Research Synthesis: Aerobic Exercise Training Effects

Artifact

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

Reviewer panel scores

Review verdicts

Why

Review decision

To resubmit, address

1. Fix the Immune and Inflammation narrative contradiction: either correct the count (the bundle contains at least Sloan 2018, Ghojazadeh 2024, Emmel 2025) and report a coherent per-source tally, or restrict the narrative to the single source actually discussed and move the others to the structured table only.
2. Reconcile the 'negative signals are not the dominant direction' abstract claim with the Swift 2012 review (effect_direction: negative) and the Konopka 2019 negative-finding paper on metformin–exercise interaction; either cite and integrate these sources into the cardiometabolic class or explain why they were excluded from the negative-signal claim.
3. Add the bundle entries that are admitted but uncited in prose (Konopka 2019, Swift 2012, Hansen 2021, Tanaka 2018, Davis 2017, Wyngaert 2018, Morita 2019, Han 2016, Prior 2014, Gelinas 2017, Raichlen 2020 only partially used) to the Findings Map, or document the omission criterion.
4. Resolve direction-coding inconsistencies between the manuscript's prose (e.g., 'direction=null' for Oliveira 2018) and the bundle's effect_direction for the same source, especially when a significant p-value is reported; either re-extract direction from the source text or explain the conservative-coding rule explicitly.
5. Replace the boilerplate 'limited corpus depth in this outcome class' limitation strings with per-class, materially distinct limitations (e.g., for the Deficiency Prevalence n=1 single-source slice, name the population and outcome specificity).
6. Tighten the Tensions and Gaps section: the current two-sentence version does not surface the high-density pairwise disagreement described in the Evidence Landscape (e.g., positive cardiometabolic RCTs vs. Swift 2012 negative-dose-response review vs. Konopka 2019 metformin-antagonism negative finding vs. Gelinas 2017 null COPD vascular finding).

Major issues

• Immune and Inflammation outcome class has a counting error: the manuscript states '1 included source were assigned to this outcome class' (and '1/1 sources') immediately after listing Sloan 2018, Ghojazadeh 2024, AND Emmel 2025 as source-level findings — three sources are presented but the narrative count says one, creating an internal contradiction.
• The bundle contains the canonical metformin-vs-exercise paper (Konopka 2019, 'Metformin inhibits mitochondrial adaptations to aerobic exercise training in older adults') and the systematic review by Swift 2012 explicitly tagged with effect_direction 'negative', yet the abstract claims 'negative signals are not the dominant direction in any outcome class' — this contradicts the bundle's own coding and an important negative signal in the cardiometabolic class.
• Several source-level findings listed under the Findings Map (e.g., Ghojazadeh 2024, Emmel 2025, ZURE 2025, Oliveira 2018, Santos 2024) are tagged with representative statistics (P=0.05, P<0.001, P=0.013, p<0.05) yet are coded as 'direction=unclear' or 'direction=null' in the manuscript prose, while the bundle's effect_direction field for the same sources is often coded differently (e.g., Ghojazadeh is a review with a significant TNF-α finding; Oliveira 2018 reports significant 6MWT improvements) — the manuscript's null/unclear coding is inconsistent with the bundle-extracted direction.
• Some cited_as identifiers in the manuscript do not cleanly match bundle entries (e.g., 'ZURE 2025' is used in the prose but the bundle entry is for a fibromyalgia BFR-aerobic study with no clear 'ZURE' author; 'Konopka 2019', 'Swift 2012', 'Gelinas 2017', 'Hansen 2021', 'Tanaka 2018', 'Wyngaert 2018', 'Morita 2019', 'Davis 2017', 'Han 2016', 'Prior 2014' are present in the bundle but never cited in the manuscript prose, despite being within the admitted 39 — several admitted sources are invisible in the map).

Minor issues

• The 'Findings Map' table header reads 'Evidence domain | Corpus slice | Strongest signal | Directness | Main limitation' but every row's 'Main limitation' is the generic string 'limited corpus depth in this outcome class' or 'single-source slice; hypothesis-generating' — not a meaningful per-class limitation.
• Search summary retrieval window is dated 2026-07-02 and one admitted source is Zhou 2026, which is plausible but should be flagged as a forward-dated record relative to the stated knowledge cutoff.
• The 'Source-context map' buckets (Aging/geroscience, Skeletal/muscle, Oncology/cancer, Pulmonary/rare-disease) overlap with the outcome-class taxonomy but use a separate, ad-hoc classification without explicit cross-walk.
• 'Significant source statistic in X/Y sources' is repeated as a strength-of-signal proxy across the map without defining what counts as a 'significant source statistic' or how it differs from the per-source p-values already listed.
• The Limitations section's second paragraph ('endpoint heterogeneity...') appears to be auto-generated boilerplate ('the contextual adjacent evidence, cardiometabolic and muscle function outcome classes, the cardiometabolic outcome class, and the immune and inflammation outcome class') and is internally redundant.
• Risk-of-bias framework is described in Search Summary but no populated risk_of_bias sidecar is actually reported in the manuscript, so the claim of RoB-2/ROBINS-I/AMSTAR-2 appraisal is unverified.

Reviewer note

This evidence map covers 39 retained sources across six outcome classes for aerobic exercise training effects, with explicit scope, a deterministic search summary, and an honest 'evidence-honesty note' about the 26/39 indirect/review/adjacent share. The taxonomy, source-context map, and per-class Findings Map are appropriate architecture for an evidence-map article type, and the manuscript correctly resists collapsing into a single pooled efficacy claim. Source attributions are present at the finding level and most cited DOIs match real PubMed records in the bundle, which keeps source_grounding at a competent level. However, several issues push this below the accept threshold. First, the Immune and Inflammation section has an internal counting contradiction (three sources listed, narrative claims one). Second, the abstract's claim that 'negative signals are not the dominant direction in any outcome class' is contradicted by the bundle itself, which codes Swift 2012 with effect_direction 'negative' and includes Konopka 2019 with a clear negative metformin–AET interaction. Third, direction coding in the manuscript prose (unclear/null) is inconsistent with both the source statistics reported in the same row and the bundle's effect_direction field for several sources. Fourth, several bundle entries (Konopka, Swift, Hansen, Tanaka, Wyngaert, Morita, Davis, Han, Prior, Gelinas) are admitted into the 39 but never cited in the manuscript prose, which weakens the 'every mapped finding is attributed to specific cited sources' check even though they appear in the bundle. Fifth, the per-class 'Main limitation' column is generic boilerplate rather than materially specific. Because the map's overall architecture is sound and the issues are concentrated in (a) a narrative count fix, (b) integration of bundle sources already in hand, and (c) tightening the Tensions and Gaps section to actually surface the disagreement map it advertises, this is a competent-but-fixable revise rather than a reject. The manuscript does not need a scope reset; it needs bounded corrections to specific sections.

Panel metadata

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

Route: consensus

Prompt: reviewer-v11-research-synthesis