# Adversarial review log - magnesium-sleep (shipped review)

Canonical file: `magnesium-sleep.review.md` (package `hsef_v53_traffic_first_five_2026-10-12`).
This is a second, independent adversarial pass. The review already carries an approval section from a pass dated 2026-10-14
(findings A1 to A5). Findings here were re-derived from the sources and the score card, not copied from that pass; where an
earlier correction was claimed but is not present in the shipped files, that is recorded as a finding.

Level before -> after: Δ -> Δ (unchanged, conditional on M2). Confidence Moderate -> Low (proposed; not applied).
Findings: 2 Major, 6 Minor, 2 Editorial. PMIDs verified 59/59. Claims spot-checked 14. Validator: not re-run (recorded PASS - 0 fail, 0 warn, 31 pass).

## ADVERSARIAL REVIEW AND APPROVAL

**Authoring model:** Sol 6.1. **Adversarial reviewer:** Claude Opus 5.5 (Anthropic). **Date:** 2026-11-02.

**Scope of checks.** All 59 REFERENCES PMIDs verified against PubMed `get_article_metadata`: 59/59 resolve to the stated first
author, journal and DOI (seven references correctly print "DOI: not available"); 0 failures; no article is typed as retracted
or carries an expression of concern. Nine references print an issue year one apart from the PubMed e-pub year (refs 5, 7, 10,
15, 19, 26, 41, 42, 56), within the house convention. The attached correction 40567408 resolves as a Published Erratum to ref 1.
Fourteen load-bearing claims were spot-checked against PubMed abstracts and, for PMIDs 40918053, 39252819 and 23853635, the
PMC full text; the publisher page for 40918053 was read for its funding, disclosure and registration statements. The checks
concentrated on the single positive qualifying unit that sets the level (bisglycinate trial: n 155/153/134, ISI -3.9 vs -2.3,
p=0.049, d=0.2, 250 mg Mg plus 1523 mg glycine, diary completion below 10 per cent, registration wording), the threonate ISI
interaction (p=0.39) and wearable results, the oxide trial (magnesium oxide 2 x 250 mg elemental, 46 entered, 43 completed,
renin-powered, total sleep time p=0.37), the citrate trial's equal PSQI improvement, the newer threonate trial's endpoint
nulls, the diabetic four-arm single-blind design, the Mg+B6 versus Mg comparison, the rat deficiency/repletion groups, the
PCOS factorial, the postoperative threonate null, the elderly EEG crossover and the three-week deficiency study with unchanged
sleep architecture. Twelve of the fourteen matched their sources; the exceptions are M2 and m3. Scoring was re-derived: the
declared `primary-subclaim` rule gives SC2 = Δ; weakest-link Α and scope-matched Δ sensitivities are correct; the score card's
`decisive_perturbation` flag holds for exactly the eight printed inventory items ([1], [2], [3], [4], [7], [9], [19], [27]),
`discriminating_contrast` for 18, and one positive qualifying unit ([3]), matching the tally. The confidence band was re-derived
against the framework text and does not survive (M1). A targeted PubMed search for magnesium sleep/insomnia RCTs and
syntheses was run to look for omitted contradictory literature (m7).

**Findings**

| ID | Severity | Location | Original to corrected | Evidence |
|---|---|---|---|---|
| M1 | Major | AT A GLANCE table and "Belief versus evidence"; Subclaim matrix aggregation paragraph; "Final score, causal direction and M/S/D"; scorecard `final_score.confidence` and `confidence_rationale`; reader-summary.json `implications`; web-page.md | "The registered clinical subclaim SC2 supports Δ with Moderate confidence." to "The registered clinical subclaim SC2 supports Δ with Low confidence."; rationale "Multiple independent studies across 59 primary reports with partial replication, but marked methodological heterogeneity..." to "One positive qualifying study unit (PMID 40918053; d=0.2, p=0.049) without independent replication of that endpoint and formulation; the other controlled sleep trials differ in salt, dose, population and endpoint and include primary or global nulls (PMIDs 21199787, 39252819, 37520407). Evidence density and reproducibility set the band at Low, the framework's (Δ, low) pattern." | `hsef-framework.md`, Confidence band: Low = "Small number of studies; little independent replication"; Moderate = "Multiple independent studies; some replication". Its worked example reads "(Δ, low) - A single causal perturbation supports the mechanism, but independent replication is lacking; the Δ assignment could move either direction with further work." The review's own tally is 1 positive qualifying unit; web-page.md says the trials "are not interchangeable replications"; and the 2026-10-14 approval's residual states "The Δ rests on a single registered positive trial, so the level is one replication away from moving in either direction" - the Low definition nearly verbatim. A 2026 systematic review of 12 RCTs (PMID 42661485) grades certainty "low to very low". Level is unaffected; only the band changes. |
| M2 | Major | "Belief versus evidence"; aggregation paragraph; inventory item [3]; Cap audit first paragraph; SC2 matrix row; Final score; Momentum; steelman; evidence card [3]; reader-summary.json; web-page.md (15 occurrences of "prospectively registered" in the review; 11 describe the bisglycinate trial, at lines 16, 67, 71, 79, 95, 117, 231, 235, 239, 520 and, in the 2026-10-14 approval log, 366) | "A prospectively registered, placebo-controlled bisglycinate trial" to "A registered (DRKS00031494, registered 'at the time of recruitment' per the report), double-blind, placebo-controlled bisglycinate trial"; inventory [3] "A prospectively registered four-week bisglycinate trial" to "A registered four-week bisglycinate trial (registration timing relative to first enrollment not verified)" | The PMC full text and publisher page of PMID 40918053 say only: "This trial was registered at the German Clinical Trials Register (DRKS) (DRKS00031494) at the time of recruitment", with recruitment advertising from March 2023 and first enrollment on 20 April 2023; neither text calls the registration prospective or gives a registration date. The DRKS record could not be retrieved (robots-blocked). The review scrupulously flags retrospective registration for the 2024 threonate trial, so asserting prospective registration for the bisglycinate trial without evidence is an asymmetric overclaim, and it is load-bearing: the SC2 Δ ladder requires "a pre-registered primary outcome". Δ is retained conditionally; if the DRKS record shows registration after first enrollment, the adapted SC2 Δ crossing must be re-adjudicated. |
| m3 | Minor | "Final score, causal direction and M/S/D", Momentum; scorecard `extended_scoring_MSD.momentum.artifact` and `.justification` | "the prospectively registered bisglycinate trial DRKS00031494 and the newer threonate trial ACTRN12624000379516 used controlled clinical endpoints, objective wearables and allocation-guess checks" to "the bisglycinate trial DRKS00031494 used a controlled validated symptom endpoint without objective sleep measurement, and the newer threonate trial ACTRN12624000379516 added ring-derived wearables and an allocation-guess check" | Full text of 40918053: outcomes are questionnaires only, the authors state "exclusive reliance on self-reported data without objective verification", and no allocation-guess assessment is reported. The joint sentence attributes both features to both trials. |
| m4 | Minor | review.md Evidence section and steelman, REVIEWER NOTES, timeline; web-page.md; scorecard `causal_tally_audit.items_rejected` | "Rubric07 is the closest clinical comparator route and06 covers" to "Rubric 07 ... and 06 covers"; "about1.6points withd=0.2 andp=0.049" to "about 1.6 points with d=0.2 and p=0.049"; "Εcould" / "Γcould" to "Ε could" / "Γ could"; "bisglycinate4-week" to "bisglycinate 4-week"; "trial37520407" / "to39252819" to "trial 37520407" / "to 39252819"; web-page.md "serumMg<0.95mmol/L" to "serum Mg <0.95 mmol/L"; scorecard "the58original reports must not be marketed as58sleep RCTs" to spaced form (see also m5) | Known renderer defect. The 2026-10-14 pass recorded these as fixed (A3, 104 replacements; A5, the identical web-page.md string), but the shipped review.md (lines 46, 48, 227, 229, 267, 317), web-page.md and the score card still contain them. No value changes. |
| m5 | Minor | Decisive-perturbation inventory, tally paragraph; scorecard `causal_tally_audit.items_rejected` | "the 58 original reports must not be marketed as 58 sleep RCTs" to "the 59 original reports must not be marketed as 59 sleep RCTs" | Banner, REVIEWER NOTES, scorecard `provenance_header`, `evidence_inventory` (59 entries) and reader-summary all give 59. |
| m6 | Minor | reader-summary.json `limitations` | "59 primary reports with 6 selected partial-body examinations; density, heterogeneous controls and incomplete access set the confidence, not correctness." to "59 primary reports, one positive qualifying trial without independent replication, and heterogeneous salts, doses, comparators and endpoints set the confidence; reviewer access depth does not." | Lead policy and `hsef-framework.md`: confidence reflects evidence density and reproducibility. The 2026-10-14 pass corrected this wording in the score card (A4) but the public reader summary still prints it; the score card no longer contains "incomplete access". |
| m7 | Minor | REVIEWER NOTES (add); Knowledge gaps / Consensus reconciliation (optional) | Add: "Syntheses examined as context, not primary credit: a 2026 systematic review of 12 stand-alone-magnesium RCTs (PMID 42661485) rated certainty low to very low and concluded current evidence does not support oral magnesium as routine insomnia treatment; a 2021 meta-analysis of three RCTs in 151 older adults (PMID 33865376) found a 17-minute sleep-onset-latency reduction on low to very-low-quality evidence. A 10-week magnesium oxide (250 mg) versus placebo trial in women with PCOS (PMID 36620514) found no sleep-quality effect (p=0.85); like the postoperative threonate trial it is a population boundary, not a direct test of the registered poor-sleep claim." | Retrieved and verified via PubMed metadata (DOIs 10.1080/19390211.2026.2719670, 10.1186/s12906-021-03297-z, 10.1002/hsr2.1013). 42661485 was published 2026-08-28, before the 2026-10-12 cutoff, and bears directly on the confidence band. It shares an author (Lopresti) with refs 3 and 19. |
| m8 | Minor | Evidence and causal interpretation, after the oxide-trial paragraph; Upper ladder ("null objective outcomes remain") | Add: "The only placebo-controlled polysomnographic EEG study, a 12-person randomized crossover in healthy older adults (ref 23), reported more slow-wave sleep and higher delta and sigma power with magnesium; it is small, was not in poor sleepers and has not been replicated." Upper ladder "null objective outcomes remain" to "mixed objective outcomes remain (one small positive EEG crossover, null ring-derived outcomes in the newer threonate trial)" | PMID 12163983 abstract: slow-wave sleep 16.5 vs 10.1 min, delta and sigma power increased (p<=0.05). The review cites it only as a timeline milestone, so the objective-evidence summary omits the one positive objective result. This favours the hypothesis but does not change the level (healthy elderly, n=12, outside the poor-sleep scope). |
| E9 | Editorial | Existing 2026-10-14 approval section, rejected finding 4 | "Six references print the issue year where PubMed carries an e-pub year one apart (refs 5, 7, 10, 15, 19, 41, 56)" to "Nine references ... (refs 5, 7, 10, 15, 19, 26, 41, 42, 56)" | The sentence says six and lists seven. This pass's metadata comparison finds nine one-year offsets, all within the stated convention. Historical log; amend by note rather than rewriting the earlier pass. |
| E10 | Editorial | REFERENCES, ref 19 | Append "Correction: PMID 42959782 (2026-10-15; published after the literature cutoff; not examined)." | PubMed lists a Published Erratum to ref 19 (DOI 10.3389/fnut.2026.2010644) dated after the 2026-10-12 cutoff. Its content was not examined; it is not a retraction. Attach to the parent at the next edition. |

**Rejected findings (checked, no change).**
(i) Tally: 1 positive qualifying unit, 8 printed decisive items and 18 discriminating contrasts all match the score card's
`evidence_inventory` flags.
(ii) Aggregation: the preregistered `primary-subclaim` rule applied to SC2 gives Δ; weakest-link Α (SC3) and scope-matched Δ
are correctly reported.
(iii) SC1 quotes the metabolism-aging rubric's lifespan/healthspan criteria and ROS/NAD/senescence hard rules. This is the
declared governing rubric (`06-metabolism-aging.md`) quoted unchanged with an explicit applied adaptation and "Satisfied: no";
not template bleed, though it reads oddly to a lay reader.
(iv) Competing-model tournament: the rivals are substantive - coingredient/expectancy and biomarker-without-benefit models are
given full Δ standing and specific supporting results (equal citrate-placebo PSQI change, glycine in the salt). Not a strawman.
The nominal is no longer printed as its own best-supported rival.
(v) Role "reinforcing / stabilizing factor" follows derivation step 5; Momentum "Mechanistic only", Stability "Stable" and
Divergence "Contested" do not alter level or confidence.
(vi) Bisglycinate responder statement: the paper reports at least 6-point ISI reductions in 26 verum versus 15 placebo
participants with no between-arm test, so "no established clinically important between-arm responder result" is accurate.
(vii) Glycine handling: the 1523 mg glycine is salt ligand, not an added ingredient, exactly as the review says.
(viii) Δ itself: the adapted SC2 threshold (identical placebo for an ordinary supplement without unmistakable acute effects)
is explicitly declared and reasonable; the level is retained subject to M2.
(ix) Falsification patterns flagged "Additional downgrade: False" are correctly not used to lower the level.
(x) No template bleed from an unrelated hypothesis was found in the prose; the "cell-only" phrase in the open-label rule
application is generic framework wording.

**Scorecard change.** None applied. Proposed: `final_score.confidence` and `confidence_rationale` (M1), the Momentum artifact
(m3), and the spacing and count in `causal_tally_audit.items_rejected` (m4, m5).

**Residual limitations.** The DRKS registration date could not be retrieved, so M2 is an overclaim finding, not proof of
retrospective registration. Most references were checked at abstract depth; three were read in full text. The diabetic-trial
baseline-imbalance statement rests on the earlier pass's Table 6 reading and was not re-opened. The search for omitted
literature was targeted, not systematic.

**Approval:** Not approved as shipped. Approval is conditional on applying M1 and M2 (or verifying prospective registration
in the DRKS record and documenting it), and the Minor corrections. - Claude Opus 5.5 (Anthropic), 2026-11-02.