The paper

Methods

Everything below was fixed in a registered plan before any number was extracted from a trial, except what the last part of this section, "What this version corrects", describes. The plan, its four codebooks, the coding done before registration, and the analysis code as registered are under plan/, unchanged. deviations.json lists every departure from the plan and every decision it left open, and the last paragraphs of this section summarize them. Every coder and reader was a separate instance of Claude Opus 5.5, working alone from its codebook and its input.

Question and groups. A trial is masked if every dose of ketamine or esketamine was given under general anesthesia, from its induction until the end of surgery, and awake if any dose was given while the patient was conscious: before induction, during a cesarean delivery under spinal or epidural anesthesia, or by an infusion pump after surgery. The design codebook counts a dose given with the induction drugs as given under anesthesia, though it may reach a patient who is not yet unconscious. A masked trial is strictly masked if its last dose is known to have ended more than 30 minutes before the end of surgery, so that the drug should have worn off before the patient woke; a trial with a dose in the last 30 minutes of surgery, at its end, or by an infusion run until its end, or whose timing is unclear, is not (late_dosing). The first primary question (P1) asks whether masked trials show an effect on depressive symptoms one to three days after the procedure. The second (P2) asks whether, among trials in surgery, the effect is larger with awake dosing.

Search and screening. On 2026-10-13 we searched PubMed (through E-utilities) and Europe PMC (titles and abstracts, which adds preprints) for ketamine or esketamine, a depression measure, randomization or placebo, and surgery, anesthesia, sedation, or childbirth: 703 and 719 records, 759 distinct. We searched ClinicalTrials.gov the same day for interventional studies of ketamine, esketamine, or arketamine that mention depression or mood: 634 registrations. data/search.json holds the queries, counts, and identifiers, and data/records.json each record's identifiers, title, journal, year, publication types, and the retractions, expressions of concern, errata, and updates PubMed links to it. Two coders decided from titles and abstracts which records might report an eligible trial, a randomized trial of either drug against an inert control around surgery or childbirth with a depression measure afterwards (the screening codebook). They agreed on 758 of the 759 records (Cohen's κ = 0.996), and a third reader kept the other. Two more coded each registration's eligibility and whether any dose was given to conscious patients (the registry codebook; κ = 0.99 for eligibility), finding 56 eligible registrations (data/registry.json).

Trials and their sources. The 138 kept records were grouped into trials by the registrations their abstracts name, by near-identical titles (a preprint and its journal version), and by four links, each with its reason (data/groups.json). With one report the search missed, which PubMed links to an eligible registration, and five eligible registrations whose only report is their posted results, this gave 126 trials. A trial's source is the first that exists of: its journal report's full text, open or obtained through a library subscription; its preprint's open full text; its journal report's abstract; its preprint's abstract; or, for the five, the registration's posted results. Full texts, abstracts, and the library's PDFs stay outside this bundle.

Design coding. Two coders coded each trial's design from its source's Methods section alone (the design codebook): eligibility; the arms, their drugs, doses, and the timing of each dose; the anesthesia; conscious_dosing (none, some, or unclear); late_dosing; the setting (surgery, cesarean delivery, or vaginal delivery); whether only patients with depression or depressive symptoms were enrolled (selected); the depression and pain measures planned, with fixed codes; registration; enrollment dates; and country. In the first round they agreed on eligibility for all 126 trials and on conscious_dosing for all 104 that both found eligible, and a third reader decided the nine trials with any difference. The 19 trials whose full text arrived before registration were coded again from it, with agreement on every compared field. Nine more arrived after registration and were coded again the same way (see the deviations below): the coders agreed on every compared field for the eight both found eligible, and a third reader found the ninth, T017, eligible. T072, a preprint of T054, was dropped. data/trials.json holds each trial's decided design and its integrity facts; the codings are in plan/coding/ and, for the round after registration, data/coding/design3/.

Extraction. Two coders copied, for every eligible trial, the numbers its source printed, in the form printed (the extraction codebook): the number randomized and analysed in each arm; every depression measure at every time, as a mean with an SD, SE, confidence interval, or range, a median with quartiles or range, a change from baseline, a count of participants above a threshold, or a difference between arms; pain at 24 and 48 hours; and each arm's psychotomimetic adverse events. Tables came before text and text before figures, and a figure only where its values were printed on it. A third reader decided every trial on which the coders differed in any fact, checking each against the source. In the first round the coders gave identical extractions for 52 of 106 trials and agreed on 4,567 of 5,357 facts (85.3%), and the third reader decided 54 trials. The nine trials coded again after registration were extracted from their full texts in the same way (eight of them for the second time), with 209 of 247 facts in agreement and six trials decided by the third reader. Their work is in data/coding/, and code/assemble_extraction.py rebuilds data/extraction.json from it. The author recorded in data/annotations.json which counts are of responders or remitters rather than of participants with depression, and which arm each between-arm difference subtracts from which, as each source's definitions and wording state.

Outcomes. For each trial and window, one depression effect: the measure the trial named primary, else the first it reports in that window of MADRS, HAMD, EPDS, PHQ, BDI, the HADS depression subscale, CES-D, GDS, Zung SDS, POMS depression, HADS total, or any other; the earliest day it reports in the window; and the first available form, in the order of form_order in data/parameters.json. The windows run from day 1 to day 3 after the procedure (early, primary), from day 4 to day 14 (middle), and from day 15 to day 90 (late). Effects are Hedges' g, ketamine against control, negative when the ketamine arm had fewer depressive symptoms (metafor::escalc, SMD), and several ketamine arms are combined into one (Cochrane Handbook, section 23.3.4; Higgins et al. (2019)). A mean and SD are estimated from a median and quartiles or range by Luo et al. (2018) and Wan et al. (2014); a log odds ratio becomes g on multiplying by $\sqrt{3}/\pi$ (Chinn (2000)), with its sign turned for counts of improvement; and a mean difference with its confidence interval gives g through the pooled SD the interval implies. T097 randomized esketamine two hours before induction, esketamine after induction, and placebo; it gives two comparisons, each with half its placebo arm (Cochrane Handbook 23.3.4): after induction against placebo, masked (T097m), and before induction against placebo, awake (T097a). Pain, the positive control, is the first of pain at rest, on movement, or unspecified, on day 1, else day 2. As the plan leaves out opioid consumption, it leaves out entries that are not ratings of the patient's pain, which the author marked in data/annotations.json: analgesic use (pump presses, rescue analgesia, or opioid consumption the design coding didn't list), an experimental pressure-pain threshold, and pain's interference with activity. The manipulation check takes, in each trial, the psychotomimetic event (hallucination, nightmare, dissociation, visual disturbance, agitation, or a combined psychiatric count) with the most participants in both arms together, chosen without regard to arm, among the events its source reports for every arm; each arm's count is the largest over the windows its source reports, at most the arm's size, which is the fewest participants the counts allow.

Models. Pooled effects are random-effects estimates by restricted maximum likelihood with Knapp–Hartung confidence intervals and tests (metafor::rma, method = "REML", test = "knha"; Viechtbauer (2010); Knapp and Hartung (2003)), reported with $\tau^2$, $I^2$, and the 95% prediction interval. Every interval in this paper is a 95% interval, and every test is two-sided. P1 pools the early effects of masked trials in every setting. P2 is the coefficient of awake in a meta-regression of the early effect on awake and selected among masked and awake trials in surgery; a positive coefficient means a smaller reduction with awake dosing, and the pooled masked and awake effects in surgery are reported beside it. Holm's correction covers the two primary p values. The plan calls a primary robust if its Holm-adjusted p is below 0.05, its p is below 0.05 among trustworthy trials and with strict masking, and its largest p with any single trial left out is below 0.05.

Integrity. A trial is trustworthy unless a retraction or expression of concern is linked to its report; its reports disagree on whether allocation was random; another trial of a different size, start, or setting gives the same registration number; it has no registration, in its design coding (from its Methods), in its extraction (from anywhere in its source), or among the ClinicalTrials.gov registrations whose linked reports include one of its records; its registration is dated after its enrollment began, by the month; two or more of its depression means fail the GRIM test, which asks whether a mean of whole-number scores given to $d$ decimals could come from $n$ participants, checked only where $n$ is known exactly and $n < 10^d$ (Brown and Heathers (2017)); or its early effect is beyond $|g| = 2$. Before extraction no kept report had a linked retraction or expression of concern, one trial's reports disagreed on randomization (T039: its preprint calls it a case-control study and its journal report a randomized trial), and three trials of different sizes, dates, and settings (T033, T044, T052) gave one registration number. Prospective registration shown by dates defines a stricter set.

Sensitivity analyses. Both primaries are repeated among trustworthy trials; among prospectively registered trials; with strict masking; without the four trials whose ketamine was given only by a pump after surgery; with full texts only (open, from the library, or preprints); with open sources only; with effects that needed no conversion; among trials enrolling only depressed patients, and only unselected patients (whose P2 models drop selected); with esketamine alone; and in the middle and late windows. Outside the plan, both are also repeated among masked trials whose doses all began after induction rather than with it, in every window, and without T011, whose own printed tests show that its printed dispersions are not standard deviations (data/annotations.json); the positive control is repeated without T011 too.

Other analyses. For pain, the masked and awake pooled effects and the awake-minus-masked coefficient in surgery, modeled as for P2. For psychotomimetic events, the Mantel–Haenszel odds ratio among masked and among awake trials, and a z test of their difference on the log scale; and, outside the plan, the same difference allowing the odds ratio to vary between trials, as a random-effects meta-regression of each trial's log odds ratio on awake dosing (Thompson and Higgins (2002)). Among masked trials, Egger's regression test for small-study effects (Egger et al. (1997)) and the PET-PEESE estimate, which takes PEESE when PET's intercept has p < 0.05 (Stanley and Doucouliagos (2014)); and the slope of the early effect on the log of the total dose in ketamine equivalents in mg/kg, esketamine counting twice. For the file drawer, of the eligible registrations whose primary completion date was at least 12 months before the search and that were neither withdrawn nor still enrolling, how many had a report (a PubMed record naming the NCT number that isn't a protocol, or a trial in this review whose Methods name it), how many had only posted results, and how many had neither. Within T097, esketamine before induction against esketamine after induction on days 1 to 3, as Hedges' g with a normal-approximation interval.

An exploratory check. One check is outside the plan: after an extraction coder noticed that two trials printed the same values, code/summarize.R lists every pair of trials that print an identical depression mean with an identical SD on the same scale on the same day.

Blinding the analysis. The code was tested on simulated trials and then, on the extracted data, with each trial's arms swapped at random (--blind, seed 20261013), before it was run once with the arms as randomized.

Deviations. Nine full texts arrived from the library after extraction had begun. Their trials, which had been coded and extracted from abstracts, were coded and extracted again from them, as the plan provides for texts that arrive before extraction; no outcome had been analysed with the arms as randomized. The plan didn't say how counts of responders or remitters enter (their sign is turned, and where a trial reports both kinds, the count of participants with depression is used), how to read a difference whose estimate the extraction put on the entry rather than among its values (it is read from either, and used only with the contrast its source states), or how to convert a percentage, a mean with a range, or a median change from baseline (each as its nearest planned form). The analysis also combines the ketamine arms that report at a time when a trial reports only some of them; takes, for psychotomimetic events, the largest count over the windows a trial reports, where the registered code added them and gave counts above arm sizes; and counts as reported a registration that a trial in this review names in its Methods.

Reproducing the work. code/run runs code/analyze.R and then code/summarize.R in the environment env/Dockerfile builds (R 4.6.1 with metafor 5.2-1 and jsonlite 2.0.0 from a dated CRAN snapshot), reading data/ and writing results/ in under a minute. Every trial's effect in each window, with its measure, day, form, and conversion, is in results/effects.csv; every model is in results/analyses.csv; the registration each eligible trial counts with, and where it was found, is in results/registrations.csv; the declared results are in results/R1.json to results/R8.json; and the descriptive counts are in results/D.json and results/S.json.

What this version corrects. The first version's three reviews found faults in the registered code and coding, unseen until then, and correcting them showed one more; deviations.json lists each. The integrity screen read registrations only from the design coding, taken from Methods text, so it missed T093's, which its extraction gives, and T040's, which ClinicalTrials.gov links to its report; registrations now come from the design coding, else the extraction, else such a link, which also finds three trials without an early effect, and a registration date counts only beside a registration number, so T049, whose report gives a date and no number, is no longer in the stricter set of prospectively registered trials. The plan's text counts a masked trial as strictly masked only if its last dose is known to have ended early, but its code also counted the trials whose timing was unclear; strict masking now follows the text. The manipulation check read an event count missing for one arm as zero; it now compares only events a source reports for every arm, which leaves out three awake trials. And pain entries that are not ratings of pain entered the positive control, because the design coding listed them as other pain measures or not at all; they are now left out. Neither primary result changes. The first version also described its masked trials as dosing only unconscious patients, which trials that dosed with induction don't show, and read P2 as excluding an awake advantage larger than its interval's bound, which a comparison between trials can't establish; the claims now say only what the trials show, and the analyses outside the plan named above answer the reviews' questions. Reading T011 again for this version, the author found that its printed test statistics don't fit its printed dispersions; it stays in every registered analysis, with the analyses without it beside them.

Results

Trials. Of the 126 trials, 109 were eligible: 42 masked, 64 awake, and 3 whose dosing was unclear. Most gave esketamine (78 trials, against 29 with ketamine and 2 with both), 83 were done in China, and 88 had a registration: 83 found by the design coding, 2 only by the extraction, and 3 only through ClinicalTrials.gov's link to a report. Their sources were open full texts for 80, library full texts for 18, preprints for 3, abstracts for 3, and posted registry results for 5. A depression effect could be computed for 82 trials in some window and for 51 in the early window, giving 52 comparisons there, since T097 gives two: 23 masked (3202 participants), 27 awake (4146 participants), and 2 unclear. Most of the eligible trials without an early effect reported depression only at other times; the rest printed no numbers the analysis could use for it, giving values only in figures, or a difference between arms without its estimate, its interval, or the arms' sizes.

P1, masked trials. Among masked trials, depressive symptoms one to three days after the procedure were lower with ketamine or esketamine than with an inert control: g -0.61 (confidence interval -0.81 to -0.4; p < 0.001; Holm-adjusted p < 0.001; 23 trials, 3202 participants). The trials differed ($\tau^2$ = 0.18, $I^2$ = 82.5%), and the prediction interval for a new trial's effect runs from -1.51 to 0.29. The estimate met all four of the plan's conditions: among the 17 trustworthy masked trials g was -0.48 (-0.63 to -0.33; p < 0.001); with strict masking, -0.67 (-1.04 to -0.3; p = 0.0018; 13 trials); and with any one trial left out, every p stayed below the threshold (largest p < 0.001). Outside the plan, among the 19 masked trials whose doses all began after induction rather than with it, g was -0.6 (-0.85 to -0.36; p < 0.001), and without T011 it was -0.6 (-0.81 to -0.38; p < 0.001; 22 trials).

P2, awake against masked dosing. Among trials in surgery (37 comparisons: 22 masked and 15 awake), the coefficient of awake dosing, allowing for depressed-only enrollment, was 0.12 (confidence interval -0.24 to 0.48; p = 0.51; Holm-adjusted p = 0.51). The pooled effect in surgery was g -0.62 (-0.83 to -0.4; p < 0.001) in masked trials and -0.45 (-0.77 to -0.12; p = 0.01) in awake trials. Leaving out any one comparison gave p values for the coefficient up to 0.72.

Positive control and manipulation check. Masked ketamine lowered pain one or two days after the procedure (g -0.55, -0.97 to -0.14; p = 0.012; 22 trials), as did awake ketamine (-0.33, -0.54 to -0.11; p = 0.0046; 32 trials). T011 gives the largest masked pain effect, from dispersions that are not standard deviations; without it the masked effect was -0.38 (-0.54 to -0.22; p < 0.001). In surgery the awake-minus-masked coefficient for pain was 0.32 (-0.21 to 0.86; p = 0.23), and 0.17 (-0.07 to 0.41) without T011. Ketamine raised the odds of each trial's most frequent psychotomimetic event in masked trials (Mantel–Haenszel odds ratio 2.11, 1.48 to 2.99; p < 0.001; 22 trials) and more in awake trials (3.63, 2.97 to 4.43; p < 0.001; 40 trials): the awake odds ratio was 1.72 times the masked one (p = 0.0083). That registered comparison takes one odds ratio for all masked trials and one for all awake trials, but the trials differ in which event is most frequent and in the windows over which they count it. Allowing the odds ratio to vary between trials, outside the plan, the awake odds ratio was 1.69 times the masked one (0.74 to 3.84; p = 0.21; 62 trials), so the difference is less certain than the registered comparison suggests.

Sensitivity analyses. Table 1 gives both primaries in each analysis the plan names and the two outside it. The masked effect's interval excluded zero in every one of them, and the awake-minus-masked coefficient's interval included zero in every one. In the later windows, the masked trials whose doses all began after induction gave g -0.61 from the fourth day to two weeks (-1.33 to 0.12; 10 trials), an interval that includes zero, and -0.6 from two weeks to three months (-0.97 to -0.22; 6 trials).

Table 1. Both primaries in each sensitivity analysis: the number of masked trials and their pooled effect, and the number of comparisons in surgery and the coefficient of awake dosing, each Hedges' g with its confidence interval, and its test. The first row is the main analysis; the two before the last two are outside the plan; the last two use the middle and late windows.

Within one trial. In T097 (Dai et al. (2025)), patients given esketamine two hours before induction had fewer depressive symptoms on the PHQ-9 on the first day after surgery than those given the same dose after induction: g -0.72 (-1.18 to -0.26; p = 0.0022; 77 participants).

Small studies, dose, and unreported trials. Among masked trials, Egger's test did not detect funnel-plot asymmetry (p = 0.7), which doesn't rule it out. PET's estimate of the masked effect, the intercept of its regression on the standard error, did not differ from zero, so the plan's rule kept it rather than PEESE's: g -0.47 (-1.23 to 0.29). The slope of the masked effect on the log of the dose was 0.38 (-0.19 to 0.94; p = 0.18), where a positive slope would mean smaller reductions at larger doses. Of the 56 eligible registrations, 39 were due to have reported: 16 had a report, 5 had posted results and no report, and 18 (46%) had neither that the search found.

Integrity. Of the 51 trials with an early effect, 32 passed the integrity screen and 19 failed it: 11 had no registration, 5 were registered after their enrollment began, 2 had two or more depression means that the GRIM test rejects (among 58 means it could check), and 1 reported an effect beyond the plan's bound. Failing the screen flags a concern; it doesn't show that a trial's data are wrong. The two trials GRIM flags (T088 and T116) are both awake trials. GRIM applies only to raw means of whole-number scores at the stated sample size, and T088's means come from its Discussion, beside a figure of adjusted means, so they may be adjusted means, to which it doesn't apply.

An observation outside the plan. The check for identical printed values compared 288 depression means with their SDs in 41 trials, and found 1 pair of trials that share any. T080 (Liu et al. (2025)) and T005 (Kudoh et al. (2002)) print 3 identical Hamilton scale means with identical SDs, two before surgery and one on the first day after it, though the two trials differ in country, year, and size. Neither has a registration, so both were already outside the trustworthy set, and neither enters any trustworthy-only estimate.

Discussion

The plan fixed, before any number was extracted, that the results would read as an effect that doesn't need awareness if P1 were significant after Holm's correction and the interval for P2 included zero, and both are. That reading claims more than the data can carry. P1 shows that ketamine and esketamine given under general anesthesia lowered depressive symptoms over the next three days, by g -0.61. P2 shows that, across trials in surgery, those that dosed awake patients did not show a larger effect; it doesn't show that being awake adds nothing, since it compares trials that differ in much else besides (Thompson and Higgins (2002)), and the one trial that randomized timing points the other way. The positive control rules out the reading that the masked doses were inactive. Masked patients had more psychotomimetic events than controls, and fewer than awake patients in the registered comparison, by a margin that a model letting trials differ can't confirm.

Earlier meta-analyses grouped perioperative trials by population, dose, timing, or anesthesia, but none we found by whether any dose reached a conscious patient. Two compared trials under general anesthesia with trials under spinal anesthesia (Guo et al. (2023); Wen et al. (2024)), which separates the anesthesia but not the patient's state at each dose, and one compared timings around childbirth, where most doses reach patients awake under spinal anesthesia (Hung et al. (2024)). Their pooled effects on depressive symptoms ranged from a standardized mean difference of −0.47 within a week of surgery, in a review of 17 trials of esketamine (Lou et al. (2024)), to −0.97 on the first day after surgery, in a review of 15 trials (Guo et al. (2023)); a review of 22 trials of ketamine or esketamine around childbirth found −0.94 at one week (Hung et al. (2024)), and one of 13 trials of perioperative esketamine in patients under general anesthesia found −0.92 (Yang et al. (2026)). The pooled effects here, g -0.61 in masked trials and -0.45 in awake trials in surgery, are smaller than most of these, and smaller again among trustworthy masked trials (-0.48). One effect per trial chosen by a hierarchy fixed before the numbers were seen, the integrity screen, and large trials with no effect, such as the trial in older adults by Mashour et al. (2018), may account for part of the difference; this work didn't test which. The one trial designed to mask ketamine with anesthesia found no difference in 40 depressed patients, with a mixed-model difference of −5.82 MADRS points (95% CI −13.3 to 1.64) and both arms improving (Lii et al. (2023)). That interval includes differences of the size pooled here for masked trials that enrolled only depressed patients (g -0.78), among which it is counted, so the trial and this review agree within the trial's precision. The largest of those trials reported more remissions with intraoperative esketamine three days after surgery (Zhou et al. (2026)).

The results leave other explanations open. Masked ketamine also lowered pain, so part of the effect on mood may follow from less pain, better sleep, or less opioid use in the days after surgery; trials that report both outcomes for each participant could test that path. Masked patients still had more psychotomimetic events than controls, so some may have noticed the drug's effects on waking; the strictly masked trials, in which the drug should have worn off by then, gave an effect of the same size (-0.67), which argues against the drug's action at waking but not against what patients recalled of it. Patients and assessors may also have guessed the arm from side effects after waking, and only 2 trials reported participants' guesses. Most trials enrolled patients without depression, in whom a lower score on a depression scale a day or two after surgery may reflect less postoperative distress rather than less depression; the effect among depressed patients and its persistence into the late window speak against a purely transient change.

The one randomized comparison of timing points the other way. In T097, esketamine two hours before induction lowered depressive symptoms on the first day more than the same dose after induction (g -0.72), and its authors proposed that anesthetic drugs blunt esketamine's effect (Dai et al. (2025)). Comparisons between trials, as in P2, cannot separate timing from everything else in which trials differ (drug, dose, population, procedure, and measure), and the within-trial comparison comes from one trial of 77 participants having one operation. Both can hold if the benefit of awake dosing differs between settings, as the heterogeneity among trials allows: P2's interval (-0.24 to 0.48) describes the average difference between the two kinds of trial under its model, not the effect of being awake in any one population.

The question that motivated the review is whether ketamine's effect needs the expectancy and meaning that its dissociative experience carries, which complicate blinding in trials of psychoactive drugs (Muthukumaraswamy et al. (2021)). The results show that an effect on depressive symptoms after surgery can occur without any dose reaching an awake patient. They don't show that the experience contributes nothing: masked patients still had more psychotomimetic events than controls and could have noticed the drug on waking or guessed their arm, and the comparison with awake dosing is between trials. That an effect occurs under anesthesia at all differs from what was found in chronically stressed mice, in which isoflurane anesthesia given with ketamine abolished both its dissociative-like state and its antidepressant-like effect a day later, leading the authors to conclude that wakefulness is likely required (Markman et al. (2024)). In people, ketamine given under general anesthesia keeps the high-frequency (β and γ) change in the EEG that it produces when given awake but loses its θ augmentation (Deverett et al. (2026)), so anesthesia changes some of the drug's effects on brain activity and not others; which of them the effect on mood needs is open. Ketamine's rapid effect in treatment-resistant depression (Zarate et al. (2006)) is measured in awake outpatients, without surgery or anesthesia, and these results don't test whether the same holds there. For perioperative care, ketamine or esketamine given under general anesthesia lowered depressive symptoms after surgery, with fewer psychotomimetic events than awake dosing in the registered comparison, though the prediction interval (-1.51 to 0.29) means a new trial in a different setting may find no effect. Of the registered trials due to report, 46% haven't, and the PET estimate's interval includes zero, so small-study and publication biases may have inflated the pooled effects.

What would settle the question is a trial that randomizes depressed surgical patients to the same dose before induction, while awake, or after induction, with a placebo arm; masks patients and assessors and asks both to guess the arm; and measures pain, opioid use, and sleep, so that the effect of awareness and the path through analgesia can be estimated in one population. Detecting a difference of g = 0.3 between the two ketamine arms with 80% power, two-sided at 0.05, needs about 175 patients per arm.

Limitations

Every coder and reader was an instance of one model, Claude Opus 5.5. Two worked independently on every trial and a third decided every difference against the source, but errors the model makes the same way every time would not show as disagreement, and no person checked the extraction. Before the third reader, the two extractions agreed on 85% of facts.

Some numbers come from sources others can't easily check: abstracts for 3 trials, posted registry results for 5, and library full texts for 18. Many effects needed conversion from medians, counts, or differences; the analysis restricted to effects that needed none (Table 1) gave the same answers to both questions. A difference between arms becomes g through the SD its interval implies, which is the raw SD only for an unadjusted comparison of two groups, so an effect from an adjusted difference may be off in size. A source can also print values that aren't what it says they are: T011's printed tests show that its printed dispersions are not standard deviations. The extraction coders noted that for T011 alone, and no check looked for it in every trial.

The search covered PubMed and Europe PMC. Of the 109 eligible trials, 83 were done in China, and trials published only in Chinese-language journals indexed in CNKI or Wanfang, or only in Embase, would be missing. The file drawer counts only ClinicalTrials.gov registrations, while many trials register with the Chinese Clinical Trial Registry, which was not checked; a trial that reported without its NCT number in PubMed or in a review trial's Methods counts as unreported.

Masking was coded from Methods text. A masked trial gave every dose under general anesthesia, but a dose given with the induction drugs (4 masked trials with an early effect) may have reached a patient not yet unconscious, and patients could still notice effects at emergence; whether the last dose came late was unclear for 5 masked trials, which are left out of the strictly masked set. The comparison of masked with awake trials in P2 is between trials and adjusts only for depressed-only enrollment; awake trials in surgery differ in other ways, including postoperative pumps (left out in one sensitivity analysis), drugs, and doses.

The integrity screen is coarse. It counts as untrustworthy every unregistered trial, including those from before trial registration was expected, and a trial that passes it may still hold errors; the values T080 and T005 share were found only by a check the plan didn't include. No risk-of-bias assessment by a standard tool was done.

Heterogeneity was large ($I^2$ = 82.5%), and the prediction interval includes zero. The outcomes are scale scores in the days and weeks after a procedure, mostly in patients without depression; this review didn't judge whether differences of this size matter clinically, nor pool adverse events other than psychotomimetic ones.

Nine trials were coded and extracted again from full texts that arrived after extraction began, and the analysis code differs from the registered code in the ways deviations.json lists. The first version's departures were all decided before the arms were unblinded; this version's corrections were made after its results and their reviews had been seen, and leave both primary results unchanged. The within-trial comparison rests on one trial.

Provenance

A person asked for a novel study, perhaps in psychiatry, rather than a re-check of earlier work. The agent, a Claude-family model (claude-opus-5-5), proposed several; the person chose this one and approved registering the plan. The agent designed the study, searched PubMed, Europe PMC, and ClinicalTrials.gov, wrote the plan, the four codebooks, and the code, tested the code on simulated data and with arms swapped at random, and registered the plan. To reach full texts, the person once passed a publisher's bot check in a browser, and obtained the PDFs a library subscription allowed for trials whose articles weren't open; the person did no coding. Separate instances of the same model screened the records, coded the registrations and the trials' designs, and extracted the trials' numbers, two working independently and a third deciding wherever they differed. The agent then ran the analysis, wrote the claims and the first version's paper, applied the hazard screen, and re-ran the analysis with the reference harness in a Linux container. After the first version's three reviews, the person asked the agent to assess them; the agent checked each finding against the code, data, and sources, proposed this correction, and the person asked for it. The agent wrote the correction's code, annotations, claims, and paper, ran the analysis again in the same environment, and applied the hazard screen again. The data are the published reports' numbers as data/extraction.json holds them and the ClinicalTrials.gov records retrieved on 2026-10-13; no full text or abstract is in the bundle. No person wrote, coded, analysed, or reviewed any part of the work.