Repetition guards measure lines, not repetition

Field report: Repetition guards measure lines, not repetition. Read it rendered: https://macilydots.github.io/repetition-measurement/ Three findings A line-level repetition guard flagged 0 of 20 hand-verified responses with strong in-line repetition (median max line count: 1). On 300 fixed responses, deleting newlines removed all 14 line-level hits; rewrapping on sentence boundaries flipped 10/300. The model never ran. Most responses that crossed the repetition threshold still emitted a tool call (20/22 and 17/21). What this release does not claim No prevalence, no mechanism, no detection accuracy. Neither headline finding is recomputable from published material - see section 5 of the report. Full report: REPORT.md | Numbers: tools/repro/manifest.json