Signal	Baseline	Candidate	Δ
mixed tasks (some trials violate, some don't)	6	2	-4
mean tool-set similarity across trials	0.9	0.975	0.075

completeness: Output omits material fact: "Margin mix flagged as a watch item into earnings"
efficiency: Run took 61200ms, over the 60000ms budget
groundedness: Contradicted claim in output: "Management pre-announced a $4B buyback acceleration."
groundedness: Contradicted claim in output: "Two desks downgraded AAPL to Sell overnight."

Finding diffs match on (dimension, checker, summary) and are indicative; dimension rates are exact.