Abstract
Evidence for agent products should not move by vibes. A dogfood run, an internal field case, an outside-host packet, and a benchmark-style corpus each answer different questions. This paper studies Signal Box as a staged evidence ladder for review-packet systems. The local Spinoffs research tree contains Level 1 dogfood evidence, Level 2 internal field evidence, Level 3 outside-context evidence, and a Level 4 benchmark-style scaffold with an error taxonomy. The counted source reports preserve denominators and nonclaims: Level 1 records 77 candidate/run cases with 59 accepted Level 1 cases and 18 reviewed failed-run cases; Level 2 records five accepted internal Switchboard field cases; Level 3 records three accepted outside-context cases while refusing product-market-fit, repeated-adoption, public-adoption, benchmark, and general review-time claims; Level 4 remains a draft scaffold requiring taxonomy and redacted corpus work before freezing. The contribution is a claim architecture for agent-produced review packets. Signal Box evidence becomes stronger only when the next rung's evidence exists; it does not become stronger because the story wants to move.
← Back to research papers