F408 Tools built for agents were proved in named consumer repositories, by counted changes there, and the consumer found what the producer's tests did not

Field	Value
Status	active
Topic	Environment, permissions, and agent-facing products
Grade	R3/S3
Cases	winwright, polyweave, roadkeep
Contrary	none
Harness facts	none

Claim

The projects that build tools for other agents took their evidence from a named consumer repository, recorded adoption as a counted change in that consumer (lines deleted, checks passing, a measured premise), and the consumer surfaced defects and false assumptions that the producer's own suite had passed.

Mechanism

A producer's tests exercise the inputs its author imagined; a consumer brings real artefacts, real hardware and a history the tool must fit. Deleting the consumer's own hand-built replacement is a test the producer cannot pass by construction: the consumer's cases either run on the tool or they do not. The cost is that the evidence comes from one consumer at a time, and a consumer's cases can themselves be unrun.

Observations

winwright. Adoption was closed by deletion in two consumers. In claude-tray,

"claude-tray still carries a 3,004-line interaction harness" [winwright@861b82e:docs/CHANGELOG.md#L449] was answered by deleting it, where "23 cases check 95 claims in 128-159 s" [winwright@861b82e:docs/CHANGELOG.md#L449] and the script had "checked 31 in 46 s and crashed" [winwright@861b82e:docs/CHANGELOG.md#L449] [winwright@2b452ec]. In pportal the file "is deleted, 285 lines in which every case decided for itself how long to wait" [winwright@6917d3f]. The roadmap requires this form of proof: "the number of lines removed is reported rather than described" [winwright@861b82e:docs/ROADMAP.md#L137].

polyweave. Its backlog comes from one consumer: "Nearly every symptom in the backlog was measured in" [polyweave@6d1c136:CLAUDE.md#L25]

Cottony, which "is the first consumer and not the specification" [polyweave@6d1c136:docs/ROADMAP.md#L117-L118]. Real artefacts copied from it overturned a synthetic threshold: a real mesh and its drawing agree at a "silhouette IoU of 0.4343" [polyweave@0cacaad] while "The synthetic tests clear 0.8" [polyweave@0cacaad]. A fix was proved in the consumer, whose suite "still passes 895 checks across 13 files" [polyweave@f203b0a].

roadkeep. Its premise was counted in a consumer, Shio: "92 roadmap lines averaging" [roadkeep@91754240:agents.md#L7]

142 words against a one-sentence rule. Consumers exposed its launcher's defects: the MCP server's stand-down was "Measured in Shio, where that is worse than cosmetic" [roadkeep@91754240:hooks/roadkeep-launch.py#L59], and a stale registry row in another corpus left "both guards absent at once" [roadkeep@91754240:hooks/roadkeep-launch.py#L152]. A round-trip property test runs over "Shio's and Turing's at the revision" [roadkeep@91754240:.claude/skills/roadkeep-dev/SKILL.md#L24-L26] pinned.

Independence

Three independent origins: each project names its own consumer and its own measurements there, and none of the three practices is copied from another. Shio appears as roadkeep's consumer, so it is not counted as a case of its own here; freewilly and Shio record no consumer of their own agent surface at the pins.

Counter-evidence

The proof was sometimes overstated. winwright corrected its pportal entry the same day: "The cases were written on a desk with no controller, so the two that need one had never run." [winwright@d2a9081] The size of the deleted claude-tray harness is given two ways, 3,004 lines in the ledger and "claude-tray's harness is 2,732 lines for" [winwright@861b82e:src/Winwright/Scenarios/StepDeclaration.cs#L71] eight cases in the engine, and the pin does not reconcile them. polyweave files its own limit, "every symptom in the backlog was measured on one game" [polyweave@6d1c136:docs/ROADMAP.md#L28]. roadkeep's consumer test skips where the corpora are absent, "which is what CI does" [roadkeep@91754240:.claude/skills/roadkeep-dev/SKILL.md#L26]. The deletions in claude-tray and pportal are cited through winwright's own records; those repositories are outside the corpus.

Grade

R3: three independent origins in three projects. S3: the sources count what changed in the consumer (lines deleted, cases and claims, checks passing, an IoU, word counts). No check in the producer fails when a consumer stops adopting it, so S4 is not reached.