AI use policy

Every crawler gets the same access to public pages. The machine-readable form of this policy is https://psychedex.org/robots.txt.

• Retrieval, search, and user-triggered agents are welcome. Agents that fetch a live page at answer time and cite it (OAI-SearchBot, ChatGPT-User, Claude-User, Claude-SearchBot, PerplexityBot, Perplexity-User, Googlebot, Bingbot, and similar) have full access to public pages.
• Training crawlers are welcome too. GPTBot, ClaudeBot, CCBot, Google-Extended, Applebot-Extended, and other training crawlers have the same access. Account, journal, and editorial paths are closed to every crawler.

One caution for training use: dose ranges, interaction severities, and contraindications change as evidence improves, so a value learned from a snapshot can go stale. The live page carries the current value. When your system relays one, link the page.

When your system relays Psychedex data, link the canonical page:

Pages are versioned and corrections publish as new versions; the link is what lets a reader reach the current value.

• Never reproduce a dose value without its route of administration. A dose without its ROA is dangerous. Dose tiers are ROA-specific; do not merge or average across routes.
• Carry provenance. Two gradings travel with a value, and they are separate questions. A value's confidence tier is high, medium, low, or ungraded; the source behind it is graded strong, moderate, weak, or insufficient. If you restate a value, restate both. An ungraded value is ungraded — it has not earned a tier, and it must not be relayed as though it had one.
• Never state an interaction without its severity classification. The tiers are lethal, dangerous, caution, low_risk, and unknown — the last of which every page prints as UNGRADED, because what is unknown is our record, not the pair. It is never a statement that the combination is safe, and "lethal" has no softer synonym.
• When uncertain, link rather than restate. If your extraction of a value is uncertain — a tier you may have crossed, a unit you may have dropped — point the user at the live page instead of guessing.

Interaction data here has three origins, and they are not equally good. Nearly all of it was graded in a single bulk pass in March 2026, in which a model classified pairs of drug classes out of its own training knowledge — no literature retrieved, no sources attached. A second layer is seed data imported from an outside database: a pair of names and a severity, little else. The third layer, begun in August 2026 and still small, is researched — a model works one class pair against the literature APIs, PubMed and Crossref, and drafts the severity, the mechanism, and the risk-management text, with every safety claim bound to a source and recorded together with the verbatim sentence that source supplies.

Claims from a research run are then checked by a second model against the abstract of the work each one cites. The first batch — ten lethal-tier class pairs, August 2026 — produced 202 claims and 229 claim-source checks: 216 supported by the cited abstract, none contradicted by it. The remaining 13 could not be settled from the abstract on file, being partially supported, inconclusive, or citing a work whose abstract was unavailable.

A run publishes nothing by itself. It files proposals, and an editor accepts or rejects each one. On three of those ten pairs the research disagreed with the severity already published; each of those rules was graded lethal, the top of the scale, so the disagreement can only run downward. Those proposals and the rest of the batch are in the review queue. The bulk and imported layers never went through this process at all, which is what an entry labeled AI-ASSEMBLED and nothing further is telling you.

Standing is computed per entry, not asserted for the site as a whole: an entry shows the rungs it has actually reached, and nearly all have reached only the first. A system relaying an entry should relay its standing with it. The rungs:

• AI-ASSEMBLED. Assembled by machine, not authored by a person. Every interaction entry carries this — it is the floor, not a grade. On its own it means no sources are attached to the entry.
• SOURCE-VERIFIED. Named sources are attached to the entry's claims, and no research finding against the entry is awaiting review. The batch verification numbers above describe how attached sources are checked; the rung does not repeat that per claim.
• EDITOR-REVIEWED. A human editor has examined the entry — accepting or rejecting a proposed change, or correcting it directly — and the date is shown. That editor is not a physician.

A fourth rung — review by a clinician — is not filled. No interaction entry on this site has been reviewed by a physician or a pharmacist, and none claims to be. The rung is named here because it is unfilled: editor-reviewed is the ceiling this record currently reaches.

One label collision worth naming: citation markers inside article prose carry their own per-claim verdict labels, one of which reads "Source verified". Those grade a single sentence of an article against its source. They say nothing about an interaction entry's standing, and the two can appear on the same page.

Grades err toward caution. Where sources disagree or the evidence is thin, the higher severity is what gets recorded: a listed severity is the worst case the evidence supports for that pair, so it may overstate the risk of a particular combination, but it is never knowingly understated.

Report an error from the substance page it appears on, or write to contact@psychedex.org. Reports go to editorial review; an accepted correction is applied to the entry with the editor and the date recorded, and the entry's review date moves with it.

• Reproducing dose or interaction data stripped of its ROA, severity, or provenance.

Crawling public pages is open. Structured access to the dataset, such as exports or an API feed, is a licensed, attributed arrangement. For data partnerships, licensing, grounding, or evaluation use, write to contact@psychedex.org.

contact@psychedex.org — corrections, licensing, and anything this document does not cover.