Research Synthesis: NAD+ Intervention Nicotinamide Riboside NR Effects

Artifact

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

Reviewer panel scores

Review verdicts

Why

Review decision

To resubmit, address

1. Resolve the exact-numeric inconsistencies flagged in major_issue (a)–(g) by either reconciling to bundle text or downgrading the affected claims to narrative-only with explicit 'per-bundle excerpt' attribution; in particular, correct the Airhart 2017 directness/randomization characterization, the Berven 2026 directness coding, the Diaz-Urbina 2026 dose unit, and the Friedman 2022 excerpt mismatch.
2. Add a 'Direction Coding Justification' subsection explaining why ~16 of 20 sources are coded direction=unclear despite bundles reporting directional p-values, and either re-code or add a 'direction per source' column distinct from 'outcome-class direction.'
3. Surface the Wu 2025a post-hoc/unadjusted-for-multiplicity caveat in the Results prose when reporting significant p-values from that study.
4. Add a brief but explicit note that the four 'direct clinical sources' mix one non-randomized PK study (Airhart), two safety/Phase I PK frames (Simic, Dellinger in healthy older adults only for Dellinger — verify), and one Phase III-style RCT in long-COVID (Wu 2025a) — i.e., the direct tier is not equivalent to 'direct RCT-grade hard-endpoint evidence.'
5. Add a RoB-2 rating for Wu 2025b (crossover, double-blind RCT) and note the small sample sizes (n=15+15 for Roy; n=12 for Elhassan; n=6+6 for Berven) in the Results, not only in the Limitations section.
6. Tighten the Cross-Domain Synthesis so that each tension is anchored to a specific source-pair numeric or design fact rather than to the same generic 'severity 3 indirectness gap' repeated nine times.
7. Convert the 'metabolic-functional tradeoff framework' into either a falsifiable hypothesis tied to specific endpoints (grip strength per EWGSOP2, gait speed per Studenski) or recast it as interpretive scaffolding rather than a paper-level organizing claim.
8. Add a PRISMA-ScR-style flow-diagram or at minimum a bulleted source identification/screening/inclusion summary inline in Methods.

Major issues

• Multiple exact-statistic traces in the Results/Evidence Snapshot appear inconsistent with the cited bundles. Specifically: (a) the Airhart 2017 cohort is described in the manuscript Abstract/Evidence Snapshot as having a representative p = 0.001, but the bundle excerpt reports p = 0.03 for NR and p = 0.001 for NAD+, and the Airhart study is n=8 healthy volunteers on Days 1–8 plus a Day-9 PK — yet the manuscript repeatedly characterizes Airhart 2017 [bundle:18] as one of the four 'direct RCTs' comparable to Simic 2020 and Wu 2025a; the source itself is open-label, non-randomized, n=8, which is a meaningful design weakness relative to typical 'direct RCT' framing; (b) the Berven 2026 bundle excerpt indicates Phase I PK in n=6 healthy and n=6 PD on NR or NMN, whereas the manuscript describes it inconsistently as both an observational cohort and a trial-like source — directness coding is therefore unstable; (c) Wu 2025a bundle indicates the primary outcome was cognition and that the trial was placebo-controlled, but the Results text repeatedly frames Wu 2025b as 'crossover, double-blind, placebo-controlled' (consistent with the bundle), yet the manuscript variously compares Wu 2025a and Wu 2025b as if Wu 2025b is the primary cognition RCT — the two studies should be distinguished more carefully because Wu 2025a is in long-COVID and Wu 2025b is in MCI/SCD with a cognition primary outcome; (d) the Diaz-Urbina 2026 bundle states it is a preclinical study in neonatal rats using i.p. NR at 0.8 mmol/kg — not 0.8 mmol as the manuscript states; (e) Friedman 2022 bundle excerpt reports a diabetes/CVD statistic (OR 1.73) that has no relation to the cited title on NR/coca-cola hepatic steatosis, suggesting the excerpt is mismatched to the title; (f) Simic 2020 is described as a 'safety/tolerability design, not an efficacy trial,' yet the Results section cites three exact p-values (P = 0.05, P = 0.04, P = 0.002) consistent with the bundle, but the Abstract evidence-scope statement that 16/20 sources are indirect/adjacent/mechanistic counts Simic as one of only 4 'direct clinical sources' — together with Airhart being non-randomized and Berven being Phase I observational, the 4-direct-source claim is overclaimed; (g) Roy 2026 is a ketone + NR combination trial in MCI, and the manuscript repeatedly treats its 'white matter' signals as direct support for a cognition MIND-translation claim while the bundle shows the primary mechanism was ketone uptake, with NR as a cofactor — the boundary between ketone effect and NR effect is not isolated in the corpus.
• Mechanistic-to-clinical extrapolation is incompletely hedged in several passages: the Cross-Domain Synthesis section invokes Trammell 2016 to anchor a 'mechanistic preclinical' bridge and the manuscript repeatedly infers human NAD+ repletion from animal transcriptomic data; while hedging exists, the Discussion repeatedly describes Trammell 2016 as 'anchoring the downstream evidence base' in language that approaches a clinical claim.
• The Findings Map repeatedly labels effect direction as 'unclear' even when the source bundles (e.g., Wu 2025b pTau217 p = 0.02; Roy 2026 white-matter p < 0.001; Berven 2026 NAD increases in blood and brain) report clear directional findings — the conservative 'unclear' coding strips the synthesis of its own substantive signal and produces a flat manuscript in which almost every outcome is 'mixed'; a synthesis of 20 sources that resolves to 'no clear direction anywhere' is not integrating, it is flattening.
• Exact numerics are repeated across many paragraphs (P = 0.05, P = 0.04, P = 0.002 for Simic 2020; the Reyna 2026 P-list at P < 0.05, P = 0.04, P < 0.01, P = 0.01, P = 0.02, P = 0.03; the Roy 2026 P-list at 12 different p-values) but the manuscript does not clearly indicate whether these were multiplicity-adjusted or pre-registered primary endpoints — the Wu 2025a bundle itself notes 'post-hoc and unadjusted for multiplicity,' which materially changes the interpretation; this caveat is not surfaced in the Results.
• Several boundary-condition sentences are clinically graded without adequate hedging: e.g., 'an NR-induced change cannot be benchmarked against the ~0.1 m/s substantial-improvement threshold' is fine, but 'the source-traced significance in Berven 2026 suggests that translation from peripheral blood NAD+ elevation to measurable cerebral NAD+ change is not reliably achieved' over-interprets a single Phase I PK study with n=6 per arm.

Minor issues

• Title says 'full paper' which is a draft artifact; should be removed for publication.
• Phrasing such as 'DDD' is not used, but several sentences are formulaic and template-shaped ('Until then, the responsible conclusion is to preserve uncertainty…'); these can be tightened.
• The Methods 'accountability' paragraph and AI-use disclosure are useful but slightly overstate the certification: 'machine-verifiable, internally consistent, provenance-traced' is asserted rather than demonstrated in the manuscript text.
• The PRISMA-ScR framing is applied but the manuscript does not include a PRISMA-ScR checklist or flow diagram inline.
• Several studies dated 2026 are referenced as forthcoming/just-published; the manuscript should clarify which are preprints vs. peer-reviewed in the corpus.
• The 'Next-Study Design Recommendation' of ≥200/arm and ≥12-month follow-up is a specific claim that should be defended or sourced; it appears as an unsourced editorial suggestion.

Reviewer note

This is a structurally ambitious research-synthesis manuscript with a 20-source corpus, an explicit PRISMA-ScR-style Methods section, an Evidence Landscape with per-source Findings Map, a numeric Results table, a Cross-Domain Synthesis with multiple boundary-condition paragraphs, and clear separation between mechanistic/preclinical and clinical evidence. The recommended depth sections (Background, Inferential Bridge through the Cross-Domain Synthesis, Quantitative Evidence Index via Findings Map and Results tables, Limitations, Gaps, Next-Study Design) are all present, which signals gatekeeper-tier intent. That said, several substantive issues keep the manuscript at the revise rather than accept threshold. First, exact-numeric claims in the prose do not all line up cleanly with the bundle excerpts provided: Airhart 2017 is repeatedly characterized as a 'direct RCT' although the bundle clearly states it is non-randomized and open-label with n=8; Diaz-Urbina 2026 is described with a 0.8 mmol dose that contradicts the bundle's 0.8 mmol/kg i.p. figure; the Friedman 2022 bundle excerpt reports a diabetes OR rather than anything on NR/steatosis, signaling a possible excerpt mismatch. Second, the manuscript codes 'direction=unclear' for nearly every source even when bundles report clear directional effects (e.g., Wu 2025b pTau217 reduction at p = 0.02), which flattens the synthesis into 'mixed everywhere' rather than integrating. Third, several boundary conditions are stated without sufficient hedging or with a clinical register that outruns the underlying evidence (e.g., inferring human cerebral NAD+ translation from a Phase I n=6+6 PK study). On the positive side, the manuscript's overall conservatism is appropriate and consistent with the geroscience evidence base in 2026; the explicit separation of mechanism, biomarker, and clinical endpoints is the right epistemic frame; the limitations section is substantive and specific; and the Gaps Identified and Next-Study Design Recommendation sections are actionable rather than generic. The Discussion's 'mechanistic plausibility coexists with sparse human data' is the correct verdict for this corpus. Verdict: revise — the manuscript is close to gatekeeper-tier and is salvageable with bounded edits focused on (i) reconciling exact numerics to bundle text, (ii) defending or softening the '4 direct clinical sources' framing, (iii) addressing the direction-coding choice, and (iv) tightening the Cross-Domain Synthesis so that the nine repeated 'severity 3 indirectness gap' rows become a few concrete, source-pair-anchored tensions. Once those edits land and the Wu 2025a multiplicity caveat is surfaced, the manuscript is a credible accept candidate.

Panel metadata

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

Route: fallback_tiebreak_failed_conservative

Prompt: reviewer-v12-grounded-integrity