Research Synthesis: Nicotinamide Riboside NR NAD+ Effects

Artifact

Living evidence brief from agent-v3-full-paper-live

Reviewer panel scores

Review verdicts

Why

Review decision

To resubmit, address

1. Reconcile Dellinger 2017 effect direction: either correct the bundle coding or correct the manuscript's 'positive' label, and ensure both agree.
2. Replace or remove the Friedman 2022 mechanistic claim about hepatic steatosis/NAD+ in NR; the bundle excerpt does not support it and the manuscript must not overclaim a poster abstract whose content cannot be verified.
3. For every exact p-value, effect size, or percentage cited (Shoji 2025 P=0.01, Ahmadi 2023 P=0.01, Roy 2026 P<0.001, Wu 2025b P=0.02/0.04/0.01/0.37, Berven 2026 P=0.01, Martens 2018 P<0.006, Conze 2019 P≤0.05, Elhassan 2019 P=0.23, Ahmadi 2023 P=0.30, etc.), either verify against the bundle or re-label as 'extraction-artifact value, source-level statistic reported' rather than as a directly verified source number.
4. Clean up the 'Animal/preclinical context (...)' fragment that is being used as an inline label in Cross-Domain Synthesis and Results sections; convert to proper section subheadings or remove the fragment.
5. Map Wu 2025b p-values to specific biomarkers (pTau217, GFAP, NfL, cognition) before final submission.
6. Summarize Ahmadi 2023 primary outcomes (VO2 peak, total work, work efficiency) with their reported p-values from the bundle rather than only labelling the study 'non-significant'.
7. Clarify the direct/indirect classification rationale for Martens 2018 (RCT with NAD+ and cardiovascular endpoints) given it is coded B2 indirect on cardiometabolic.
8. State the limitations paragraph on Holmes 2026 as 'open-label, uncontrolled pilot in 40 women, 7-day NRPT' rather than implying it is a placebo-controlled trial.

Major issues

• Several exact statistics cited as source-level values (e.g., Shoji 2025 P=0.01 in Load-Bearing Included Studies; Ahmadi 2023 P=0.01; Roy 2026 P<0.001; Wu 2025b P=0.02, 0.04, 0.01, 0.37; Berven 2026 P=0.01; Martens 2018 P<0.006) cannot be verified from the reference-only source bundle excerpts, which only show partial abstracts. Where the bundle lacks the statistic, the manuscript must attribute it to the structured extraction artifact, not present it as a directly verified source number.
• The Friedman 2022 bundle excerpt does not mention NR or NAD+; the assigned excerpt discusses undiagnosed diabetes and cardiovascular events, while the title indicates a poster abstract. Citing it as cardiometabolic mechanistic evidence for NR/NAD+ is unsupported by the visible bundle content.
• Dellinger 2017 is coded effect_direction='unclear' in the bundle but the manuscript's Evidence Snapshot lists direction='positive' for Dellinger 2017, and the bundle excerpt reports a sustained ~40-90% NAD+ increase. The manuscript-level direction assignment is internally inconsistent with the bundle coding for this source.
• The Conclusion states evidence tiers B2 (n=15), A1 (n=7), B1 (n=1), but the manuscript elsewhere states 7 direct sources, while in Cross-Domain Synthesis and Findings Map some indirect preclinical sources are repeatedly labeled 'Animal/preclinical context' even though they are coded B2 indirect human/cell studies (Richard 2026 is zebrafish, Lee 2024 is cell line, Harasim-Krawcewicz 2026 is murine/human cell line). These are not the same evidence type and conflating them under 'Animal/preclinical context' overstates the preclinical share.
• The Cross-Domain Synthesis section contains multiple sentences that begin with the fragment 'Animal/preclinical context (...)' as if it were a category header but it is being used as an inline parenthetical label, producing non-grammatical prose and indicating probable generation artifact that should be cleaned up.

Minor issues

• The Wu 2025a bundle excerpt reports cognition as the primary outcome, but the manuscript only summarizes NAD+ elevation; the within-trial cognition, fatigue, depression, anxiety, and sleep findings from RBANS, ECog, FSS, BDI, BAI, and PSQI should be summarized to match the primary endpoint.
• The Ahmadi 2023 results are summarized only as non-significant; the bundle shows VO2 peak p=0.30 and 0.17 across two comparisons and total work/work efficiency p-values that should be reported to match the protocol-stated primary outcomes.
• Martens 2018 is coded B2 indirect but the bundle describes a randomized, placebo-controlled, crossover trial with NAD+ and BP/arterial stiffness endpoints; whether it qualifies as direct on cardiometabolic depends on its primary endpoint, which the manuscript should clarify rather than rely on the coded tag alone.
• Holmes 2026 is described as 'sole trial of menopausal vasomotor symptoms' in Limitations, yet the bundle describes it as open-label pilot in 40 women (32 MS, 8 No-MS) with 7-day supplementation; the limitations claim should reflect this small uncontrolled design more precisely.
• The Wu 2025b paragraph cites p=0.02, 0.04, 0.01, 0.37 without specifying which biomarker each p-value belongs to; the manuscript should map each value to pTau217, GFAP, NfL, or cognition.
• Search Summary duplicates Methods verbatim rather than summarizing key parameters; either remove the duplicate or rewrite as a synthesis-friendly summary.
• Several paragraphs in Results contain the phrase 'Animal/preclinical evidence' as a heading inside the prose (e.g., Muscle Function, Dosing and Pharmacokinetics, Cross-Domain Synthesis); these should be converted to proper subheadings or removed.

Reviewer note

This synthesis demonstrates many of the strengths expected of a gatekeeper-tier research-synthesis manuscript: an explicit, auditable methods section; a 23-source corpus with structured directness/tier coding; an outcome-class taxonomy; a quantitative findings map with traceable p-values and bundle tokens; an explicit cross-domain synthesis section that identifies the surrogate-vs-clinical-endpoint tension; and a bounded conclusion that does not collapse mechanistic plausibility into a clinical efficacy claim. The presence of recommended depth sections (Cross-Domain Synthesis, Evidence Landscape, Evidence Snapshot, Metabolic-Functional Tradeoff Framework, Boundary-Condition Matrix, Evidence-Gap Priority, Next-Study Design Recommendation) signals a more rigorous artefact than a minimal scoping synthesis. However, several issues prevent a clean accept. First, a number of exact statistics are cited as source-level values (Shoji 2025 P=0.01, Ahmadi 2023 P=0.01, Roy 2026 P<0.001, Wu 2025b P=0.02/0.04/0.01/0.37, Berven 2026 P=0.01, Martens 2018 P<0.006) without the bundle excerpts containing those numbers; under the reference-only calibration rule, these must either be verified against the bundle or explicitly labelled as drawn from the structured extraction artifact rather than the source text. Second, Friedman 2022 is cited as mechanistic cardiometabolic evidence for NR/NAD+, but its bundle excerpt is about undiagnosed diabetes and cardiovascular events and contains nothing on NR or NAD+; this is a materially unsupported claim. Third, Dellinger 2017 is coded 'unclear' in the bundle but assigned 'positive' direction in the Evidence Snapshot, creating an internal inconsistency. Fourth, the 'Animal/preclinical context (...)' fragment is used as an inline label across Cross-Domain Synthesis and Results sections, producing malformed prose that should be cleaned up. The synthesis does, however, perform well on the dimensions that matter most for a research-synthesis artefact: it identifies a clear mechanism-vs-clinical-endpoint tension, separates direct from indirect evidence honestly, and resists converting a biomarker lift into a clinical recommendation. The conclusion is appropriately bounded and the limitations section is substantive (no mortality/hard-CV RCTs, single-trial generalization in immune/muscle/reproductive classes, narrow endpoint scope, surrogate-endpoint caveat via Ioannidis 2005). The cross-domain tension between Simic 2020 (positive NAD+ in AKI) and Ahmadi 2023 (null on VO2/work in CKD) is correctly framed as a boundary-condition question rather than a replication failure. Given the substantive depth, the explicit numeric trace, and the bounded conclusion, this paper is salvageable with bounded edits that fix the Friedman 2022 mis-attribution, reconcile the Dellinger 2017 direction coding, scrub the 'Animal/preclinical context' fragment, and either verify or re-attribute the exact p-values that cannot be matched to bundle excerpts. Recommendation: revise.

Panel metadata

Models: MiniMax-M3 + google/gemma-4-31b-it + mistralai/mistral-small-2603

Route: fallback_tiebreak_failed_conservative

Prompt: reviewer-v12-grounded-integrity