Adjacent Evidence Brief: Mitochondrial DNA damage

Artifact

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

Reviewer panel scores

Review verdicts

Why

Review decision

To resubmit, address

1. Reconcile the admission funnel arithmetic so the bucket counts sum to the 15 admitted sources; remove or correct the '1/1 strict high-confidence' figure that contradicts the 15-source count.
2. Add the two missing sources (Chakraborty 2026, Chan 2012) to the Findings Map outcome-class tables, or explain why they were admitted but not mapped; ensure all 15 bundle entries appear in the body.
3. Recode outcome classes so mechanistic/animal studies are not silently treated as the corresponding clinical outcome class — e.g., Ng 2019 should be 'Mechanism/Longevity (C. elegans),' not 'Longevity outcomes'; Reid 2023 is a blood-based biomarker AD-risk study, not a clinical cognitive outcome study. Label these as biomarker/adjacent rather than outcome-equivalent.
4. Surface direction heterogeneity in the Findings Map itself, not only in a separate 'direction profile' row — explicitly state, e.g., that cardiometabolic is negative (Shimizu), muscle function is mixed/negative (Luo negative; Picca unclear), mechanism is mixed (Liang null; Ng unclear), and contextual adjacent evidence is null/unclear (Hsiao null; others unclear).
5. Write a substantive Tensions and Gaps section that names the actual disagreements: (a) mFRTA tension (Ng 2019: mtDNA damage does not determine C. elegans lifespan); (b) biomarker-vs-clinical-outcome gap (significant 8-OHdG/mtDNA-deletion signals without hard-endpoint translation); (c) direction divergence between negative cardiometabolic/cachexia findings and null/unclear contextual findings.
6. Correct the '0 direct human evidence' statement: several human observational sources are present (Picca 2019/2020, Roca-Bayerri 2020, Reid 2023, Chan 2012, Kennedy 2025 human TM cells). Distinguish 'direct interventional hard-endpoint' (which is indeed 0) from 'human evidence' (which is non-zero and should be characterized as observational/indirect).
7. Rescale the next-study design recommendation to match the actual evidence gap: a first-in-human biomarker validation or small interventional study, not a 200-per-arm 12-month RCT, unless the gap is explicitly framed as needing a phase-3-style trial.

Major issues

• Search Summary lists 10 queries but the cited source bundle (15 sources) spans heterogeneous topics (LRRK2/Parkinson's, Nrf2/ARE in mice, bronchopulmonary dysplasia, glaucoma TM cells, HIV brain aging, PUFA diet in mice, exercise in rats, C. elegans lifespan, cancer cachexia, etc.) without an explicit conceptual anchor tying them to a single evidence-map thesis on 'mitochondrial DNA damage.' The scope/boundary is not auditable as stated.
• The admission funnel (47 → 18 → 9/3/14/2/1 → 15) is internally inconsistent: 'Classified source candidates' is 18 but subtractions (9+3+14+2+1 = 29) exceed it, and the final 15 does not follow from the listed buckets. This makes the search/screening summary unauditable.
• Outcome-class mapping conflates domains: 'Longevity' is assigned to a single C. elegans mechanistic study (Ng 2019), 'Cognitive' to a single blood-based AD biomarker study (Reid 2023), and 'Cardiometabolic' to a PUFA-diet mouse study (Shimizu 2026). These are not the same construct as clinical longevity/cardiometabolic/cognitive outcomes; the map silently recodes mechanistic and adjacent biomarker studies as outcome-class evidence, which overstates outcome coverage.
• The Findings Map repeatedly states 'significant source statistic in N/N sources' while coding direction as 'unclear' for almost every source, despite the source bundle itself recording effect_direction (unclear/null/negative). The map does not surface the actual direction heterogeneity (e.g., Shimizu 2026 negative; Hsiao 2026 null; Luo 2024 negative; Liang 2022 null) at the finding level — it buries direction in a separate 'direction profile' row, weakening honest heterogeneity mapping.
• Chakraborty 2026 is listed in the source bundle (DOI 10.1016/j.jbc.2026.111156, Huntington's disease, F2,6BP) but does not appear in the Findings Map outcome-class tables or the Boundary-Condition Matrix, despite being an admitted source. This is an internal inconsistency between bundle and manuscript.
• Chan 2012 is in the bundle but not in the Findings Map tables either. The manuscript says 15 admitted sources but only enumerates 13 in the outcome-class sections.
• The Tensions and Gaps section says 'resolving 0 disagreement(s) narratively' — but there are at least two load-bearing tensions: Ng 2019 (no simple relationship between mtDNA damage and lifespan in C. elegans) vs. the broader framing of mtDNA damage as a longevity-relevant endpoint; and Shimizu 2026/Luo 2024 negative signals vs. the contextual adjacent evidence classed as 'significant.' A genuine Tensions section should surface these rather than assert zero disagreements.
• The Evidence Landscape states '0 provide direct human evidence,' but the source bundle includes multiple human studies (Reid 2023 human blood; Picca 2019 and Picca 2020 human muscle biopsies; Roca-Bayerri 2020 human brain samples; Chan 2012 human lymphocytes; Kennedy 2025 human TM cells). These are indirect/contextual rather than interventional, but the landscape conflates 'direct interventional hard-endpoint' with 'direct human,' which is misleading.
• The next-study design recommendation specifies 'at least 200 participants per arm' and 'at least 12 months' follow-up, but the underlying gap is mechanistic/animal; the recommendation is disproportionate to the actual evidence-base need (which is first human dose-finding/biomarker validation, not a large RCT).

Minor issues

• Abstract says '15 included source papers and 387 high-confidence extracted claims,' but the claims totals across outcome classes sum to 387 only if the boundary-condition matrix rows are not double-counted; the Breakdown is not clearly reconciled.
• Several sources have years beyond the stated retrieval window context (Shimizu 2026, Hsiao 2026, Chakraborty 2026, Kennedy 2025, Perez-Perez 2025) — retrieval date is 2026-06-28, so these are forward-dated but plausible.
• Repeated boilerplate 'Outcome-class note' and 'Source-context map' sentences add length without adding analytic content; trimming would improve readability.
• The Limitations section's third bullet about 'unsafe source-level numerics' is generic and does not materially constrain the conclusion.

Reviewer note

Evidence-map review. Scope: the manuscript declares a search/screening protocol and a thin-corpus evidence-brief framing, which is appropriate, but the admission funnel arithmetic is internally inconsistent (buckets sum past 18) and two bundle sources (Chakraborty 2026, Chan 2012) are not mapped in the body, so the auditable boundary claim is partially undermined. Source attribution: each mapped finding is tagged with a specific cited_as entry and DOI, which is good; citations match bundle entries on title/author/year, and source_grounding is adequate. Heterogeneity mapping: the map states there are '0 load-bearing cross-study disagreements,' which is the opposite of what the bundle actually shows — Ng 2019 directly contradicts the mitochondrial free radical theory of aging's longevity claim, Shimizu 2026 and Luo 2024 report negative effects, Hsiao 2026 and Liang 2022 report null, and the rest are unclear. The 'direction profile' row in the Boundary-Condition Matrix captures some of this but the prose and tables consistently code direction as 'unclear' and never surface the negative/null tension. The Tensions and Gaps section is one sentence and asserts 0 disagreements, which is a fidelity failure for an evidence map. Overclaim: the map correctly avoids a causal/clinical/policy headline and includes an evidence-honesty note about no direct interventional hard-endpoint evidence — good. However, outcome-class labels (Longevity, Cardiometabolic, Cognitive) are applied to single mechanistic or biomarker studies without flagging that this is adjacent/biomarker evidence, not clinical-outcome evidence; this is a mild form of the very collapse the review checks warn against. Limitations are present and partially material; gaps are actionable but the recommended RCT design is disproportionate to a mechanistic gap. Triage: not elite-tier accept; not fundamentally broken enough to reject. The manuscript is salvageable with bounded edits to the funnel, the missing-source reconciliation, the outcome-class coding, and a real Tensions and Gaps section. Recommendation: revise.

Panel metadata

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

Route: consensus

Prompt: reviewer-v11-research-synthesis