An AI reporting method that fixes the evidence and numbers before generating prose showed far greater repeatability in an fMRI test. Claim-locked reporting had 98.5% reproducibility for report-visible numerical content across runs, compared with 61.1% for a hybrid template. That was a 37.4 percentage-point gap; the reported 95% confidence interval for the difference ran from 15.1 to 59.7 points.
The setup fixes which evidence-bound claims may appear, what numbers and directions they carry, and how strong the language is allowed to be before prose is generated. Prose then connects those approved pieces. The comparison treated reproducibility, rhetorical governance, numerical support and direction correctness as separate tests.
Two very different evidence sets
The study drew on an obesity-focused institutional fMRI cohort of 428 participants, including 248 people classified as obese and 180 as normal-weight. It also used a public HCP S1200 cohort of 712 adults, with 232 obese and 480 normal-weight participants after the overweight group was excluded.
A second test used 200 deduplicated Evidence Inference 2.0 records: 112 described statistically significant increases and 88 significant decreases. Records with no significant difference were excluded because a null direction made direction-inversion auditing undefined. The RCT result therefore applies to increased and decreased labels, not to inconclusive findings.
Stability was only one measure
In a governance check, which counted strong statements that were not qualified, claim-locked reports had 0.75 such sentences versus 2.25 for the hybrid template. The difference was minus 1.50 sentences, with a 95% confidence interval from minus 2.05 to minus 1.00. Both methods recorded a 0.00 numerical audit flag.
The RCT comparison sharpened the picture. Claim-locked reports reached 100.0% reproducibility, compared with 79.5% for the hybrid template, a 20.5-point difference. But the hybrid produced the lower raw numerical-flag value, 0.11 versus 0.23. The study therefore separates repeatability from numerical support rather than treating one as a substitute for the other.
A direction audit of 120 reports per method found that claim-locked reporting preserved 116 directions, inverted none and left four unresolved, for a reported inversion rate of 0.0%. The baseline methods had inversion rates between 5.6% and 14.3%. In this test, direction meant whether the evidence pointed to an increase or a decrease.
Manual calibration also changed how the raw flags should be read. It found harmful statistical-result fabrication in 0.0% of claim-locked outputs, compared with 2.5% for the hybrid template and 5.9% for free-form generation. Claim-locked outputs still generated 181 raw flags, but one was a regex false positive and 180 were benign mismatches.
The data still set the limits
The fMRI analysis included a separate warning about the evidence itself. In a group-covariate stress test, the number of FDR-surviving edges was 10,041 before BMI adjustment and 8 afterward in the institutional cohort, a 99.9% drop. In the HCP comparison, the count fell from 550 to zero, a 100.0% drop. This should be read as a covariate stress test, not as a biological finding about obesity.
That distinction is central to the paper's claims. Claim-locked reporting controls how structured evidence is verbalized; it does not validate the upstream statistics or guarantee truth beyond the evidence record. Errors in preprocessing, model specification, covariate selection or evidence extraction can pass into the resulting prose. The study does not establish that reproducibility alone makes a report faithful or correct, or that the method will transfer beyond the tested settings.
Operationally, claim-locked reporting was also the leanest option in the listed DeepSeek fMRI comparison, using 12,482 input tokens, 4,209 output tokens and 77.3 seconds of latency. Those figures were the lowest observed among the methods listed. This efficiency result is specific to the tested setup, not a promise about every model or reporting task.
The paper is a preprint whose front matter says it was accepted to EMNLP in August 2026 and identifies version 2 on arXiv dated 19 September 2026. The institutional fMRI cohort cannot be redistributed, while the public HCP and Evidence Inference 2.0 settings, prompts, generated reports, permitted evidence records, audit scripts and code are described as released. The acknowledgments report support from the National Natural Science Foundation of China, the National Key R&D Program of China, the Natural Science Basic Research Program of Shaanxi, Xidian University, the China Postdoctoral Science Foundation and the Postdoctoral Research Program of Shaanxi.
For language-model statistical reporting, the study's practical lesson is narrow: fix the provenance, numbers, direction and permitted language before prose is written, then assess correctness separately. The tests support more stable reporting under the configurations examined, not a guarantee that the underlying evidence or the final report is right.
Paper data and sources
Original title: Provenance Before Prose: Claim-Locked Reporting
Authors: Xiao Fan, Jingyuan Li, Hongbin Guo et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text