Hypothesis-Generating Brief: Low dose naltrexone inflammation

Artifact

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

Reviewer panel scores

Review verdicts

Why

Review decision

To resubmit, address

1. Remove or reclassify the off-topic 'contextual_other' sources whose acronym 'LDN' denotes laparoscopic donor nephrectomy (Srivastava 2018, Khajeh 2023, Brunschot 2018, Pegat 2025) and Brunschot 2018; these are not low-dose naltrexone evidence and their inclusion as load-bearing contextual signals inflates disagreement counts and misrepresents the corpus.
2. Reconcile the source bundle against all inline citations: either add bundle entries for the missing author-year references or remove uncited citations. Maintain a 1:1 audit trail between prose and manifest.json as the Search Summary promises.
3. Reclassify Parkitny 2017 in the Findings Map to reflect the positive direction reported in the bundle (15% pain reduction, cytokine reductions) rather than 'null' or 'no extracted directional signal' in Dosing/Pharmacokinetics; flag it as mechanistic/pilot rather than dosing evidence if appropriate.
4. Tighten the abstract and Scope to state that the 'geroscience case' framing rests on a single Long COVID survey (Livieratos 2024) and indirect mechanistic plausibility, not on longevity RCT evidence; avoid any language suggesting the corpus supports anti-aging claims.
5. Disambiguate Moloney 2025 (conference abstract, salience network) from Moloney 2026 (full RCT report, MADRS null). Specify that Moloney 2026 is the direct-evidence trial rather than mixing the two.
6. Replace the generic Tensions and Gaps paragraph with specific, actionable gaps (e.g., 'Tsui 2024 and Bruun 2021 report short-duration pilot RCTs; an adequately powered Phase III efficacy trial for pain or inflammation endpoints is absent').
7. Verify the '0 mechanistic/model-system' claim in the abstract against the actual bundle composition, which includes McKenzie 2026, Toljan 2018, and Parkitny 2017 as mechanistic/dosing reviews.

Major issues

• Source-grounding concerns: several cited claims do not align cleanly with the bundle. Parkitny 2017 is reported as 'no extracted directional signal' in the Findings Map table for Dosing/Pharmacokinetics, but the bundle excerpt reports a clear 15% pain reduction and reduced cytokines — the manuscript understates this. Plank 2022 is cited as indirect with unclear direction, consistent with bundle; Moloney 2026 reports null efficacy (p=0.97) which supports the characterization but the cited_as variant (Moloney 2025 vs Moloney 2026) creates confusion across two distinct MDD/LDN RCT sources. Srivastava 2018 is labeled 'contextual_other' with effect_direction 'negative' but its negative signal is on surgical complications, not on LDN efficacy — its inclusion as a load-bearing 'negative signal' on LDN outcomes is misleading. Khajeh 2023 and Brunschot 2018 are nephrectomy/surgical-context papers bearing the acronym 'LDN' for laparoscopic donor nephrectomy — not low-dose naltrexone — and are misclassified as contextual LDN evidence. Pegat 2025 is similarly a length-dependent neuropathy diagnostic study unrelated to LDN. These misattributions inflate the 'contextual adjacent evidence' tier with off-topic sources that bear only the coincidental acronym 'LDN'.
• Several author-year citations in prose (McKenzie 2026, Leiber 2025, Radi 2023, Parkitny 2017, Moloney 2025, Frech 2011, Lim 2020, Moser 2025, Britton 2025, Driver 2023, Marcus 2024, Paula 2022, Isman 2024, Raknes 2018, Raknes 2020, Vatvani 2024, Nazir 2025, Bested 2023, Rungkitwattanakul 2025, Colomer-Carbonell 2022, Paulides 2022, Zapata 2025, Cabanas 2021, Bolton 2020, Sullender 2024, Ciwun 2024, Toljan 2018, Srivastava 2018, Brunschot 2018) are not present in the source bundle; only a subset of citations can be matched. The bundle contains 27 entries; the manuscript references materially more named sources. This breaks the audit trail the search summary promises.
• The 'Longevity' outcome class is mapped to a single indirect Long COVID survey (Livieratos 2024) — this does not constitute anti-aging or longevity evidence and the framing as a 'geroscience case' in the abstract overstates what the corpus contains.

Minor issues

• Britton 2025 is a case report with effect_direction 'unclear' in the bundle; the manuscript discusses it in the dosing section without flagging its anecdotal nature. Cabanas 2021 is labeled TRPM3/ME-CFS but is not in the bundle — its inclusion rests on inline citation only. The abstract states 'no sources classified primarily as mechanistic or model-system evidence', yet Parkitny 2017 and Toljan 2018 are explicitly mechanistic/immune-modulation reviews and studies. The 'Shae 2025 narrative review' style self-citation density and the unusual LDN-as-acronym confusion (laparoscopic donor nephrectomy mis-categorized as contextual LDN evidence) need editorial cleanup. The Tensions and Gaps section is generic ('Run adequately powered human studies') rather than surfacing specific within-source contradictions; it does not name the surviving trials needed. Pagination/study identification: Moloney 2025 and Moloney 2026 appear to refer to overlapping MDD/LDN RCTs and should be disambiguated in revision.

Reviewer note

This evidence map attempts to map a heterogeneous LDN-inflammation corpus across dosing, contextual, immune, and longevity outcome classes without collapsing to a single claim — which is the right posture for an evidence map on this topic. The Tensions narrative (e.g., null Bruun 2021 FINAL trial vs. positive signal in Paula 2022 vs. sparse Plank 2022 reporting) is faithfully heterogeneous and the abstract appropriately resists a pooled efficacy verdict. However, the manuscript has three substantive integrity problems that fall short of accept and require bounded revision: 1. Off-topic source inclusion. At least four 'contextual_other' entries (Srivastava 2018, Khajeh 2023, Brunschot 2018, Pegat 2025) appear to use 'LDN' to denote laparoscopic donor nephrectomy rather than low-dose naltrexone. Their inclusion as load-bearing contextual signals distorts the disagreement count (109 cross-study disagreements is partly inflated by misclassified sources) and misrepresents the LDN evidence base. 2. Source-grounding gaps. Roughly half the author-year citations in prose are not in the supplied bundle, which breaks the audit trail the Search Summary promises. A few in-bundle sources are also mischaracterized (Parkitny 2017 described as null/dosing when the bundle reports positive mechanistic findings; Moloney 2025 vs 2026 conflation). 3. Mild longevity overclaim. Framing a single Long COVID scoping review as a 'bounded geroscience case' overstates what one indirect observational survey on fatigue/PASC supports. The synthesis itself is adequate, the heterogeneity mapping is honest, and the limitations section correctly identifies the mechanism-to-clinic gap and absence of hard endpoints. With the four off-topic sources removed, the prose citations reconciled against the bundle, and the longevity framing tightened, this would be a competent evidence map. In current form it is partially supported and warrants revision.

Panel metadata

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

Route: consensus

Prompt: reviewer-v11-research-synthesis