Research Synthesis: NAD+ Effects

Artifact

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

Reviewer panel scores

Review verdicts

Why

Review decision

To resubmit, address

1. Re-classify every included source by whether its primary content is an NAD+ precursor (NR/NMN/Nam/NAD+/NRPT/RiaGev/trigonelline/acipimox) intervention study in humans or relevant preclinical models, versus an off-topic record matched only on the 'NAD' string. Remove or explicitly flag off-topic records (Han 2022, Zhao 2024, Simonis 2025, Kim 2017, Cho 2020, Gao 2025 hearing, Bai 2022 sperm, Mevenkamp 2024 MRS methods, Ren 2023 CNM-Au8 secondary readout) and rebuild the outcome-class denominators from the cleaned set. The current map's claim of a tiered clinical evidence profile is not credible while these records are pooled with on-topic RCTs.
2. Report per-outcome class the actual extracted direction counts (positive/null/negative/unclear/mixed) with named source attributions for every cell, and reconcile the abstract's '6 direct / 24 adjacent' split against the bundle after the cleanup. Direct on-topic RCTs (Martens 2018, Airhart 2017, Katayoshi 2023, Yi 2022, Liao 2021, Elhassan 2019, Connell 2021, Xue 2022, Vreones 2022, Membrez 2024 human arm, Christen 2026, Yu 2025, Holmes 2026, Simic 2020, Reyna 2026, Weijer 2015, Baichuan 2023 meta-analysis) should appear in the findings table with their actual reported directions, not be collapsed into 'no extracted directional signal'.
3. Reconcile the claim 'no sources classified primarily as mechanistic or model-system evidence' with the actual bundle, which contains Curran 2025 (animal meta-analysis), Membrez 2024 (multi-species mechanistic), Sun 2026 (mechanistic review), and Nobile 2026 (mechanistic review). Either move these to a clearly labeled mechanistic tier and adjust the direct/adjacent counts, or remove them and explain.
4. Replace the three Curran 2025-based 'surfaced tensions' with genuinely comparable within-outcome tensions (e.g., Martens 2018 vs Connell 2021 on NR/niacin and muscle/physiological endpoints; Yi 2022 vs Katayoshi 2023 on NMN dose-response in middle-aged adults; Baichuan 2023 meta-analysis vs individual null RCTs on weight/BMI). A cross-species, cross-population, cross-endpoint disagreement is not a meaningful tension to surface.
5. Provide a full in-manuscript References section mapping every cited_as string (including those in the additional-corpus list) to a bundle entry with DOI, year, and study design, so source-attribution is auditable without leaving the manuscript.

Major issues

• Outcome-class slice labels do not match the actual content of the underlying sources. The bundle contains many clearly on-topic NAD+ precursor clinical trials (Martens 2018 NR, Airhart 2017 NR PK, Katayoshi 2023 NMN, Yi 2022 NMN, Liao 2021 NMN, Elhassan 2019 NR, Connell 2021 niacin, Xue 2022 RiaGev, Membrez 2024 trigonelline, Vreones 2022 NR, Ren 2023 CNM-Au8, Yu 2025 NAD+ in HF, Christen 2026 NR/NMN/Nam comparison, Holmes 2026 NRPT menopause, Simic 2020 NRPT AKI, Reyna 2026 IV NAD+/NR, Weijer 2015 acipimox, Baichuan 2023 meta-analysis), yet the Findings Map places most of them under 'Contextual Adjacent Evidence' and codes the strongest signal as 'no extracted directional signal in 9/16 sources'. The map systematically underweights direct clinical evidence in favor of a 'contextual' classification, contradicting both the abstract's own 6 direct/24 adjacent split and the stated evidence-honesty note.
• Several included sources are clearly off-topic for NAD+ effects and appear to have been admitted by a string-match on the abbreviation 'NAD' (Han 2022 'non-accidental death' air pollution mortality; Zhao 2024 'non-anxious depression' SSRI acupuncture trial; Simonis 2025 'non-acute decompensation' cirrhosis ascites; Kim 2017 MB12066 beta-lapachone NQO1 PK with no NAD+ precursor outcome; Ren 2023 CNM-Au8 gold nanocrystals with NAD+/NADH as a secondary MRS readout; Cho 2020 CYBA/NAD(P)H oxidase exercise redox study; Mevenkamp 2024 31P-MRS methods paper; Bai 2022 sperm NAD+ observational; Gao 2025 NAD+ in sudden sensorineural hearing loss - a different clinical context). The evidence map's stated topic is NAD+ effects (presumably on aging/longevity given the domain_slug), and these off-topic records inflate the corpus, distort the outcome-class denominators, and undermine the claimed 'tiered clinical evidence profile'. The Search Summary's eligibility criterion 'Sources whose primary content addresses nad effects' is not enforced against these records.
• The single 'survived' positive signal is Frailty (n=1, positive=1, indirect=1) on the strength of a single indirect source, and the 'no extracted directional signal' is asserted for 9/16 of the largest slice without showing the underlying extraction. The map therefore reads as a protocol artifact rather than a faithful synthesis: a reader cannot tell from the manuscript alone why direct NR/NMN RCTs were classified as 'no directional signal' or 'indirect' or why off-topic air-pollution and acupuncture studies were retained.
• Tensions and Gaps lists only three example tensions (all involving Curran 2025 against Katayoshi 2023, Connell 2021, Gao 2025), and two of those pairings (Curran 2025 vs Katayoshi 2023, and Curran 2025 vs Connell 2021) appear to be fabricating disagreement: Curran 2025 is a meta-analysis of animal infectious-disease models, Katayoshi 2023 is a human NMN arterial-stiffness RCT, and Connell 2021 is a human niacin muscle RCT - these are not directly comparable populations, interventions, or endpoints, so labeling the directions as 'positive versus null/unclear' disagreements is a category error rather than a genuine surfaced tension.
• Source bundle contains 29 entries but the manuscript refers to '30 retained source(s)' and lists additional off-topic preclinical/animal sources in the Search Summary (Simic 2020, Simonis 2025, Weijer 2015, Airhart 2017, Reyna 2026, Zhao 2024, Elhassan 2019, Bai 2022, Holmes 2026, Christen 2026, Kim 2017, Sun 2026, Nobile 2026, Vreones 2022) that mix mechanistic/animal content with clinical content without distinguishing them in the outcome table. The evidence-honesty note claims 'no sources classified primarily as mechanistic or model-system evidence', which is contradicted by Curran 2025 (animal meta-analysis), Membrez 2024 (multi-species incl. C. elegans and mice), Sun 2026 (mechanistic review), Nobile 2026 (mechanistic review of NAD compartmentalization), Simic 2020 (AKI), and the cited animal/preclinical list itself.

Minor issues

• The Search Summary states 'No records were excluded at the gates instrumented for this run' and yet the admission funnel shows 40 sources with 'No extractable claims', 11 'None-only claim binding', 38 'Mixed partial-or-none', and 18 'Partial-only' were dropped between the 44 candidates and the 30 admitted. This contradiction is not reconciled.
• The Scope section repeats the abstract nearly verbatim and adds no additional boundary information (date range, population, intervention definition) that would make the landscape auditable.
• The reference list is incomplete: Christen 2026, Holmes 2026, Reyna 2026, Sun 2026, and Nobile 2026 are all in the source bundle but have 2026 publication years relative to a 2026-06-21 retrieval window; several do not appear in the in-text Findings Map or Tensions section, and no full reference list is provided to reconcile cited_as strings to bundle entries.
• The Limitations section conflates pipeline mechanics (unsafe source-level numerics excluded, comparability-across-topics framing) with substantive content limitations, which obscures the actual boundary of the synthesis.
• The Frailty outcome class with a single positive indirect source is described as 'hypothesis-generating' in the table but the abstract's closing sentence uses it as a stepping stone for the broader 'bounded geroscience hypothesis' framing, which slightly overclaims the single-source signal.

Reviewer note

Evidence-map review of 'Research Synthesis: NAD+ Effects'. The manuscript's framing as a tiered, hypothesis-bounding geroscience evidence map is undermined by a corpus-admission problem: multiple retained sources are matched on the 'NAD' abbreviation rather than on NAD+ precursor effects, and the outcome-class table systematically downgrades clearly on-topic human RCTs (Martens 2018, Yi 2022, Katayoshi 2023, Liao 2021, Elhassan 2019, Connell 2021, Xue 2022, Vreones 2022, Christen 2026, Yu 2025, Baichuan 2023) to 'no extracted directional signal' inside a 'Contextual Adjacent Evidence' bucket. The abstract's evidence-honesty note (6 direct / 24 adjacent / 0 mechanistic) is contradicted by the bundle, which contains animal meta-analyses and mechanistic reviews. The Tensions section fabricates disagreement between an animal infectious-disease meta-analysis and human NMN/niacin trials, which is not a comparable evidence pair. The Search Summary claims no post-screening exclusions while the admission funnel drops ~14 of 44 candidate sources. Net: the search summary is auditable in principle, but the findings are not faithfully attributed and the heterogeneity is not honestly represented. Revise after a corpus cleanup, a per-source direction table with named attributions, and a replacement of the Curran 2025-based tensions with within-outcome ones. Do not accept until the off-topic records are removed and the on-topic RCTs are classified by their actual reported directions.

Panel metadata

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

Route: consensus

Prompt: reviewer-v11-research-synthesis