Not another yes-or-no verdict. A verifiable map of what the evidence actually says, how certain it is, and why studies disagree.
The signal is more consistent for processed than unprocessed meat. Individual risk still depends on baseline consumption and personal risk factors.
Association ≠ causation. A statistical relationship may still reflect confounding or measurement error.
Hazard ≠ individual risk. An IARC classification does not tell you how much one person’s risk changes.
Compared with what? Replacing meat with legumes is not the same as replacing it with refined carbohydrates.
A living conclusion needs rules set before the next result arrives. These thresholds follow the selected outcome; they are editorial criteria, not a forecast.
A consistent association with increased risk
Low–moderate certaintyIndependent cohorts with repeated diet measures and explicit replacement foods, a convergent dose–response, narrower prediction intervals and mechanisms at ordinary exposure levels.
Comparable high-quality cohorts or harmonized participant-level analyses that repeatedly approach the null across definitions, regions, lag periods and bias adjustments.
One new association, a p < 0.05 result, or a mechanistic experiment at an unrepresentative dose.
One cohort reported model-estimated absolute differences as well as hazard ratios. Change the horizon and exposure to see why a percentage without time and comparator is incomplete.
cardiovascular events
per 1,000 people over 30 yearsModel-associated difference
95% confidence interval: +8.5–+26.3 / 1,000 · +1.7 percentage points
2 vs 0 servings of processed meat per week
No explicit replacement food. Other foods were adjusted for, so this cannot answer “replace with legumes, fish or refined grains?”
A model-estimated association for a cohort with average covariates—not an individual forecast and not proof of a causal effect.
The interface does not fit a new risk model or invent a personal baseline.
Fully adjusted model 3 risk differences reported in Table 2 at 10, 20 and 30 years.
The authors applied the mean values of included covariates across six pooled US cohorts. That profile is not your baseline risk.
Published percentage-point difference × 10 = associated events per 1,000. The same conversion is applied to both confidence limits.
The 95% interval came from 500 bootstrap samples. It represents statistical uncertainty—not residual confounding or exposure-measurement error.
A preregistered meta-analysis compared specified substitutions. Select what replaces 50 g/day of processed meat; the outcome is total cardiovascular disease.
associated with a 27% lower relative hazard
95% confidence interval: 0.59–0.91 · 8 cohorts
50 g/day processed meat → 28–50 g/day nuts
The source did not provide one shared baseline risk and time horizon. Converting this relative estimate into an absolute number would require assumptions the study did not report.
GRADE was moderate for these three estimates, but every included publication was observational and none was judged at low risk of bias.
Nine claims across four articles. We preserve the valid methodological criticism, mark where the reasoning outruns the evidence and link an internal decision-source audit only when the match is exact.
The warning about relative risk is useful; the conversion is not.
Headlines should show baseline and absolute risk, not only a large relative number.
The paper followed 216,695 people and reported 22,761 cases. Its headline 1.62 was a hazard ratio comparing extreme intake quintiles in a model that intentionally omitted BMI as a possible mediator. Adding time-varying BMI reduced the estimate to 1.23.
Cases divided by person-years are incidence rates, not each participant’s cumulative absolute probability. A hazard ratio is not a risk ratio. Calling roughly 0.29 versus 0.52 cases per 100 person-years “absolute risk” mixes incompatible quantities and hides the important BMI sensitivity analysis.
Exposure misclassification is real; the proposed alternative cause was not tested.
Mixed dishes and self-reported diet make it harder to isolate the effect of meat itself.
The exposure definition did include processed-meat sandwiches and beef, pork or lamb in sandwiches or mixed dishes. But diet was reassessed every 2–4 years, averaged cumulatively, adjusted for refined grains and overall diet, and calibrated against two 7-day weighed records in 1,207 participants.
FFQ error weakens precision; it does not show that bread or oil generated the association. That would require a component-level or substitution analysis. The published exposure list supports “mixed dishes,” not the claim that lasagna was the causal driver.
Baseline imbalance is a limitation, not a complete refutation.
Higher-meat participants differed in smoking, alcohol, BMI and diet. Residual confounding cannot be eliminated.
The 474,996-person UK Biobank analysis adjusted for smoking, BMI, activity, deprivation, alcohol, fibre, fruit, vegetables, dairy and fish; calibrated amounts with repeated 24-hour recalls in 69,076 people; excluded early cases and reran analyses in never-smokers; and corrected for 22 outcome tests.
After those checks, colorectal cancer was the only robust cancer signal; most other associations disappeared after multiple-testing correction. The honest conclusion is “observational signal with residual uncertainty,” not “the baseline table disproves it.”
A rat surrogate experiment became a human treatment implication.
Animal models can test mechanisms and sometimes produce results that challenge a prior hypothesis.
One hundred carcinogen-initiated rats ate diets containing 30% or 60% freeze-dried cooked meat for 100 days. The endpoint was aberrant crypt foci, a putative precancerous marker—not human cancer, survival or treatment of an existing tumour.
The authors themselves proposed that bacon’s salt increased water intake and diluted promoting compounds; they explicitly warned about transfer from this rat model and surrogate endpoint. The study cannot support advice that bacon prevents or treats human colorectal cancer.
No exact match in the 21-card decision corpus. The external evidence links remain the authoritative trail for this claim.
Particle modification matters, but it is not a prerequisite for LDL causality.
Oxidation, glycation, inflammation and particle characteristics affect atherosclerotic biology and individual risk.
More than 200 prospective, genetic, Mendelian-randomization and randomized-treatment studies—over 2 million participants and 150,000 cardiovascular events—show concordant dose- and time-dependent effects of LDL/apoB particle exposure.
That triangulation is not explained away by saying only modified LDL matters. Nor does LDL causality prove that one food has a fixed effect: cut, saturated-fat content, replacement food and the whole diet still determine the dietary comparison.
A country-level correlation cannot identify an individual dietary effect.
Cross-country data can challenge simplistic universality and generate hypotheses about context.
The comparison uses national meat supply/consumption and national life expectancy. Those variables also move with income, sanitation, infant mortality, healthcare access, smoking history, urbanization and many other exposures.
Inferring what happens to individuals from grouped national averages is the ecological fallacy. It is methodologically weaker than the individual-level cohorts Hoole rejects and cannot establish that meat caused longevity—or that moderate intake is harmful.
No exact match in the 21-card decision corpus. The external evidence links remain the authoritative trail for this claim.
Large individual-level international comparison ↗Hoole’s article ↗
Two national trends moving in opposite directions do not isolate causation.
Long-term trends should be compatible with a proposed causal story and can reveal contradictions worth investigating.
The argument compares aggregate intake with a broad, undefined category of chronic disease over decades. It also excludes pork from red meat by personal definition, unlike standard epidemiologic classifications.
Population aging, obesity, smoking latency, diagnosis, survival, exercise, total calories and the rest of the diet all changed. Without outcome-specific age-standardized rates and a design controlling those trends, the graph neither proves harm nor safety.
No exact match in the 21-card decision corpus. The external evidence links remain the authoritative trail for this claim.
Hoole’s time-trend argument ↗IARC’s standard red-meat definition ↗
A real sampling limitation is fused with the wrong historical dataset.
The Seven Countries cohorts were purposively, not randomly, selected; external validity and unmeasured cultural differences are legitimate limitations.
The famous 22-country comparison was a separate, earlier ecological dataset. The later Seven Countries Study began in 1958 as a prospective study of 16 cohorts and 12,763 men using standardized individual measurements and follow-up.
Saying Keys removed 15 outcome-inconvenient countries from the prospective study conflates two projects. Correcting that history does not turn the Seven Countries Study into a randomized trial or make it sufficient, by itself, to settle diet–heart causality.
No exact match in the 21-card decision corpus. The external evidence links remain the authoritative trail for this claim.
Seven Countries Study design ↗Historical review ↗Hoole’s article ↗
Nutrient density is not evidence of long-term completeness or safety.
Red meat is protein- and micronutrient-dense, and adequacy depends on cuts, organs, total energy and individual needs.
The article relies on nutrient composition, evolutionary narrative and personal experience. It does not cite a long-duration controlled comparison of an all-red-meat diet measuring deficiency, cardiovascular, cancer and mortality outcomes.
A table of nutrients cannot establish absorption, requirements across populations, long-term clinical outcomes or the safety of excluding all other foods. The categorical prescription exceeds the evidence even if short-term experiences are positive.
No exact match in the 21-card decision corpus. The external evidence links remain the authoritative trail for this claim.
The 2023 diabetes cohort was NIH-funded, declared no author conflicts, repeated diet measures and used calibration—but remains observational and its headline estimate is BMI-sensitive.
A 2022 meta-analysis found no overall short-term glycemic harm. Beef Checkoff funded it, commented on early design, and the authors’ employer had beef and pork-board funding. Most trials lasted 4–8 weeks.
A 2026 preregistered review found meat-industry-tied studies far more likely to publish favorable conclusions (OR 16.4). That does not invalidate each paper; it makes methods and wording—not sponsor labels alone—the unit of audit.
Every metric uses the cards shown after the evidence filter. Counts describe public source cards—not independent study families; overlap remains possible until the definitive corpus is deduplicated.
These 21 decision-relevant sources and 9 claim audits demonstrate the product; they are not yet a complete systematic review. The full corpus will be deduplicated, checked by two reviewers and published under a preregistered protocol.
The interactive table answers the selected question. This complete register keeps all 21 source audits in the initial page HTML so source unit, methods, limitations, editorial audit concern, causal-interpretation boundary, registration/protocol, comparator, absolute-risk and sensitivity status, funding, conflicts and sponsor role, data access and code status can be found, linked and checked without running JavaScript.
Processed meat was classified as carcinogenic and red meat as probably carcinogenic. This is a hazard classification, not an estimate of an individual’s risk.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Rates the evidence as convincing for processed meat and probable for red meat in colorectal cancer risk.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
No uniform difference: plant protein sources were more favorable for some lipids than red meat.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Issued a weak recommendation to continue current consumption, based on low certainty of effects and individual preferences.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Small increases in relative and absolute risk associated with processed and unprocessed meat; no association for fish.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Rated evidence as weak for several outcomes and absent for some, using a conservative evidence-strength scale.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Positive associations for processed and unprocessed meat across many populations, but not proof of causation.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Higher intake was associated with CVD and T2D in both meat categories. The pooled T2D estimates were HR 1.27 for unprocessed and 1.44 for processed meat, but between-study heterogeneity was 88.7% and 90.5%, respectively.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Supports biological plausibility through multiple mechanisms; cannot quantify the clinical effect in humans by itself.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
The headline HR was 1.62 without BMI; it fell to 1.23 with time-varying BMI. Both estimates are observational, not an absolute-risk calculation.
After multiple-testing correction, the robust signal was specific to colorectal cancer; most other cancer associations did not survive.
No overall effect on most short-term glycemic markers; this does not test years-long diabetes incidence.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
The legume arm had lower LDL cholesterol and weight, but no between-group glycemic differences; a 2026 companion found no major inflammatory, TMAO or gut-barrier differences. Vitamin B12 and iodine markers were lower.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
The plant-burger diet lowered the registered primary TMAO endpoint and LDL, but LDL was secondary and hypothesis-generating. Weight and NT-proBNP were slightly higher, alongside much more sodium in the supplied plant patties.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Most direct pork-versus-legume differences were null. HDL was higher after pork, but fell from baseline in both diets; homocysteine was 1.6 μmol/L higher after pork. Both low-UPF diets improved several markers versus habitual baseline, without proving that the protein source caused those changes.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Independent evidence streams converge on a dose- and time-dependent causal role for LDL-containing particles in ASCVD.
Industry-tied studies were much more likely to reach favorable conclusions. This is a reason to audit design and wording, not to auto-reject every result.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Associations span several chronic outcomes, but all GRADE ratings were low or very low; the review does not raise causal certainty.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
The conservative model estimated at least 11% higher average type 2 diabetes risk and 7% higher colorectal-cancer risk across observed intake ranges; every processed-meat relationship received a two-star rating for weak or inconsistent evidence.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Replacing 50 g/day of processed meat with nuts, legumes or whole grains was associated with lower total CVD hazard; the exact estimate changed with the replacement food.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
The moderate substitutions produced small, sex- and follow-up-sensitive associations rather than one stable effect; equalizing follow-up to 7.9 years changed several results.
None of the nine current claim audits points to this card by exact match—this is not evidence that the source is irrelevant.
Relevant-looking papers can answer the wrong question. These sources remain public, with the exact reason they do not enter the selected evidence count.
Direct eight-week comparison of Beyond Meat products with provided animal-meat products; TMAO, LDL and weight favored the plant phase.
The animal arm mixed beef, pork, processed sausages and chicken, while some fish was allowed. It cannot isolate unprocessed or processed red meat for the exact-question table.
Prospectively registered with a public analysis plan. Beyond Meat funded the study through an unrestricted Stanford gift; a blinded third party ran the main analysis. A significant TMAO order effect and no washout limit interpretation. Data, codebook and code are request-only. No linked Crossref notice or indexed exact-DOI PubPeer page found.
Nine prospective cohorts modeled replacing 3% or 5% of energy from animal protein with plant protein for mortality outcomes.
The exposure aggregates all animal protein and does not identify red meat, processed meat or a specific replacement food. Its estimates cannot answer the pilot’s food-level question.
PROSPERO CRD42024614889; observational input with substantial heterogeneity for all-cause mortality. One author disclosed Danone consulting outside the work. No linked Crossref notice or indexed exact-DOI PubPeer page found.
A current review of young-onset colorectal cancer in Mexico repeatedly discusses red, pork and processed meat as modifiable exposures.
It is a narrative review of selected literature, not an eligible primary study or a systematic review with a reproducible selection flow. It reports no original red-meat effect estimate that can enter the causal synthesis.
The publisher full text describes searches of PubMed, SciELO and ResearchGate, prioritizing 2019–2026, but no prospective registration or systematic selection flow. The authors declared no funding and no financial or other conflicts; no original dataset, analysis code or separate review supplement was identified. It remains a discovery lead, not a counted source.
Each source now exposes the identifier that was checked, any linked notice or formal discussion, the web-indexed PubPeer signal and what can actually be reanalysed. The register distinguishes ‘nothing found’ from ‘nothing exists’.
The versioned runner queried Crossref REST API v1 once for each of the 20 article records. All 20 were retrieved with zero request errors; update-to exposed no linked retraction, expression of concern or correction. The known NutriRECS correction, publisher-linked discussion and indexed PubPeer search remain explicit separate checks.
A questionnaire completed once twenty years ago is not equivalent to repeated diaries or biomarkers.
Potential impact: highMeat versus legumes can yield a different answer than meat versus white bread or a similar diet.
Potential impact: very highSmoking, income, physical activity and overall diet quality may move together with meat intake.
Potential impact: highA sponsor can shape the question, comparator, duration and conclusion without falsifying the numbers.
Analyze it; never use it as an automatic vetoMature review infrastructure, living synthesis products and newer AI workbenches already cover much of the production stack. The honest comparison separates infrastructure, public output and responsibility for a durable health conclusion.
They commission, synthesize and publish evidence—not merely search it.
Rigorous reviews, living evidence or unusually strong consumer nutrition summaries.
What it does not replaceUsually organized around review questions or reference topics—not a public audit trail from a viral claim through calculations, funding, comparators and sensitivity choices.
They maintain updatable corpora and publish interactive, traceable evidence views.
This is the closest product overlap: continuous screening, structured extraction, interactive synthesis, source traceability and public living outputs.
What it does not replaceThey invalidate any claim that living, public or traceable synthesis is unique. The remaining distinction must be narrower: editorial responsibility for a contested health conclusion plus claim fidelity, funding, absolute-risk arithmetic and result-changing sensitivity choices in one public record.
Mature platforms manage screening, extraction, audit trails and living updates at team or enterprise scale.
The baseline is already high: independent screening and conflict resolution, structured extraction, risk-of-bias tools, provenance, living imports and reproducible or human-verified automation. EPPI-Reviewer also supports synthesis and live public views; ASReview makes active-learning screening open and inspectable.
What it does not replaceThese systems make reviews; they do not automatically become the accountable public editor of a contested conclusion. Their capabilities also differ materially—screening-focused ASReview is not equivalent to enterprise DistillerSR or end-to-end EPPI-Reviewer.
They support evidence teams from search and screening through extraction, appraisal and synthesis.
End-to-end workflows, human review controls, audit trails, collaboration and increasingly interactive outputs. Cochrane selected Nested Knowledge and Laser AI for a 2026 platform study.
What it does not replaceSelection for evaluation is not an endorsement, and the later Cochrane results are still pending. These systems help teams produce reviews; they do not automatically own a public conclusion or its ongoing corrections.
They assess influential public claims and reasoning.
Expert scrutiny, clear verdicts and excellent checks of whether citations support a claim.
What it does not replaceMostly reactive claim, article or book reviews; not a continuously updated, outcome-by-outcome evidence model with sensitivity analyses.
They help users find papers, summarize corpora or inspect citation context.
Scale and speed: discovery, multi-paper summaries, citation context and evidence filtering.
What it does not replaceThey are research instruments. They generally do not assume editorial responsibility for a durable public conclusion and its correction history.
They turn evidence and publisher content into point-of-care guidance or answers for clinicians.
Clinical context, continuous updating, graded guidance and rapid AI answers grounded in licensed journal content.
What it does not replaceDesigned for professional decisions—OpenEvidence is limited to verified US healthcare professionals—and not built as a public audit of polarizing narratives, topic-wide funding or result-changing calculations.
They expose post-publication criticism, corrections, retractions or participant-level data.
Essential checks on the scientific record and paths to independent reanalysis.
What it does not replaceThese are inputs to an evidence audit, not consumer-facing syntheses of what a disputed health claim currently means.
The complete sourced matrix is now public, with capability boundaries and strategic implications.
Open the full competitor matrix →A plain-language conclusion, certainty level and the most important caveat appear before the technical detail.
Readers can see how comparator, study design, funding or exclusion choices change the conclusion.
Material claims point to the study, method, data availability and revision history behind them; missing links are stated rather than inferred.
AI currently accelerates search, extraction and editorial checks; that does not make them independent human review. Sources, versions and confidence remain visible, while dual verification is still pending.
Databases, registries, gray literature and multiple languages
One study, even when it produces many papers
Methods and numbers linked to the original page
Design-specific tools; independent dual review remains pending
Alternative analyses, funding and sensitivity
New studies without erasing version history
The PECO question, eligibility, study-family rules, appraisal and synthesis remain public. Version 0.8 binds editorial completion to one exact successful automated receipt and requires separate English and Italian improvement summaries: a failed control, missing language or later same-day rerun keeps the state pending. Definitive searches, independent dual review and registration remain pending.
Executed queries, exclusions, extracted fields and calculations are exposed through the protocol, cards and dated receipts. The application source is not currently published as a public code release.
Funding and sponsor roles become queryable data, not footnotes.
Every evidence card can be challenged with sources; revisions stay public.
Every daily review records what was searched, what changed and whether the conclusion moved. An unchanged conclusion is still a result—not a reason to erase the check.
Audited the public evidence-type taxonomy against all 21 decision cards. The five cards assigned to the former ‘human studies’ category are cohort analyses or pooled observational evidence, while five randomized human-evidence cards already sit in the separate trials category. The broad label therefore implied an incorrect partition: that trials were not human studies, or that the category covered all human research. Renamed both the public filter and its internal identifier to ‘observational studies’ / ‘studi osservazionali’ everywhere the classification drives filtering, coverage counts and causal boundaries. The five observational, five trial, one mechanism and ten review assignments are unchanged; only the taxonomy became explicit. The default processed-meat/colorectal-cancer evidence mix now visibly reads one observational study, one mechanism and four reviews. Bilingual regressions reject the former internal identifier and pin all five observational assignments. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no source appraisal, claim verdict, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Closed a split-population state in the exact-selection summary. After an evidence-type filter, the rendered rows and source-unit mix used the visible cards, but the question label omitted the active evidence type and the headline count, publication span and evidence-design mix still described the broader meat/outcome match. The summary now names the active evidence type and derives all four metrics—card count, publication span, evidence-design mix and source-unit mix—from the same filtered card set. The default server-rendered selection therefore reads processed meat, colorectal cancer and all evidence, with six cards from 2015–2026, while subsequent filter changes cannot leave a stale count or date range beside a narrower table. The non-independence warning remains adjacent, and bilingual regressions pin the default state plus the single filtered-data dependency used by every metric. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no source appraisal, claim verdict, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Extended the exact-selection summary so the new source-unit classification is visible at the point where readers interpret a filtered evidence count, not only inside the complete register. The source-unit mix is computed from the currently displayed cards and changes with the evidence-type filter; zero categories stay absent instead of implying evidence that is not shown. For the server-rendered default selection—processed meat and colorectal cancer—the six matching cards are now visibly composed of three official assessments or evidence streams, two syntheses or guidelines, and one primary study or linked report family. The existing boundary continues to state that card counts are not independent study families, so the dynamic mix is descriptive rather than additive evidence weighting. Bilingual regressions pin the default 3/2/1 composition and the filtered-data dependency. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no source appraisal, claim verdict, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Audited the unit represented by every decision-source card because the public count could still be mistaken for a count of independent studies. Added a named source-unit field to the permanent register and study detail: three official assessments or evidence streams, eight syntheses or guidelines, four pooled or consortium analyses, four primary studies or linked report families, and two contextual method sources. The boundary now warns that reviews may contain primary studies shown elsewhere, pooled analyses may share cohorts, and one study family may include multiple reports; card counts therefore cannot be added as independent evidence units. These labels describe the current card structure and do not complete the pending study-family deduplication. Bilingual regressions pin all 21 assignments and the 3/8/4/4/2 census. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no source appraisal, claim verdict, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Re-audited the 21-card register after exposing the former ‘audit risk / causal status’ field and found that it conflated two distinct questions. Split it into an editorial audit-concern label and a separate causal-interpretation boundary on every permanent card and study detail. The 13 existing lower-concern and eight higher-concern assignments were preserved without rerating, while the causal boundary now follows the recorded evidence design: five observational human cards cannot establish causality alone; five randomized marker cards do not establish long-term clinical events; one mechanistic card supports plausibility without quantifying a human effect; and ten syntheses or assessments inherit their causal scope from included designs and the stated framework. The audit-concern label remains editorial triage, not a formal ROBINS-E, RoB 2 or AMSTAR 2 result. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no appraisal, claim verdict, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Audited the complete 21-card register against the method-assessment link in the public evidence chain. Every study already had an editorial risk category used by the filtered table and modal header, but the server-rendered permanent audit omitted it. Added a named audit risk/causal-status field to every permanent source card and the study detail, preserving the existing 13 ‘Some concerns’ and eight ‘High / non-causal’ labels without re-rating any source. The register boundary now states that this label is an editorial triage category, not a substitute for a source-specific ROBINS-E, RoB 2 or AMSTAR 2 judgment; this avoids presenting a mixed causal/risk signal as a formal risk-of-bias score. Bilingual regressions pin all 21 labels and the 13/8 boundary. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no appraisal, claim verdict, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Audited all 21 decision-source cards for a separate registration/protocol state. Registration information previously appeared only inside design, methods, funding or limitations, so the public source chain did not expose one consistent status. Preserved eight states already supported by the recorded audit: six exact identifiers (three PROSPERO records and three ClinicalTrials.gov records), the WCRF pre-specified review protocol without a separately recorded registry ID and the substitution review’s reported PROSPERO protocol without its identifier in the current audit. The other 13 cards now say that registration or protocol status was not separately verified and explicitly warn that this is not evidence that no protocol or registry record exists. The field appears in the complete server-rendered register and study dialog; bilingual regressions pin the 8/13 boundary and all six identifiers. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no registration timing was upgraded, no claim verdict, source assessment, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Audited the permanent claim and source fragments after making the evidence relation bidirectional. The target IDs were correct, but a direct URL or internal link could land on a closed HTML disclosure without revealing its audit content. Added one shared, tested rule that opens the targeted details element on initial page load, browser hash navigation and same-page fragment clicks; malformed, missing and non-disclosure fragments remain no-ops. Claim and source targets also receive a visible target outline without replacing keyboard focus styles. Pure regressions cover direct and nested targets, encoded fragments and invalid input, while rendered checks preserve all nine claim and 21 source IDs. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no claim verdict, source assessment, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Closed the reverse-navigation gap in the public evidence chain. The four exact internal links introduced in review 63 now also appear from the corresponding permanent source cards back to their claim audits: two from Gu 2023, one from UK Biobank and one from the EAS LDL consensus. Because two claims share the Gu card, these four backlinks occupy three distinct source cards; the other 18 cards explicitly state that none of the nine current claim audits points to them by exact match, without implying that those sources are irrelevant. Every source card also links to the public revision history. The reverse map is generated from the already validated claim-source IDs, so a separate relation cannot drift from the forward links; rendered regressions pin four backlinks, 18 explicit no-match states and 21 revision-log links in both languages. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no claim verdict, source assessment, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Audited the nine influential-claim checks against the complete 21-card decision corpus. Each claim already linked its external evidence, but none exposed whether that source also had a permanent internal audit, so the public claim→source chain stopped before the fields for method, comparator, absolute risk, sensitivity, funding/COI, data/code and integrity. Added stable claim fragments and a separate matching-internal-source block. Four claims have an exact internal match: the two 2023 diabetes claims link to Gu 2023, the healthy-user-bias claim links to the UK Biobank card and the LDL-oxidation claim links to the EAS consensus card. Five claims deliberately show ‘no exact match’ rather than being linked to a merely related mechanisms, country-comparison, historical or nutrient source. Module-level validation rejects duplicate or unknown study IDs, and rendered regressions pin nine claim fragments, four internal links and five explicit gaps in English and Italian. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no claim verdict, scientific estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Audited all 21 decision-source cards against the sensitivity-analysis link in the public evidence chain. Alternative models, subgroup checks and robustness findings previously appeared only inside findings, methods, limitations or the two interactive lenses, leaving no source-level state. Added a separate sensitivity/robustness field to the complete server-rendered register and study dialog. Ten findings already supported by the recorded audit are now explicit: comparator stratification, competing-risk modelling, regional/subgroup analyses, ROBINS-E sensitivities, time-varying BMI adjustment, multiple-testing correction, funding/comparator subgroups, a post-hoc high-responder subgroup, cross-design LDL triangulation and follow-up/sex sensitivity. Eleven cards now say no source-specific sensitivity or robustness analysis is separately recorded and warn against assuming none was performed. The field distinguishes a performed check from a stable result: several explicit checks attenuated, crossed the null, failed multiplicity control or were post hoc. Rendered regressions pin the 10/11 boundary in English and Italian. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no analysis or result was invented, no estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Audited all 21 decision-source cards against the comparator and absolute-risk link in the public evidence chain. Exposure prose previously carried some comparisons while the cards had no separate comparator or source-specific absolute-risk state. Added both fields to the complete server-rendered register and study dialog. Ten comparisons already supported by the recorded audit are now explicit—including current intake, zero intake, lower/no meat, named plant foods and three randomized whole-diet or product contrasts—while eleven cards say that no source-specific comparator is separately recorded and warn against assuming none exists. Only the JAMA pooled cohort retains a recorded source-specific absolute-risk estimate, with its 10-, 20- and 30-year model boundary; the other twenty cards state that no estimate is separately recorded and that this is not evidence of zero risk. Rendered regressions pin the 10/11 and 1/20 boundaries in English and Italian. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no comparator or risk number was invented, no estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Audited all 21 decision-source cards against the funding/COI link in the public evidence chain. The cards previously combined funder identity, author disclosures and sponsor role in one field, allowing a funding classification to look like a completed conflict assessment. Split the concepts in the complete server-rendered source register and the study dialog. Thirteen existing disclosure or sponsor-role findings remain explicit—including corrected disclosures, declared conflicts or no conflicts, external ties and stated sponsor involvement—while eight funding-only cards now say that conflicts and sponsor role were not separately verified; the fallback explicitly warns that funding class alone is not a COI finding. Rendered regressions require both labels on every source in English and Italian and pin the 13/8 information boundary. The 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no disclosure, sponsor-role or absence claim was invented, no risk-of-bias judgment, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Audited all 21 decision-source cards against the product requirement to expose data and code availability. The public model previously stored one combined data field: nine cards already embedded a code statement there, while twelve did not say whether code had been checked. Split the two concepts across the complete server-rendered source register, the 23 integrity records and the study dialog. The nine existing code findings remain explicit—public, request-based, inconsistent or no statement identified—while the other twelve now say ‘not separately verified in the current audit’, which is deliberately not presented as evidence that code is unavailable. Rendered regressions require a data field and an analysis-code status for every source in both languages and pin the 9/12 boundary. The same 16:59 UTC PubMed, Crossref/Retraction Watch, PubPeer and official-source competitor retrieval was not repeated; no availability claim was invented, no source assessment, estimate, certainty judgment, competitor conclusion or health conclusion changed, and independent human review remains absent.
Audited the structured editorial receipt behind the published competitor analysis. It required an explicit completed official-source sweep and consistent material-change flags, but the boolean did not identify the source register being asserted. Editorial receipt schema 6 now pins the exact COMPETITOR-LANDSCAPE.md corpus with its SHA-256 digest, 29 mapped product rows and 49 unique official-source URLs; finalization fails if the matrix is missing, malformed, empty, contains a product without an HTTPS source or repeats a source URL. The English and Italian competitor pages expose the same scope and full fingerprint, and regressions compare the rendered receipt with the structured status. This improvement reuses the completed same-day official-source sweep and the 16:59 UTC automated receipt: PubMed, Crossref/Retraction Watch, PubPeer and competitor retrieval were not repeated; no source assessment, estimate, certainty judgment, landscape conclusion or health conclusion changed, and independent human review remains absent.
Audited the private scheduler trail after the verified release and found that latest-run.json correctly described the most recent duplicate trigger, but no longer retained the path to the successful cycle’s narrative report. The status writer now mirrors only completed outcomes to latest-completed-run.json, while the launcher copies the non-empty final narrative to the stable latest-completed-report.md before recording success. Later already-completed, overlap or failure outcomes may still replace the last-trigger receipt without overwriting the last-success receipt. Unit regressions prove that a duplicate changes latest-run.json and leaves the completed mirror byte-for-byte unchanged. The 16:59 UTC scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated; no source, estimate, certainty judgment, landscape assessment or health conclusion changed, and independent human review is still absent.
Removed the last manually copied current-review dates from the evidence page. The hero now identifies the latest completed editorial review from the exact selected receipt; if a newer automated run is still pending, it keeps the last completed editorial date rather than presenting unfinished work as current. The integrity execution label and caveat derive from the automated receipt timestamp, are formatted in English or Italian at render time and expose semantic machine-readable values. Invalid or missing receipt dates now fail the build instead of silently falling back to stale copy. The 16:59 UTC scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated; no estimate, certainty judgment, landscape assessment or health conclusion changed.
Corrected a future-dating flaw before the first multi-day revision. Every rendered time element previously used one hard-coded 13 August value, so changing it for a later review would silently redate the complete history, while leaving it unchanged would misdate the new entry. Revisions now belong to explicit ISO-date groups validated at render time; a later day is prepended as a new group and existing groups remain untouched. The visible localized label stays separate from the machine-readable date, and regressions require all 55 current entries to expose the correct value while rejecting the old hard-coded JSX attribute. The 16:59 UTC scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated; no estimate, certainty judgment, landscape assessment or health conclusion changed.
Closed the gap between the workflow receipt and the public change it represents. Editorial receipt schema 5 now requires a positive public revision number, and finalization compares it with the complete revision count rendered from the evidence dashboard. The Atom entry includes the exact English and Italian editorial summaries, localized links to #revision-54 and a machine-readable public-revision category; a pending newer automated receipt falls back to the generic history and cannot inherit the previous change. Feed regressions cover both states and escaping. This same-day improvement reuses the completed 16:59 UTC receipt: no scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval was repeated, and no estimate, certainty judgment, landscape assessment or health conclusion changed.
Added a stable numbered fragment to every entry in the bilingual revision history, from #revision-1 through #revision-53, plus a visible localized permanent link and a machine-readable date. Prepending a later review preserves every existing number, so a citation to a prior decision keeps resolving to the same entry. Rendered regressions require the complete unique sequence in both languages. This same-day editorial improvement reuses the completed 16:59 UTC receipt: no scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval was repeated, and no estimate, certainty judgment, landscape assessment or health conclusion changed.
Audited the public path from the daily control to the deployed release. The protocol visibly linked the compact automated receipt, but the exact release manifest and Atom update feed were only discoverable to a reader who inspected page metadata or already knew their URLs. Added a sixth bilingual control item with the full SHA-256 source digest from the canonical release manifest, plus localized visible links to the JSON receipt, release manifest and Atom feed with their real media types and new-tab context. The displayed digest is the same value embedded once in every editorial page and enforced by the production gate. Rendered regressions compare it with the local manifest in both languages and require all three endpoints. This reuses the same successful 16:59 UTC receipt and does not repeat scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval. The competitor landscape, source assessments, estimates, certainty judgments and health conclusion remain unchanged; independent human review is still absent.
Extended the Open Graph versus JSON-LD audit to the English and Italian protocol and competitor landscape. All four pages correctly declared an Article type and localized language, but none emitted the publication and exact current modification times already present in its structured document. Added one shared publication constant for the protocol and one for the landscape, reused the dated editorial receipt for modification time, and exposed those values as Article metadata on all four routes. Rendered regressions now compare type, locale and both dates with each page’s JSON-LD. The protocol and competitor cards deliberately remain image-free because no page-specific asset exists; the fix does not recycle the red-meat image. This reuses the same successful 16:59 UTC receipt and does not repeat scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval. The competitor landscape, source assessments, estimates, certainty judgments and health conclusion remain unchanged; independent human review is still absent.
Compared the canonical English and Italian evidence pages with their Article JSON-LD and found a concrete social-metadata mismatch. The English page still emitted og:type=website from the parent layout, while the Italian override emitted no Open Graph type; neither exposed the publication and modification times already declared in its structured Article. Both canonical pages now emit localized Open Graph article metadata with the correct URL, site name, image, publication time and exact current editorial modification time. A shared publication-date constant and the dated editorial receipt drive both Open Graph and JSON-LD, and rendered regressions fail if their dates or article type diverge. Canonical and hreflang translation links remain unchanged. This reuses the same successful 16:59 UTC receipt and does not repeat scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval. The competitor landscape, source assessments, estimates, certainty judgments and health conclusion remain unchanged; independent human review is still absent.
Audited the server-rendered interaction semantics across the evidence audit, public protocol and competitor landscape. External sources consistently opened a new tab, but only a visual arrow signalled that behavior, so assistive technology could not reliably announce the context change. Added one localized English or Italian assistive note per page and associated every new-tab source link with it, covering decision rules, risk and comparator sources, claims, all 21 source audits, integrity records, revisions, surveillance records, method references and official competitor pages. All 26 evidence-page controls now also declare type=button so their behavior cannot silently become form submission if the interface is embedded or refactored. Rendered regressions enumerate every external link and button in both languages and fail on an undescribed new tab or implicit button type. This reuses the same successful 16:59 UTC receipt and does not repeat scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval. The competitor landscape, source assessments, estimates, certainty judgments and health conclusion remain unchanged; independent human review is still absent.
Audited the rendered social metadata for both protocol and competitor routes and found a concrete inheritance bug: their Open Graph titles were localized, but all four pages still emitted the parent red-meat Twitter title, description and image. A shared link could therefore preview the wrong product surface and language. Added route-specific English and Italian summary cards for the protocol and competitor landscape, and deliberately removed the unrelated red-meat image instead of presenting it as a page-specific asset. Rendered regressions verify the exact localized title and description, summary card type, absence of twitter:image and absence of the inherited red-meat title on every secondary route. This reuses the same successful 16:59 UTC receipt and does not repeat scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval. The competitor landscape, estimates, certainty judgments and health conclusion remain unchanged; independent human review is still absent.
Compared the newly enforced bilingual receipt rule with the public governance record and found that the site still called the protocol v0.7 even though finalization had gained a new fail-closed requirement. Published protocol v0.8 with a dated English and Italian amendment: both improvement summaries are mandatory, either missing or undersized field stops completion, and each protocol renders only its own language. A single shared version constant now drives the protocol badges, status blocks, evidence-page callout and JSON-LD, while tests fail if the visible and structured versions diverge. This reuses the same successful 16:59 UTC receipt and does not repeat scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval. The competitor landscape, estimates, certainty judgments and health conclusion remain unchanged; independent human review is still absent.
Audited the remaining bilingual protocol surface after localizing interface semantics and found a persistent workflow defect: the daily editorial receipt stored only one English improvement summary, so the Italian protocol rendered that English sentence inside an otherwise Italian control record. The finalizer now requires separate English and Italian summaries, writes both into a versioned schema-4 editorial receipt and fails when either is missing or too short. Each protocol selects its matching field, and regressions prove that neither language leaks into the other. The unattended instructions and documentation now preserve this requirement for future 08:00 cycles. This reuses the same successful 16:59 UTC receipt and does not repeat scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval. The competitor landscape, estimates, certainty judgments and health conclusion remain unchanged; independent human review is still absent.
An accessibility audit found that the Italian language switch was still announced in English, while the study results and method pipeline used generic labelled containers, the visual method sequence was not exposed as a list and competitor-matrix symbols had no explicit image semantics. The switch now has a localized accessible name; the study region is tied to its visible heading; the six method steps use a native ordered list; and every matrix symbol exposes its localized status. This reuses the same successful 16:59 UTC receipt. Scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated; the competitor landscape, estimates, certainty judgments and health conclusion remain unchanged. Independent human review is still absent.
Compared the new complete source register with the Article JSON-LD and found a concrete SEO mismatch: the page exposed 21 decision cards and 22 unique underlying publications or official sources, while structured metadata still declared five hand-picked citations. The Article citation list is now generated from the canonical 20-DOI integrity manifest plus the two unique IARC and WCRF official reports. English and Italian pages expose the same 22 unique URLs; tests fail if the manifest and JSON-LD diverge or a duplicate enters the structured list. Corrections and discussion records remain linked as integrity context rather than being misrepresented as decision-corpus studies. This reuses the same verified corpus and does not repeat scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval. No estimate, certainty judgment or health conclusion changed.
Completed the missing connection in the new crawlable source register. Every one of the 21 decision cards already had integrity coverage, but its permanent audit did not identify or link the corresponding Crossref/article or official-report row. All 23 integrity records now have unique stable fragments; each source card lists its exact record or linked publication family and the current status—no linked notice, known correction, formal discussion or official report. Multi-publication BeanMan and PRODMED2 cards link both records separately. Tests prove complete 21-card coverage, 23 unique bidirectional targets and the NutriRECS correction path in both languages. This is navigation over the same verified August 13 integrity evidence, not a new integrity search. Scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated, and no estimate or health conclusion changed.
Audited the server-rendered evidence page against the product promise and found that the initial selection exposed only six source rows while decisive fields for the full 21-card corpus—methods, limitations, funding and conflicts, and data or code availability—depended on client-side interaction. Added a bilingual complete source-audit register that is present in the initial HTML, uses native expandable sections and gives every source a stable fragment link. Each entry now keeps finding, design, population, exposure, methods checked, limitations, funding or sponsor role, data or code access and the original source together in one crawlable record. The existing interactive question filter and focused study dialog remain available. Today’s scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated, and no estimate or health conclusion changed.
Audited the durable prompt that the 08:00 session receives and found it still described production verification generically after the gate became stricter. The unattended instructions now name all twelve canonical routes and state that deployment success or HTTP 200 is insufficient. Before reporting completion, the agent must confirm the exact public receipt, isolated-clone release digest, Atom MIME and byte-identical feed, all seven per-page source markers, sitemap and robots; its final report must retain receipt, feed and page-identity results. The scheduler continues to enforce these postconditions independently, so prompt compliance is not the only control. Today’s scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated, and no estimate or health conclusion changed.
Closed the remaining partial-deployment ambiguity. The exact release manifest could be current while an individual HTML route was merely returning 200 from older cached output. Every English and Italian editorial page now embeds one machine-readable SHA-256 source marker generated from the verified release manifest. The completion gate fetches all seven HTML routes and requires exactly one marker equal to the isolated review’s local digest before promotion. Tests cover every route and reject a wrong digest, missing marker, duplicate marker or invalid expected digest. Status, feed and manifest retain their separate exact-identity checks. Today’s scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated, and no estimate or health conclusion changed.
Corrected a transaction-boundary bug before the next daily run. The feed verifier was initially anchored to canonical main, but canonical intentionally stays on the previous review while tomorrow’s work is built and published in an isolated clone. That would have rejected a correct new feed against yesterday’s receipt before promotion. The completion gate now passes the explicit review root and executes the verifier and receipt files from that same clone. A regression creates a next-day isolated receipt, proves it passes against the isolated root and proves the identical feed fails against canonical old state. The already-completed path remains pinned to canonical main. Today’s scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated, and no estimate or health conclusion changed.
Closed the production-side gap left after making feed freshness fail-closed. The completion gate now requires the live Atom MIME type and compares the entire production response byte-for-byte with the feed generated locally from the exact verified scientific and editorial receipt. A stale timestamp, altered count, pending/completed mismatch, changed wording, empty body or any other drift stops completion even when the route returns HTTP 200. Regression tests accept the exact feed and reject tampered content, a mismatched receipt and an empty response. Today’s scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated, and no estimate or health conclusion changed.
Turned the Atom feed’s pending behavior from an assumption into an independently tested gate. The feed now accepts ‘completed’ only when the automated control passed and the editorial date, time zone and exact automated timestamp all match the current receipt. A newer or mismatched receipt emits a bilingual pending entry, uses the newer timestamp and withholds integrity, competitor and health-conclusion outcomes from the older review—even if a malformed caller supplies the old completed object directly. Missing core receipt fields fail instead of producing a plausible feed. Today’s scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated, and no estimate or health conclusion changed.
Added an auto-discoverable Atom endpoint generated from the exact combined daily receipt—not from a second editorial copy. Its bilingual entry exposes PubMed and DOI retrieval completeness, newly eligible count, integrity, competitor and conclusion-change outcomes, editorial completion and the independent-human-review boundary. It contains no abstracts, retrieved record titles or personal medical guidance. The endpoint now belongs to the production completion gate and will update with each finalized daily receipt. Today’s scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated, and no estimate or health conclusion changed.
Narrowed the fail-closed boundary without weakening it. Before claiming the Los Angeles date, the user-local bootstrap now verifies that the project scheduler is readable. If the external volume is still unavailable, it writes the private failure receipt and alert but leaves the date unclaimed, so a later five-minute tick can try again before any scientific retrieval has begun. Once the scheduler is reachable, the atomic daily claim remains in place; a later failure is still not retried automatically, protecting against duplicate retrieval. A regression proves unavailable-then-recovered behavior. Today’s scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated, and no estimate or health conclusion changed.
Closed a diagnostic gap left by the new Los Angeles gate. The readiness command now refuses success unless the installed bootstrap and plist are byte-for-byte identical to the versioned files, both plists parse, the LaunchAgent is loaded, scheduled mode is present, and the hourly wake signal, five-minute interval and RunAtLoad are active. This turns a previously plausible ‘ready’ result with stale installed files into an explicit failure. The installed launcher passes the stronger check. Today’s scientific, Crossref/Retraction Watch, PubPeer and competitor retrieval was not repeated, and no estimate or health conclusion changed.
Corrected a scheduling-control mismatch: setting TZ for the process did not itself prove that launchd’s Hour=8 calendar event followed Los Angeles after the Mac changed time zone. The installed agent now emits periodic and calendar wake signals, while the user-local bootstrap reads the America/Los_Angeles clock, exits before 08:00 and atomically allows one attempt per Los Angeles date. Run-at-load and wake signals provide catch-up after login or sleep; a failed claimed attempt is reported but not retried automatically. Today’s completed scientific, Crossref/Retraction Watch, PubPeer and competitor checks were not repeated, and no health conclusion changed.
Connected the visible meat-type, risk-outcome, time-horizon and replacement labels to their button groups with explicit accessible names; linked the ‘how we got here’ toggle to its expandable panel; and increased the smallest language and risk-control text. Keyboard focus, pressed states and dialog focus management were already present and remain unchanged. This accessibility correction did not repeat today’s scientific, Crossref/Retraction Watch, PubPeer or competitor retrieval and did not change any estimate or health conclusion.
Corrected three claims that overstated the current product: changed ‘every sentence’ to decision-relevant claims, removed the implication that independent human review already has final authority, and replaced ‘public code’ with a precise availability boundary. The public product exposes executed queries, exclusions, extracted fields, calculations, receipts and revisions; it does not currently present the application source as a public code release, and independent dual review remains pending. The same 16:59 UTC receipt remains authoritative: no retrieval was repeated and no scientific estimate or health conclusion changed.
Published dedicated English and Italian routes for the complete competitor analysis already verified in today’s official-source sweep. The page separates eight product categories, links each product to its official source, exposes a capability matrix and states the narrower Onde Evidence hypothesis without claiming uniqueness or independent performance validation. COMPETITOR-LANDSCAPE.md remains unchanged because the sweep found no substantive market change. The same 16:59 UTC scientific receipt remains authoritative: Crossref/Retraction Watch retrieved 20/20 DOI records with no linked notice, indexed PubPeer searches produced no publication page, and the health conclusion did not change.
The single 16:59 UTC control retrieved four complete PubMed records. Three were overlapping records with unchanged exclusions. The new Cureus article discusses red and processed meat in young-onset colorectal cancer in Mexico, but its publisher full text confirms a narrative search rather than a registered systematic review and supplies no original red-meat effect estimate; it is now logged publicly as a discovery lead, with its no-funding/no-conflict declarations and absent original data or code. All 20 canonical DOI records were retrieved with no linked Crossref/Retraction Watch notice, exact-DOI web-indexed PubPeer searches produced no publication page, and known correction and publisher-discussion records remain visible. A current official-source sweep found no substantive competitor change. No record advanced to the decision corpus, and the health conclusion did not change.
Closed the remaining scheduling-control gap. A loaded macOS LaunchAgent now starts a fresh fail-closed Codex session at 08:00 America/Los_Angeles in the canonical Evidence repository. It refuses a dirty worktree or overlapping run, executes the complete scientific, integrity, PubPeer and official-source competitor workflow, finalizes the exact receipt, tests the site and publishes only the verified commit. The schedule still depends on the Mac being on, the user logged in, the external volume mounted, network access and valid Codex/Sites authentication. This operations change reuses today’s completed 15:17 UTC evidence receipt; it does not claim a new search, independent human review or a changed health conclusion.
Closed a provenance gap left by date-only matching. Each editorial receipt now stores the exact automated-control timestamp it reviewed, and finalization is rejected if its completion precedes that control. A failed automatic control or any later rerun—even on the same calendar date—therefore returns the public editorial section to pending until the new receipt is explicitly finalized. Today’s receipt remains complete for the 15:17 UTC run: 3/3 PubMed records screened, 20/20 DOI records retrieved, no new eligible source, no material integrity or competitor change and no change to the health conclusion. Independent human dual review remains incomplete.
Replaced manual receipt editing with an atomic finalization command. Every daily cycle must now declare screening counts, scientific and integrity outcomes, PubPeer mode, the official-source competitor result, the published improvement and whether independent human review occurred. The command checks these declarations against the date-matched automated receipt and refuses incomplete DOI retrieval, unscreened PubMed results, stale timestamps or contradictory competitor flags before writing anything. Today’s validated receipt confirms 3/3 records screened, 20/20 DOI records retrieved, no new eligible source, no material integrity or competitor change and no change to the health conclusion. Independent human dual review remains incomplete.
Added a second, date-matched layer to the public daily status. The automated section still proves PubMed and DOI retrieval completeness; the editorial section now records scientific screening counts, integrity and PubPeer review, the official-source competitor sweep and the bilingual improvement. It is returned as completed only when its review date matches the automated receipt, so tomorrow’s run cannot inherit today’s editorial completion. The receipt identifies the reviewer as the Codex editorial workflow and states that independent human dual screening, information-specialist review and external peer review remain incomplete. No scientific source, estimate, competitor finding or health conclusion changed.
Corrected a scheduling flaw in the operational PubMed layer. A one-day Entry Date query run at 08:00 could miss a citation added later that same calendar day, because the next run moved directly to the next date; PubMed also documents that Entry Date may be backdated to publication date for citations added more than twelve months later. Query v0.3 now uses Create Date—the field PubMed identifies for checking newly added citations—and retrieval v0.4 repeats the preceding calendar day. The schema 2 public receipt exposes the field, exact window and overlap. The corrective 08:17 run retrieved 3/3 complete records and abstracts; all three PMIDs were already screened exclusions, 20/20 DOI integrity checks passed with no linked notice, and no source, estimate, competitor finding or health conclusion changed.
Added server-rendered JSON-LD that describes only content visible on the site. Every route now exposes the Onde Evidence organization and website; the English and Italian evidence pages identify themselves as free Articles with MedicalWebPage context, canonical URLs, real publication and modification dates, language, translation links, publisher, topics and five visible core citations. The two protocol routes identify version 0.3 as a draft DigitalDocument with their real language, translation, source guidance and publication history. The root duplicate emits the same canonical article identity as the English evidence route. No ClaimReview rating, medical recommendation, person byline or breadcrumb was invented. No scientific or competitor conclusion changed.
Added a compact public status endpoint generated atomically by the unified daily runner. It exposes the America/Los_Angeles review date, execution time, PubMed result, retrieval, MEDLINE and abstract completeness, plus canonical DOI retrieval, request errors and linked Crossref update counts. It deliberately excludes titles, abstracts and the full child receipts. The protocol page renders the same status and states its boundary: a passing automated check proves retrieval completeness, not eligibility, methodological validity, competitor coverage or completion of human editorial review. No source, estimate, competitor finding or health conclusion changed.
Closed an operational gap between the daily PubMed and DOI-integrity checks. A single timezone-safe command now derives or accepts the America/Los_Angeles review date, runs both versioned controls sequentially, preserves their complete machine-readable receipts and fails if either child run fails or its completeness boundary is not met. The accompanying runbook keeps the non-automated work explicit: human eligibility decisions, publisher and PubPeer checks, official-source competitor research, bilingual editorial improvement, tests and verified publication. This command makes a Scheduled task deterministic; it is not evidence that the task has been activated in the ChatGPT interface. No evidence card, estimate, competitor finding or health conclusion changed.
Corrected a material bias in the competitor analysis: the map covered emerging AI workbenches but omitted five established or open review platforms. Covidence documents human-screened automation and human verification for extraction suggestions; Rayyan combines blinded collaborative screening with PICO, extraction and risk-of-bias tools; DistillerSR exposes configurable dual review, cell-level provenance, versioned audit trails and automated living imports; EPPI-Reviewer spans screening through synthesis, OpenAlex-assisted living updates and public visualisation; ASReview provides open-source, reproducible active-learning screening and simulation. These incumbents raise the operational baseline and weaken any workflow-as-moat claim. Onde Evidence’s narrower position remains public editorial accountability across claim fidelity, methods, funding, absolute risk, comparators, sensitivities and bilingual revisions. No evidence card or health conclusion changed.
Upgraded the operational retrieval protocol to v0.3 without changing the v0.2 scientific query. The machine-readable runner now combines official NCBI ESearch and ESummary output with EFetch MEDLINE, preserving the full abstract, publication type, language and full-record completeness behind each title/abstract decision. It fails explicitly if PubMed truncates the result set or a returned identifier lacks its full MEDLINE record. The 14:32 UTC run returned the same two already screened records, with 2/2 full records and abstracts, zero full-text forwards and no change to the corpus, conclusion or competitor map. This remains one operational PubMed layer; the definitive multi-database search, information-specialist review, dual screening and prospective registration are still pending.
Added a prospectively registered European Heart Journal meta-analysis separating processed and unprocessed red meat across cardiovascular disease and diabetes. Its CVD synthesis covered 43 studies and 4.46 million participants; the diabetes synthesis covered 23 T2D studies plus four gestational-diabetes studies and 1.76 million participants. Higher intake was associated with both outcomes, but every eligible clinical-outcome study was observational, no eligible randomized trial was found, and T2D heterogeneity reached 88.7% for unprocessed and 90.5% for processed meat. The card exposes ROBINS-E appraisal, dual screening, public funding, no declared conflicts, residual confounding, FFQ exposure error, English-only searching through June 2022 and the absence of a code statement. Crossref retrieved the new DOI with no linked notice; no exact-DOI PubPeer publication page appeared in indexed search. The daily PubMed rerun returned only two already excluded records, and no material competitor change was identified. The direction of the conclusion is unchanged; causal certainty is not raised.
Created a canonical manifest for all 19 article DOIs and ran a versioned Crossref v1 check against every record. The execution retrieved 19/19 records with no request errors; Crossref update-to metadata exposed zero linked retractions, expressions of concern or corrections. That empty field is only a monitoring signal: the known NutriRECS disclosure correction remains retained from its correction record, and the Nature Medicine Matters Arising exchange remains a separate publisher check because neither appeared in the original articles’ update-to metadata. PubPeer also remains a separate exact-DOI indexed search, not an automated absence claim. No source status or scientific conclusion changed.
Upgraded the public protocol to v0.2 with an operational PubMed query that uses Entry Date rather than publication date, so delayed database processing is not silently missed. A machine-readable runner now preserves the exact query, PubMed translation, date window, result count and identifiers. The first public receipt covered 12–13 August and returned three records: a narrative review of nutrition and cancer, a tuberculosis-burden model and a breast-cancer risk study. All three were excluded at title/abstract for prespecified reasons; none entered full-text review or the decision corpus. The definitive multi-database search, dual screening and preregistration remain pending. The scientific conclusion and competitor map are unchanged.
Rechecked all 12 displayed point estimates and their 24 confidence-interval bounds against model 3 in Table 2 of the pooled JAMA cohort analysis; all 36 numeric values match the publication. The interface now states that it does not fit or extrapolate a new model: it converts the published percentage-point risk difference to events per 1,000 by multiplying by 10. It also exposes the authors’ mean-covariate profile, the 10-, 20- and 30-year horizons, 500-sample bootstrap confidence intervals and the fact that statistical intervals do not account for residual confounding or exposure-measurement error. The Crossref record still exposes no linked correction, expression of concern or retraction; no exact-DOI PubPeer publication page was found in indexed search. No scientific estimate or conclusion changed.
Added the PRODMED2 study family as one decision card with two linked article records. In 36 completers, most pork-versus-legume biomarker differences were null; HDL and homocysteine moved in opposite interpretive directions, while improvements from habitual baseline occurred under both low-ultra-processed diets and cannot isolate the protein source. National Pork Checkoff and Meat Foundation funding, attrition, surrogate outcomes, post-hoc work and registry timing are explicit. The competitor map now separates living synthesis products from review workbenches: LivingMeta, Epistemonikos SK-E2E and LIvE materially overlap with public living output, while Nested Knowledge, Laser AI, Elicit and EVID AI cover end-to-end review workflows. Cochrane’s selection of two tools is recorded as an evaluation—not an endorsement. The corpus is now 20 cards and 19 article DOI records; the health conclusion is unchanged.
Published the first bilingual register of sources screened but not counted in the decision corpus. SWAP-MEAT is retained as relevant product-comparison evidence, but its animal arm combined beef, pork, processed sausage and chicken and cannot isolate either red-meat category. A 2026 mortality meta-analysis remains excluded because it aggregates all animal protein. Each row now exposes relevance, the exact exclusion rule, funding or conflict context, registration, data/code access and current integrity checks. The 19-card corpus and scientific conclusion remain unchanged.
Added FOOD-1, a 41-person randomized crossover comparison of two daily Beyond Burgers with two beef patties for six days per diet. The registered primary endpoint, TMAO, and LDL favored the plant product, while weight and NT-proBNP were slightly higher; the supplied plant patties contained about seven times the sodium of the beef patties. LDL and the other secondary outcomes were not multiplicity-corrected and remain hypothesis-generating. The audit also exposes four sequence changes, weak effective blinding, unavailable participant data and an unresolved timing discrepancy: the article reports enrollment from June 2020, while the registry was submitted in August and lists a November start. No health-outcome conclusion changed.
Added the BeanMan randomized study family: 102 healthy men compared 760 g/week of red and processed meat with 200 g/week plus non-soy legumes for six weeks. LDL cholesterol and weight favored the legume arm, while glycemic markers did not differ; a linked 2026 paper found no major differences in inflammation, TMAO or gut-barrier proxies. Vitamin B12 and iodine markers were lower. The card discloses public funding, company-provided foods, request-only data/code and that trial registration was submitted 24 days after the declared start. The short surrogate-outcome trial sharpens the comparator analysis but does not change the health-outcome conclusion.
Added explicit processed/unprocessed applicability to every source card and made the study table require an exact match for meat type, outcome and evidence design. Previously the meat selector did not filter the corpus, and an empty evidence-type result was filled with unrelated outcomes. The page now reports the exact card count, publication span and evidence mix, preserves true empty states, clarifies the 2019 trial as unprocessed-red-meat evidence and keeps LDL and funding meta-research as context rather than direct meat-exposure evidence. No scientific conclusion changed.
Published protocol v0.1 with the PECO question, evidence streams, eligibility criteria, planned databases, a versioned PubMed seed query, study-family linkage, dual-review process, appraisal tools, synthesis rules, update triggers and amendment log. The page states prominently that registration, complete searches and dual screening remain pending. No scientific estimate or conclusion changed.
Published outcome-specific criteria for evidence that could raise certainty, weaken or reverse the current conclusion, and evidence that would not move it alone. The rules follow Cochrane guidance on effect interpretation, heterogeneity and certainty, and WCRF evidence-judgement criteria. No scientific estimate or conclusion changed.
Added the 2025 processed-meat Burden of Proof analysis, including its public code, Gates Foundation funding statement and two-star weak/inconsistent ratings. The direction of the processed-meat signal did not change, but colorectal-cancer certainty moved from moderate to low–moderate because recent formal frameworks disagree on strength. The integrity register now also links the peer-reviewed critique and author reply attached to the 2022 unprocessed-meat analysis; these are discussion, not a correction or retraction.
Added OpenEvidence after verifying its point-of-care positioning, free access for verified US healthcare professionals and JAMA Network content agreement. Its speed, clinician distribution and licensed publisher corpus raise the benchmark for clinical answers. Onde Evidence remains differentiated by a public claim-to-source audit, explicit absolute-risk and comparator calculations, funding context, sensitivity choices and versioned corrections. No scientific estimate or conclusion changed.
Published identifiers, Crossref / Retraction Watch notice status, indexed PubPeer search results and data/code availability for all 16 evidence cards. Thirteen article DOI records were checked; no linked retraction or expression of concern was found, while the known NutriRECS disclosure correction remains explicit. The 2019 randomized-trial meta-analysis was also corrected from an erroneous journal label to Circulation.
Corrected the server-rendered language attribute, added a bilingual sitemap and crawler instructions, and improved keyboard behavior. Study dialogs now move and contain focus, close with Escape and return focus to the opening control. Filters expose their selected state, the study list no longer uses invalid table roles, and reduced-motion preferences are respected.
Added a preregistered substitution meta-analysis and a 2025 Finnish pooled-cohort study. The new comparator lens shows that replacing 50 g/day of processed meat with nuts, legumes or whole grains produced different relative estimates. It also states why those estimates cannot be turned into events per 1,000 without a shared baseline risk and time horizon, and why the observational design does not establish causality.
Added the study’s fully adjusted 10-, 20- and 30-year absolute risk differences per 1,000 people. The view keeps exposure contrast and implicit comparator visible, and flags that the unprocessed-meat/CVD association crossed the null in a competing-risk model. These are cohort estimates, not personal predictions or causal effects.
Added a processed-meat umbrella review, while keeping its low to very low certainty visible. Disclosed the NutriRECS correction, checked nine core DOIs against the current Crossref / Retraction Watch dataset, and reviewed indexed PubPeer signals. No retraction or expression of concern was found in the records checked.
First concept for onde.media · Pilot topic: red meat