Human review · 8 October 2026
Human review of the results
The human reading retains all 50 historically positive citations. Mentions become 18/294 (6.12%). Treating the four disputed cases solely as source labels gives 14/294 (4.76%). That second calculation is a hypothesis; it does not replace Isabelle’s choices.
On 7 and 8 October 2026, Isabelle Frassy reviewed all 50 historically positive citation or mention answers and 20 stratified negatives, five per platform. She declares a personal relationship with the author, no participation in the initial collection and no remuneration. Permission to name her is recorded. Independence is not established.
Version 1.2.0 · Counts, known denominators and coding sensitivity| J+30 scenario | Mentions | Citations |
|---|
| Historical edition retained | 18/295 · 6.10 % | 50/295 · 16.95 % |
|---|
| Human reading after discussion | 18/294 · 6.12 % | 50/294 · 17.01 % |
|---|
| Sensitivity: four cases as sources only | 14/294 · 4.76 % | 50/294 · 17.01 % |
|---|
| Sensitivity: two entire rows removed | 18/293 · 6.14 % | 49/293 · 16.72 % |
|---|
The two denominators of 294 cover different answers. The mention is unknown for J30-P001-CGPT-R1; the citation is unknown for J30-P005-GGL-R3. Each metric excludes its own unknown value. The last row removes both entire answers and their contributions to the numerator.
J0 retains historical coding: 32/295 mentions (10.85%) and 35/295 citations (11.86%). The 225 valid J+30 answers outside the sample also retain historical coding; this review does not newly validate them.
The ten cases reconsidered
Six mention ratings change from YES to NO: Perplexity R1, prompts P009, P010, P011, P013, P016 and P022. Four YES ratings remain: J30-P018-PPLX-R1 (review document), J30-P014-PPLX-R1, J30-P001-CLDE-R1 and J30-P015-CLDE-R1 (confirmation relayed by Bertrand). The latter confirmation is not presented as a new signed document.
Edikka is present in all four excerpts. The disagreement concerns its function: a mention in the answer or a source label. Text archives do not reliably preserve this visual distinction, and no original MIA-FR J+30 screenshot is available. The human-reading scenario retains the four YESs; source-only treatment remains hypothetical.
One supported correction and two unknown values
P025-PPLX-R3: the archived text supports NO for mention; citation remains YES. P001-CGPT-R1: missing text, indeterminate mention; the source URLs support citation YES. P005-GGL-R3: answer text exists, but the source file is empty; citation is indeterminate. For P007-CLDE-R1, the archive explicitly records no displayed sources, supporting citation NO independently of whether the answer names the brand.
The human total of 18 follows removal of the unverifiable P001-CGPT-R1 mention, addition of the P014/P018-Perplexity YESs, and removal of the P025-PPLX-R3 mention. It happens to equal the historical total of 18, but its members differ. No unknown value is imputed as NO.
First reading: 60/69 agreement on mentions and 68/68 on citations, with one and two indeterminate values respectively. After discussion: 66/69 and 69/69, with one indeterminate value per metric. Post-discussion agreement is not independent reliability; this targeted sample does not estimate accuracy across all 295 answers.
The conclusion about mentions depends on the four disputed codings. The interval includes zero in the human-reading scenario but is entirely negative in the source-only sensitivity. Citation intervals include zero in all four scenarios. No scenario is selected to obtain a more favourable conclusion.
25 French prompts, four platforms, three repetitions: 300 planned runs and 295 valid runs per wave. The additional off-plan Claude calibration remains excluded. The bootstrap uses 10,000 paired draws of 25 prompts with replacement; all platforms and repetitions within each prompt stay grouped. Identical draws for J0/J+30, seed 20260826, mulberry32, R-7 quantiles.
Private Word documents, correspondence, platform-session URLs and full third-party answer transcripts are not redistributed. Document hashes and the provenance of choices are published. Calculation reproducibility therefore does not, by itself, enable a complete repeat of the documentary review or certify its validity.
