Harness observation runs on high-observability metrics

A human catching a wrong number in agent output is not quality control — it is an existence proof that observability is missing. The catch is the anecdote; the principle:

A quantitative claim in agent output is harness-grade only when it carries its provenance: value + source + measurement scope. Not 8, but 8 (organic search — app property 493986654, paths, 90d). Typed this way, misuse becomes a type error the gate rejects mechanically — a web-SEO verdict citing an app-typed number fails on its face — instead of a subtle wrong conclusion that survives until a sharp-eyed reader frowns at it.

Corollaries:

• The eyeball is fallback, never design. If the harness's verification story is "the reviewer scans numbers for thorns," the harness has no verification story.
• The two failure cases that motivated this note were the same defect: a conclusion's load-bearing number had an unpinned origin (one invented, one property-confused). Both would have died at the type layer.
• This is certificate-is-the-subagent-boundary applied to arithmetic: a bare number is self-narration; a sourced number is a checkable certificate.

kin: certificate-is-the-subagent-boundary · the FlexMesh method scar (never judge web SEO from app attribution) is this note's live specimen