Foomax

The Man in the Lab Coat

September 2026 · Erowid

Part 2 of a three-part series from the same corpus.

On a corpus where one report in five is written as a timestamped log — what the apparatus of rigour buys, and what it quietly doesn’t.

I. An apple, and a control group

In September 2010 a man took LSD in a shared house and went downstairs for a snack.

I donned my white labcoat… Just wearing it provokes me to think within a scientific framework. Suddenly, Im not just getting a snack from the kitchen. Now I am performing a perceptual test… I wander around the house, feeding slices of apple to people. Control group: People not on LSD! How would you describe this apple?

The coat is wonderful — that he owns one, that he admits to using a costume as a cognitive prosthetic. But the next line matters more:

My control group told me is was sour, and crisp. Those subjects on LSD were in agreement.

Sober people: sour and crisp. Tripping people: sour and crisp. He ran the comparison, got a null, wrote it down, and moved on. Nobody made him. He was in a house full of drunks at two in the morning holding a plate of apple slices.

And my classifier scored that report negative for every scientific-register theme I built. The most rigorous act in the corpus is invisible to the instrument built to find rigour. I’ll come back to that.

II. More instrument, less sermon

I categorised 25,171 experience reports — 26.8 million words across four decades. The top of the list is what you would guess: nausea, closed-eye geometry, the world breathing, time coming apart.

Third is not a topic at all. It is a format. Roughly one report in five is written as a timestamped log — a document whose spine is a column of time-marks rather than sentences.

The classifier catches these at 93% precision — the cleanest signal I measured, cleaner than nausea, because prose almost never begins a line with T+2:30.

And the form is winning. Standardising for report length, the timestamped log rose about 60% in relative terms between 2000 and 2026, while the advice-to-the-reader coda — be careful, respect the substance, always have a sitter — fell 47%. Those are the only two era trends that survived length-standardisation; four others were nothing but word count. In twenty-five years the reports stopped preaching at you and started handing you a dataset.

III. What the timestamp buys

The property that made the classifier work is the point: the timestamp is legible to a machine. “After a while things got weird” is a story; a column of time-marks is data, and it is data because the author accepted a constraint that cost him something in the moment.

From a 2018 LSD report, “Famous Last Words”:

I set a bunch of alarms on my phone, one every 30 minutes for 10 hours, to remind me to take note of what I was experiencing.

Twenty alarms, on LSD. The entries record not visionary content but calibration:

T: 0hr30m — I feel a little light. I’m probably imagining it. T: 0hr:44m — Not imagining it. Real. I like it.

Fourteen minutes. He registers an effect, discounts it as expectancy, then revises. That is live placebo discrimination on oneself, timestamped — and only the format preserves it. Written as prose the next morning it becomes “it came on slowly.” The clock keeps the epistemics.

IV. And what it doesn’t

I ran a recall audit — thirty reports the classifier had rejected, read in full — and asked a broader question of each: does this narrator deliberately record or measure in any form?

Seventeen of thirty. Roughly four times the rate of the timestamp format itself. Cassette recorders. Digital scales against a remembered reference dose. Volumetric dosing — 265mg in 500ml, 50ml decanted — with a note flagging a contaminant as a confound. The timestamped log is not the phenomenon; it is the visible tip. Treating your own experience as an experiment is close to a majority practice, and the format is merely the part a regex can see.

And yet. Almost nobody has a comparison condition — the lab-coat man is remarkable precisely because he is nearly alone. Nobody is blinded. The dose stays a guess where it matters most: null rates track dose uncertainty, 15.7% for wild-picked Amanita against 5.1% for nitrous. Timestamps do not tell you what was on the blotter.

Most sharply: the instrumentation does not prevent the escalation. Reports containing a null are four times more likely to contain a redose, and that is no lower among the meticulous. The man with twenty alarms is doing something real. He is not thereby protected. The apparatus records the decision; it does not improve it.

V. Cheap rigour and expensive rigour

This is not a story about drugs. It is about what a community gets when it adopts the forms of science without the institutions — and the honest answer has two halves people want to collapse into one.

Rationalist culture is unusually invested in visible apparatus: epistemic-status headers, confidence intervals on casual claims, calibration curves, Brier scores. People wearing lab coats because the coat provokes them to think within a scientific framework. This mostly works. It is not cargo cult: twenty-five years of drift toward the timestamp and away from the sermon is a community teaching itself something true.

The failure mode is that the apparatus is cheap and the substance is expensive, and they feel identical from the inside.

A timestamp costs nothing. A confidence interval costs a keystroke. A control group costs you a friend, an apple, and the willingness to find out that the sober people and the tripping people said the same word. Blinding costs a collaborator and your own certainty. Pre-registration costs the option of deciding afterwards what you were testing.

Every cheap thing produces the feeling of rigour; only the expensive ones produce the thing. And because the cheap ones are what is visible in a document, they are what gets rewarded and imitated.

So: which of your epistemic habits would survive if nobody could see them? The timestamp is legible, and that is most of why it spread. The control group in the kitchen at two in the morning was legible to nobody, produced a null, and was written down anyway.

VI. My own lab coat

I built 356 regular expressions, ran them across 26.8 million words, and produced a beautifully ranked table of the ten most common themes in the largest archive of first-person drug experience there is.

Then I had the matches read by hand, and roughly half the constructs turned out to be measuring something else. bad trip indicated genuine terror one time in eleven — a topic label, appearing mostly inside “I have never had a bad trip.” My detector for lye, a real extraction reagent, fired on people who lye down on the bed, and on the filmmaker Len Lye.

The table had the form of a result: columns, percentages, a ranking. Every statistical check passed. A bad instrument does all of that perfectly well.

What exposed it was 670 matches read by a person, one at a time, asking of each: is this actually what I said it was? That is the expensive thing. It produced no new numbers — it only corrected old ones, downward.

I had spent a day in a very good lab coat.

The apple is the hard part. It always was.

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