Editorial policy

This page collects the editorial conventions used across Initiative publications. The intent is twofold: to make our reasoning transparent to readers, and to make it straightforward for other groups to replicate our work.

Reporting conventions

• Agreement statistics over null-hypothesis testing. For accuracy and validity claims we report Bland-Altman analysis with 95% limits of agreement, intraclass correlation, and where relevant equivalence testing against a pre-specified margin. We avoid framing accuracy claims as the absence of a significant difference from a perfect reference.
• MAPE on calorie estimation as primary outcome. Where the underlying claim is about calorie estimation, we report mean absolute percentage error (MAPE) as the primary metric, on the grounds that absolute kilocalorie error fails to normalize across meal sizes. Per-meal absolute error is reported as a secondary outcome.
• Stratified breakdowns reported. Where sample size permits, we report stratified accuracy by cuisine bucket, by meal complexity (single-ingredient vs. mixed-dish), and by lighting / capture conditions. Strata with N<20 are reported but flagged for limited inference.
• Confidence intervals on every estimate. Point estimates without confidence intervals are an editorial defect.

Vendor-reported vs. independently-replicated numbers

Every accuracy number that appears in an Initiative publication is labelled as either vendor-reported (taken verbatim from a vendor's published source) or independently-replicated (measured against an Initiative reference set per the methodology brief in question). The two are never combined into a single aggregate number. Where we replicate a vendor-reported number and find that it does not reproduce, we publish the discrepancy with the same prominence as a confirmation.

Pre-registration

Validation studies are pre-registered where the design permits — typically with primary outcomes, exclusion criteria, and analysis plan fixed before data collection begins. Pre-registration links are reported on each publication that has one. Methodological work, narrative reviews, and commentary are not pre-registered.

Open data and code

We aim to release weighed-food protocols, per-meal ground truth tables, and analysis code alongside the corresponding publication, under a CC BY 4.0 license for documents and a permissive software license for code. The underlying photographic material is governed by the participant-consent terms recorded in each dataset description.

Authorship and contributorship

We use a CRediT-aligned contributor statement on each publication identifying who did what (conceptualization, methodology, data collection, statistical analysis, writing, etc.). Listed authors are those who meet ICMJE-style authorship criteria; other contributors are acknowledged.

Corrections

If a reader identifies a methodological or factual error in a published Initiative work, the report is reviewed by Initiative staff outside the named author list. If the error is confirmed and material, a correction is issued, the underlying article is updated with a visible correction notice, and the original version is preserved alongside.

Language and tone

Initiative writing avoids superlative claims about products. We use academic-register language: "demonstrated the lowest replicated MAPE", "showed statistically significant agreement", "warrants further investigation". We do not write that any product is "the best" or "the most accurate" without the qualifying methodology and the accompanying confidence interval.