Risk of bias: serious
Functional unblinding, subjective investigator attribution, and mismatched reporting windows can influence the measured outcome.
Certainty gate · framework v0.5
The strongest result is a bound, not a score.
Zero events were observed. The data are still compatible with a materially higher underlying risk because only 67 exposed participants have been studied in the direct stratum.
Observed0 / 67
Drug-related serious adverse events in the two direct MDD trials.
Upper limit53.6
Events per 1,000 at the two-sided 95% exact upper bound.
CertaintyVery low
Provisional outcome-level GRADE judgment for the underlying risk estimate.
To exclude >10‰n = 368
Requires 301 additional participants with zero observed events.
01 · Outcome-level certainty
The observed count is clear. The true risk is not.
Randomized evidence starts at high certainty, but this outcome is downgraded once for risk of bias and twice for imprecision. The result is a provisional very-low-certainty estimate of underlying risk—not a judgment that the treatment is highly dangerous.
95% exact interval
0–53.6
events per 1,000 administrations
The interval still permits approximately one drug-related serious event per nineteen administrations at its upper edge.
0100 / 1,000
Imprecision: very serious
Sixty-seven exposed participants are far too few to exclude uncommon serious harm with useful precision.
Indirectness: narrow match
The studies closely match screened adults with MDD receiving pharmaceutical 25 mg psilocybin in a highly supported clinical setting.
Missing evidence
Publication bias and unpublished safety data cannot be assessed until the full database, registry, citation, and author-contact search is complete.
02 · Fragility
One differently classified event changes the picture substantially.
| Analysis | Events / n | Risk / 1,000 | 95% exact interval / 1,000 |
|---|---|---|---|
| Primary pooled direct stratum | 0/67 | 0.0 | 0.0–53.6 |
| Leave out Yngwe 2026 | 0/50 | 0.0 | 0.0–71.1 |
| Yngwe 2026 alone | 0/17 | 0.0 | 0.0–195.1 |
| One event was missed or reclassified | 1/67 | 14.9 | 0.4–80.4 |
| Two events were missed or reclassified | 2/67 | 29.9 | 3.6–103.7 |
The primary calculation pools denominators only for this tightly matched direct stratum. It does not model between-study heterogeneity, which cannot be estimated from two zero-event studies.
03 · Information size
What would it take to make the zero genuinely informative?
These targets show the exposed sample required, assuming no events are observed, for the two-sided 95% exact upper limit to fall below a stated risk boundary.
| Upper bound sought | Total exposed n required | Additional beyond 67 |
|---|---|---|
| < 50 per 1,000 | 72 | +5 |
| < 25 per 1,000 | 146 | +79 |
| < 10 per 1,000 | 368 | +301 |
| < 5 per 1,000 | 736 | +669 |
| < 1 per 1,000 | 3,688 | +3,621 |
Important: these are statistical precision targets, not declarations of what level of risk is acceptable.
04 · Interpretation boundary
The model now distinguishes permitted claims from prohibited ones.
May report
- Zero drug-related SAEs were observed among 67 exposed participants.
- The exact 95% interval is 0–53.6 per 1,000.
- Certainty about the underlying risk is provisionally very low.
- The result applies only to the fixed controlled MDD scenario.
May not report
- “The risk is zero” or “proven safe.”
- Safety equivalence with niacin.
- Generalization to recreational or unscreened use.
- A conversion of this result into a universal 0–100 harm score.
05 · Next gate
Independent replication of the review process—not another author adjustment.
- Peer review and translate the complete search strategy.
- Run all databases, trial registries, citation searches, and sponsor/author queries.
- Have two reviewers independently screen, extract, and complete outcome-level RoB 2.
- Adjudicate the certainty profile and rerun the frozen analyses.
- Publish the PRISMA flow and every excluded full-text reason.
06 · Reproducibility files
Every judgment and calculation is inspectable.
Full analysisMarkdown · judgments and claim rulesGRADE profileCSV · outcome-level domainsSensitivity tableCSV · leave-one-out and missed-event analysesInformation targetsCSV · required zero-event sample sizesReview readinessCSV · complete, partial, and pending gatesAnalysis codePython · reproduces intervals and targets