# CMFC exact-repair follow-up — report-level result downloads

Public research working paper, not peer reviewed.

Briefing: https://www.future-shock.ai/research/when-models-fail-to-repair
Methods: https://www.future-shock.ai/research/when-models-fail-to-repair/methodology
Paper: https://www.future-shock.ai/research/when-models-fail-to-repair.pdf

`results.csv` and `study-summary.json` contain selected reported aggregates transcribed from the retained D1 primary analysis and publication-sensitivity closeout. They are not raw observations and do not regenerate historical provider calls or the full statistical analysis. CSV fraction differences must be multiplied by 100 to express percentage points. The primary rows report dependence-inflation ratios, not failure probabilities.

The full confirmatory-success rule was not met. The implemented point estimator used equal-item rather than protocol fixed equal-stratum weights; a required precollection calibration-support projection gate was omitted and retrospectively fails. Mean/tail analyses are postconfirmatory; their intervals are unadjusted percentile intervals. The unrestricted missingness envelope includes one and is not a confidence interval.

The at-least-half threshold is >=4/8, not strict majority. A2-F synthetic-fixture/API verification is not empirical reanalysis. The separate unfinished validation campaign is not declared complete.

Restricted prompts, responses, gold state, provider payloads, event databases and operational records are not distributed. Publication does not grant redistribution rights to referenced upstream artifacts. Hashes establish file identity, not scientific validity.

Funding and conflicts: Self-funded by Nicholas Zinner; no external funding or competing interests to declare.