Cold-audience read
The system-backed arm produced shorter artifacts on average while preserving source boundaries, role fit, and self-sufficiency checks. That is worth inspecting. It is not yet enough to publish a broad lift number.
Pilot matrix
Each row uses the same task class twice: once as a cold run and once with the system context available. The public view shows only sanitized readings and measured values.
Customer-facing onboarding artifact
The system-backed output was shorter, source-bounded, and clearer about what context it did not need.
Architecture and decision brief
Both outputs found useful product shape; the system-backed run added decision boundaries and reversible tradeoffs.
Service blueprint
The cold hollow-dependency failure is retained as a likely heuristic false positive; the system-backed run passed all three checks.
Cold-audience one-page artifact
The system-backed output was more compact and deliberate about audience translation.
Doubt and risk record
The system-backed output used a reusable doubt protocol and named tests that would change the verdict.
What this proves
The public claim is narrow: in this five-cell pilot, the system-backed arm produced shorter, more source-bounded artifacts that passed the public review checks.
What it does not prove
The run used one model family, one pass per cell, synthetic tasks, and no blinded multi-rater scoring. The next proof needs preregistered arms, repeats, graders, and uncertainty intervals.
Next benchmark gate
The candidate bar is a meaningful blinded rubric lift over the strongest cold baseline while maintaining equal-or-lower private-context and dependency risk.