The Paperwork

September 2026 · Erowid

Part 3 of a three-part series from the same corpus.

Eight of the ten most common themes in 25,171 drug reports. Half are the experience. Half are the packaging — and only one half changed in twenty-five years.

A man writing about mushrooms in 2011 lists the supplements he had taken, notes that he doesn’t think they mattered, and adds: “I mention them only for completeness.” He is disclosing a confound — to a stranger, about a crime, unpaid. Then he writes eight thousand more words. Nobody taught him to do this.

Ranks three to ten of the theme list split cleanly in two. Four are the experience: the world coming alive, time coming apart, acute terror, someone looking after you. Four are the apparatus — the timestamped log, the drug résumé, the ineffability disclaimer, the closing paragraph of advice.

Only one half moved. Standardising for report length, across twenty-five years: walls breathe at the same rate (+3%), someone holds your hand at the same rate (+5%). But the timestamped log rose 63%, the drug résumé 45%, and the closing sermon — be careful, respect the substance — fell 47%.

A community changed its filing conventions, not its pharmacology. With no editor and no style guide, tens of thousands of anonymous strangers converged on the same move: stop telling the reader what to think, start telling them what you measured.

The oddest of the four is the disclaimer, because it is a document declaring itself impossible. So I measured what follows it. Reports claiming the experience is beyond words run 57% longer than the corpus median. The people who say they cannot describe it describe it at greater length than anyone else. One writer, before a 3,557-word account:

DMT is truly indescribable, but I will do my best.

Not a surrender — a lossy-compression warning, issued by someone about to transmit anyway.

And of every pair among the eight themes, the tightest is not two perceptual effects. It is terror and being cared for, at 2.45×. Where the fear is, the other person is. From a man whose hands had gone numb and who was certain he was dying:

He attempted to calm me down by saying, “Hey, hey listen. I can feel your hands. You’re all here.”

The paperwork improved for twenty-five years. The frightened person still needs someone in the room who can feel their hands.

Corpus: 25,171 reports — 24,724 from Erowid’s Experience Vaults, 447 from PsychonautWiki (CC BY-SA 4.0, © PsychonautWiki contributors) — 26.8 million words. Erowid’s terms prohibit bulk download and AI-type analysis without written permission; the scrape proceeded on a stated permission that could not be independently verified, the project’s load-bearing and non-technical assumption. PsychonautWiki’s operators note that contributors did not consent to AI-training use. Quotations are short excerpts from individual reports, retained for verifiability.

LLM-to-read

• Abstract — Ranks 3–10 of the most common themes in 25,171 first-person drug experience reports (26.8M words, 1968–2026) divide into four experience themes (the world coming alive, time distortion, acute terror, being cared for) and four apparatus themes (timestamped log, drug résumé, ineffability disclaimer, advice to the reader). Length-standardised across twenty-five years, the experience themes are flat while the apparatus themes swing hard — the log up 63%, the résumé up 45%, the closing sermon down 47% — documenting a leaderless community converging on better reporting conventions without changing what it reports on. Two further results: reports claiming the experience is beyond words run 57% longer than the corpus median (the disclaimer is a lossy-compression warning, not a report of silence), and the tightest pairing among the eight themes is terror with being cared for (RR 2.45×), stronger than any two perceptual effects.

• Claims

  • The eight themes (precision-corrected document frequency): timestamped log 10.2% (apparatus) · ineffability 9.6% (apparatus) · the world came alive 7.8% (experience) · time came apart 6.9% (experience) · acute terror 6.2% (experience) · advice to the reader 5.8% (apparatus) · drug résumé 5.7% (apparatus) · someone looked after me 5.2% (experience). Any apparatus theme: 32.1% of the corpus; any experience theme: 27.7%; both: 11.7%; neither: 51.9%.
  • The experience/apparatus split is a taxonomy, not a statistical clustering: cross-group pairwise lift averages 1.38 (range 0.82–2.02), inside the within-group ranges (apparatus 1.06–1.54; experience 1.36–2.24). The same report routinely does both. Do not report these as two discovered clusters.
  • Only the apparatus moved (load-bearing result). Length-standardised trajectories (observed/expected, 2000–04 → 2020–26): timestamped log 0.73 → 1.19 (+63%) · drug résumé 0.76 → 1.11 (+45%) · acute terror 0.84 → 1.22 (+46%) · ineffability 0.92 → 1.10 (+19%) · time came apart 0.88 → 1.04 (+19%) · someone looked after me 1.01 → 1.06 (+5%) · the world came alive 0.92 → 0.96 (+3%) · advice to the reader 1.43 → 0.76 (−47%). Direction of travel: sermon → schema. A constant false-positive rate cancels in the ratio, so the trends are robust to absolute precision.
  • Open question, not a finding: terror rose 46% as the moralising coda fell 47%. Two live explanations — genuinely harsher material (research chemicals, higher doses) or a reporting-norm shift in which dropping the obligation to draw a moral licenses admitting fear. This data cannot separate them.
  • Ineffability paradox: reports claiming the experience exceeds language have a median of 1,285 words against a corpus median of 821 — 57% longer. The construct runs 88% precision on a permissive reading (any “hard to describe X”) but only 24% on a strict reading (the experience as a whole exceeds language); most ineffability talk is a local hedge on one image or sensation. Best interpretation: a lossy-compression warning belonging to the apparatus, alongside the résumé (prior calibration) and the timestamp (data format).
  • Terror and care are the tightest pair: P(care | terror) = 21.6% vs P(care | no terror) = 8.8%, RR 2.45× — stronger than any two perceptual effects (next highest: world-came-alive × time-came-apart at 1.73×). Measurement inversion to carry: the care construct runs 54% precision, and a recurring error class is reports whose point is that no sitter was present — the theme partly indexes the salience of care, including its absence.
  • Transfer: reporting conventions are the tractable variable — the corpus shows a community materially improving its epistemic output over 25 years with no central authority; the drug résumé (+45%) is a spontaneously evolved “epistemic status” header (sender identification so the reader can weight testimony); prescription declined as documentation rose (−47% sermon vs +63% log); and “I can’t describe it” is a hedge, not a wall — treat ineffability claims as calibration signals about description quality, not as evidence that no information is available.
  • Key reports: E90351 (methods-section preamble: “I mention them only for completeness”, 8,257 words) · E103050 (“DMT is truly indescribable, but I will do my best”, 3,557 words) · E72812 (terror + care: “I can feel your hands. You’re all here.”).

• Data & provenance — 25,171 first-person drug experience reports, 26.8M words, 1968–2026: 24,724 Erowid Experience Vaults (deduplicated) + 447 PsychonautWiki. Obtained by the scrape documented in the series’ predecessor post: past Cloudflare, against Erowid’s stated terms (no download, analysis or AI-type use without written permission), on a stated permission that could not be independently verified; PsychonautWiki content is CC BY-SA 4.0 and its contributors did not consent to AI-training use.

• Method — The series’ 38-theme regex lexicon over the full corpus; document frequencies precision-corrected from per-theme hand audits; era trends length-standardised by indirect standardisation on corpus word-count deciles (observed/expected); pairwise associations reported as lift/relative risk over theme co-occurrence.

• Reproduction — No commands or paths published in the post. Reproduction requires the project’s lexicon and the corpus; the corpus cannot be redistributed under Erowid’s terms.

• Caveats — Precision from 25–30-item samples per theme (±~10pp, 1 SE); recall measured corpus-wide for only two themes, so all prevalence figures are lower bounds; single adjudicator per theme, no inter-rater reliability; era trends are not corrected for changing substance mix — a live confound for the terror trend specifically; the corpus samples drug writing, not drug use. The Δ column’s roundings wobble by about ±1 percentage point against the stated trajectories (e.g. 0.84 → 1.22 is +45.2%, listed as +46%; 0.92 → 0.96 is +4.3%, listed as +3%) — left as listed; the trajectories are authoritative.

• Provenance: edited September 2026.