Research Synthesis: Acute Exercise Effects

Artifact

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

Reviewer panel scores

Review verdicts

Why

Review decision

To resubmit, address

1. Reconcile the outcome-class assignments between the Findings Map, Evidence Snapshot, and Results narrative so that each quoted statistic is attached to the same outcome class as its source row; specifically, move Schmid 2021 narrative into the Mechanism outcome class and Dorneles 2020 narrative into the Immune/Inflammation or Bioactivity class, not Muscle Function.
2. Recode or reroute Harris 2008: the source is a primary RCT (16 overweight men, mixed factorial design), not a review; either change the directness to 'direct' / 'indirect' and reset the tier to match an RCT code, or justify the 'review' classification with an explicit secondary-publication reason.
3. Provide an auditable definition of the 'pairwise disagreement map' and the 353-tension count: state the disagreement rule (e.g., 'two sources coded as different effect direction within the same outcome class=dissent pair'), enumerate the dyad-level rule, and give a reproducible tally per outcome class. The current 8-row tension list is insufficient to support the 353 claim.
4. For any source where the bundle excerpt does not contain a claim that the Results narrative attributes to it (e.g., Schmid 2021 CXCL12 post-hoc statement, Dorneles 2020 'seven of eight contrasts' enumeration), insert the exact bundle token or remove the unverifiable numeric claim.
5. Add a population anchor to the bounded conclusion: state explicitly for which population (e.g., healthy adults, healthy older adults, or adults with stable CAD) the bounded interpretation holds, and what the boundary is for each other represented population (children, adolescents, sickle cell, poly-substance, hypertensive, breast cancer survivors).

Major issues

• The Evidence Snapshot and Findings Map are presented as quantitative evidence indexes, but several quantitative claims embedded in the Results narrative (e.g., Dorneles 2020 'seven of eight contrasts reaching conventional significance and one contrast not'; Schmid 2021 'post-hoc analysis revealed a significant increase in CXCL12 levels from baseline to 0 min'; the enumerated p-values for Windsor 2018) do not all carry exact bundle tokens sufficient to verify them against the reference-only source bundle, and the Schmid narrative appears to describe results not visible in the cited source excerpt.
• The Cross-Domain Synthesis repeatedly states a 'high-density pairwise disagreement map' contains 353 cross-study tensions, but the tension list in the Evidence Snapshot only enumerates 8 named tensions, and no audit table defines what counts as a tension or how 353 is derived. The number is asserted as a corpus-level property but is not auditable.
• Several Results subsections discuss specific studies (e.g., Bjorkman 2026, Kunz 2022, Schmid 2021, Dorneles 2020) inside the Wrong outcome class header (e.g., the Muscle Function section describes Schmid and Dorneles), creating misclassification between outcome classes and the Findings Map. A reader cannot trace the narrative claim to the row it belongs to.
• The Harris 2008 source is classified as 'review' with tier B1 and effect direction 'negative', but the bundle excerpt shows it is a primary randomized mixed factorial study in 16 overweight men, not a review. The directness/tier coding is internally inconsistent with the underlying source type.
• The abstract and conclusion describe a 'bounded evidence case' for acute exercise effects, but the corpus is heterogeneous in population (children, adolescents, healthy adults, older adults, breast cancer survivors, sickle cell patients, poly-substance inpatients, hypertensive adults) and the manuscript does not adequately translate 'bounded' into a population-specific constraint on the conclusion; the gating decision is deferred to the boundary-condition framework instead of being explicit about which population the bounded claim applies to.

Minor issues

• Several bundles (e.g., Heselton 2024) lack a PMID, and the doi-to-pmid mapping is not verified; the manuscript should flag these as verification-limited rather than treat them as equivalent to PMID-anchored sources.
• The outcome class 'Deficiency Prevalence' for Dorneles 2020 is a labeling choice that does not match the source's primary endpoint (in vitro immunoregulation by post-exercise serum); the label should be reconciled with the bundle's actual outcome_class field.
• The 'next-study design recommendation' specifies ≥200 participants per arm and 12-month follow-up but no source in the corpus supports those specific thresholds; this is a paper-level inference presented without citation, which is acceptable as editorial guidance but should be marked as such.
• The 'Metabolic-Functional Tradeoff Framework' is a paper-level organizing claim presented without bundling it to a specific source; the framework is useful but should be tagged as author-introduced rather than source-derived.
• Some duplication of p-values and effect directions between the Evidence Snapshot and Findings Map tables could be consolidated to reduce reader load without losing traceability.

Reviewer note

The manuscript is a long-form research synthesis on acute exercise effects with 43 sources, a stated source bundle, an explicit outcome-class map, and visible cross-domain tensions. The bones are appropriate for a gatekeeper-tier synthesis: an outcome-class table, a Findings Map, a Boundary-Condition Matrix, and an Evidence-Gap Priority list are all present, and the manuscript separates direct, indirect, review, and mechanistic evidence roles throughout. That said, several bookkeeping defects prevent an accept call. The Harris 2008 source is coded as a review with a B1 tier, but the bundle shows a primary 16-person RCT in overweight men; the tier/directness coding is internally inconsistent with the source type. Several Results narrative passages describe studies inside the wrong outcome class (Schmid 2021 and Dorneles 2020 inside the Muscle Function section), which breaks the cross-class boundary the synthesis relies on. The '353 cross-source tensions' figure is asserted as a corpus-level property but only 8 tensions are enumerated, and the disagreement rule is not defined. Quantitative claims inside the Results narrative (Schmid 2021 'significant increase in CXCL12 levels from baseline to 0 min', Dorneles 2020 'seven of eight contrasts reaching conventional significance') are not supported by the bundle excerpts and lack the exact bundle tokens needed to verify them. Reference-only source bundles are valid, but the manuscript cannot then layer additional numeric claims on top of them without traceability. What the manuscript does well: explicit denominator across outcome classes (n=26, 5, 5, 4, 1, 1, 1), a direct-A1 RCT roster with stat anchors (Duggan 2025 P<0.0001, Kunz 2022 P=0.039, Bauer 2025 p=0.001), appropriate hedging ('mechanistic plausibility coexists with mixed or sparse human-RCT evidence'), and a genuine attempt to keep mechanistic and clinical evidence separate. The Cross-Domain Synthesis is lengthier than it is novel, but it does address the population / dose / endpoint / time-horizon / comparator / measurement stack as a boundary map, which is the right structure for the topic. The verdict is revise, not reject: the manuscript has the right architecture and the right source set, and the defects are auditable bookkeeping problems (recoding, re-attaching narrative to the right outcome class, defining the disagreement count, and verifying a small set of overclaimed numerics). Bounded edits will resolve them. A population-anchored bounded conclusion is also needed so that 'bounded evidence case' does not read as a global qualifier over a heterogeneous population corpus.

Panel metadata

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

Route: fallback_tiebreak_failed_conservative

Prompt: reviewer-v12-grounded-integrity