How it works
How PsychDoc works
PsychDoc takes one research question at a time. It finds candidate papers itself by following citations, sorts them by written rules, turns what they report into claims you can check, and keeps scanning for new work. This page shows each step, who or what does it, and what is planned next. The numbers come from the synthesis Does one dose of psilocybin rewire the brain?, version 2.2.
- CodeRuns on its own, on a schedule or when the site is built
- AI draftAn AI model drafts the decision and its reason
- PersonA person approves before anything is published (today, the publisher)
- PlannedNot built yet
The whole system
Eight steps, and a loop that keeps the search going
Step 2 · Find
It looks for papers itself, by following citations
Every Monday at 06:00 UTC a scheduled job asks OpenAlex, an open index of scholarly works and the citations between them, two questions about each seed paper: which papers does it cite, and which papers cite it, keeping the 100 most-cited. Seeds are the papers the synthesis already rests on. Any paper linked to a seed becomes a candidate.
The last accepted scan, 28 September 2026
5 seeds produced 521 candidates; 34 of them link to two or more seeds. The synthesis now lists 24 seeds, so the next scan starts from more.
Scans are now saved only when every seed resolves and every request is answered. A partial scan saves nothing, so the last accepted one stays in place. Each saved scan arrives as a pull request a person reviews, listing the papers that newly link to two or more seeds.
Why it also searches by keyword
New papers have few citations yet, and the scan keeps only the 100 most-cited papers that cite each seed, so recent work is easy to miss. So a review also runs written keyword searches on OpenAlex and Europe PMC and records the exact queries and hit counts.
In the review of 10 October 2026, an AI agent ran them. Of the 20 records that passed the rules, the keyword searches found 19 and the citation scan 3; one was found by the scan alone. A preprint and its published version count as two records. Running the searches on a schedule is planned.
Step 3 · Screen
It decides what counts from written rules, and keeps every reason
Ranking only sets the reading order. Whether a paper counts is decided against written rules, stored with a version stamp in the screening record. Every screened paper gets one of three verdicts and a one-line reason, and nothing is thrown away: excluded papers stay on file with the reason they were left out.
The rules for this screening pass
- IncludePrimary study (human or animal), systematic review or meta-analysis that measures brain structure, function or molecular markers at least one week after psilocybin dosing, with the timepoints stated.
- ExcludeNarrative reviews, commentaries, editorials, news, protocols, conference abstracts and theses.
- ExcludeRetracted papers.
- ExcludeOther drugs, unless the paper compares them with psilocybin.
- ExcludeAcute-only measurements: every brain measure under one week after dosing.
- ExcludeNo brain measure: clinical, behavioural, psychological or peripheral (blood) outcomes only.
The rules carry a version stamp, c8662963b. Each verdict records the version it was made under. If the rules change, old verdicts stop counting until someone looks at them again.
Three recorded verdicts
- Include
Effects of psilocybin on sleep quality and brain microstructure in chronic cluster headache (2026)
Chronic cluster headache patients: diffusion MRI 1 week after the third of three weekly doses; no significant group change.
- Exclude
Single-dose (10 mg) psilocybin reduces symptoms in adults with obsessive-compulsive disorder: A pharmacological challenge study (2025)
Clinical outcomes only (OCD symptoms); this report has no brain measure.
- Pending
Psilocybin’s lasting action requires pyramidal cell types and 5-HT2A receptors (2025)
Mouse two-photon spine imaging after one dose (Kwan lab). The abstract states no post-dose imaging days; full text unreachable here (PMC, publisher and lab site blocked).
This pass covered 161 of the 521 candidates in the scan of 28 September 2026, those from 2024 to 2026, and 174 records from the keyword searches. The other 360 candidates in that scan have no recorded verdict yet. Its rules ask whether a study measured the brain a week or more after dosing; the synthesis has broader rules of its own, and the trust panel counts only verdicts made under those.
Who decides: an AI model drafted these rules, each verdict and its reason; a person approves the change before it is published. Today that person is the publisher, until a scientific editor is appointed. Which studies to include is an editorial judgment, as the AI policy says.
Steps 4 to 6 · Read, claim, connect
Each finding becomes a claim you can trace to its paper
Reading pulls out the population, design, sample size, timepoints, the change measured and its effect size, with the authors' own words quoted. A claim then states one finding in plain language with its tier and the reason for it, and is linked to other claims by a typed relation: it extends one, measures the same thing differently, differs by population, or suggests a mechanism.
In rats, a single dose of psilocybin or of the 5-HT2A agonist 25CN-NBOH reduced immobility in a forced swim test for at least three months; several months after treatment, prefrontal layer 5 neurons still showed changed resting potential, firing and synaptic excitation after either drug, while no difference in synaptic density or spine types was detected between control and psychedelic-treated rats; the abstract does not say whether that structural result holds for psilocybin on its own.
Claim 19 in full, with why it has this tier- One findingStated in a sentence, with the population and timepoint in it.
- Its papersEvery claim names the sources it rests on, with DOI links.
- A tierEstablished, active debate, preliminary, speculation, or our interpretation, which is always posed as a question.
- Why this tierThe numbers and quotes behind it. The claims the argument rests on also say what result would change them.
Steps 7 and 8 · Check, publish, watch
Agreement from a single team is flagged
Computed by code
- Research programs: two papers belong to one program when they share the first or last author, or two or more authors; one shared author in the middle of both lists is a tie, not a merge. This synthesis has 24 papers from 15 programs.
- A claim whose two or more sources all come from one authorship group is flagged: its sources share authors. That is a prompt to check whether their cohorts and data are independent, not a finding that they are not. Papers that reuse one sample are not independent replications; shared authors alone do not mean shared participants.
- Counts in the trust panel, such as sources, programs, tiers and scan dates, are computed from the data.
Checked by AI, and by people
- AI stress tests: separate AI agents look for errors, confounds and overstatement in a posed reading. They are not peer review, and AI is never credited as an author or reviewer.
- Signed expert review: invited, not yet done for this synthesis. The trust panel says so.
- Each change is versioned, waits for a person's approval, and is logged on the page with its date.
Where this is going
From many papers to new questions, with shared credit
Today PsychDoc links claims inside one question. The plan is to link claims across questions, journals and fields wherever they measure the same thing, so the system can point to a pattern no single paper shows. Each such pattern would be posed as a question, checked, offered to labs to test, and carry a record of everyone whose work it rests on.
| Part | Today | Planned |
|---|---|---|
| Finding papers | A weekly citation scan from the seeds, whose pull request lists newly linked papers; keyword searches run by an AI agent during a review, with queries recorded. | Keyword searches on a schedule, and seeds added automatically when a study joins the synthesis. |
| Deciding relevance | Written rules with a version stamp; an AI model drafts each verdict and its reason; a person approves. | Screening proposals arrive with each weekly scan, for every candidate, ready for a person to approve. |
| Connecting findings | Typed links between claims within one question; our own readings are posed as questions and stress tested by AI agents. | Links across questions, journals and fields through shared measures, populations and mechanisms; the system proposes candidate questions. |
| Independence | Research programs computed from author lists; agreement from one program is flagged. | The same count across fields, so a cross-field question shows how many separate teams it rests on. |
| Credit | Every claim names its papers and authors; the page cites them with DOI links. | A contribution record for each posed question, naming papers, journals, labs, universities, reviewers and the testing lab, with credit terms agreed with partners in advance. Nothing is agreed yet. |
Numbers on this page are computed when the site is built from the synthesis Does one dose of psilocybin rewire the brain? (version 2.2), its scan of 28 September 2026 and its screening of 10 October 2026. See the evidence map and the editorial charter.