A twelve-step pipeline across archives, source identification, transcript collection, multi-pass AI analysis, two-model cross-validation, and automated quality assurance. Applied to first-hand accounts.
Long-form interview archives were assessed on how deep the interviews go (usually 60 to 180 minutes), how large the back catalogue is, whether captions exist, and whether the content is people describing their own experience rather than teaching about it. archives are in the current collection. Two former archives are no longer in it: every one of their accounts turned out to be either a near-death account, which is now studied separately, or off-topic.
Each archive's full video list is compiled automatically, then filtered by title to drop things that are not first-hand accounts: book reviews, how-to content, and narrated summaries of other people's experiences. Captions are retrieved for everything that remains.
Every transcript is screened against a written rule: does the person narrate an experience of their own, in enough detail to be read? These are interviews, so hosts digress and guests teach. That does not disqualify an account. What disqualifies it is the person never describing anything that happened to them. Panels and multi-guest interviews are kept and separated by speaker rather than dropped.
A recording is not a person. interviews turn out to be people: 53 of them were interviewed more than once, and one appears five times. Every figure on this site is a share of people. The link between recordings and people is worked out from names and checked by hand where it is uncertain, and a near-match is never merged automatically.
In August 2026 every account in which somebody nearly died was moved to a study of its own. Nobody behind the figures on this site had a brush with death. That was done because near-death accounts are a different phenomenon that deserves its own treatment, and it has a useful side effect: where a feature usually associated with dying still shows up here, it cannot be put down to dying.
Before any reading began, every category was written down in full: what it means, what counts, what does not, and which neighbouring category it is most often confused with. That rulebook is the single source of every definition. No reading pass writes its own. Each reader receives it byte for byte identical, which is the one thing that makes their answers comparable at all.
Each account is read five times over, once for each question: what the experience was like, what happened in the body, what came immediately before, what was hard afterwards, and what helped. Each reading is a separate request that sees the full transcript and none of the other answers. That matters most for the two that are compared to each other: whatever is found in the body must not be influenced by knowing what came afterwards.
No transcript is shortened. The longest is around 46,000 words and it is read in full. An earlier version of this work gave different readers different amounts of the same interview, which meant only two thirds of accounts were being read on equal terms. That is fixed and checked automatically before every run.
Every account is then read again by a different AI model, working from the identical rulebook and seeing none of the first reader's answers. This is what the two numbers on every chart come from: where both readers marked a feature, and where at least one did.
Agreement is measured for every category with a statistic that corrects for agreement you would get by chance, reported with an interval calculated by resampling people. of categories cleared the conventional bar. Which ones did is the most interesting result here, and it is on the insights page.
Agreement is not accuracy: two readers can agree and both be wrong. So 400 applied tags were drawn at random across the whole collection, and each was checked by hand against the sentence it was drawn from. 2.2% did not hold up. That is the accuracy figure, and it only counts tags that should not be there. It cannot see a feature a reader missed, which is the harder problem, because a missing tag leaves no sentence to check.
The rulebook was not written once and left alone. It went through thirteen numbered revisions before the reading that produced these figures, each one recorded with the reason for it, and every account in the collection was read under the same final version rather than under whichever version was current at the time. That is why the whole collection was re-read from scratch rather than patched: readings made under two versions of a definition are not comparable, and combining them would hide the difference inside an average.
The findings are written up as a research paper, which is a plain description of what these accounts report and how much of it existing questionnaires ask about. Every number in it is generated from the data by a script rather than typed, and a checker refuses to build the paper if a number in the text does not come from the data, if a number in a table does not belong to the category named on its row, or if a table or figure is mis-numbered. The paper states what it cannot claim as plainly as what it can.
Every feature on this site is shown as a range rather than a single figure. The range is not a margin of error in the usual sense. It is the distance between two independent readers.
Both readers marked it. They worked separately, from the same written rulebook, each seeing the full transcript and neither seeing the other's answers or anything else already recorded about the person.
This is the careful figure. It answers: how often did two readers, under the same rules, agree this was there?
At least one reader marked it. This holds everyone in the first figure plus everyone the two readers split on.
It answers: among all people, how far might this reach?
The width of the gap is the useful part. A short gap means the two readers saw the same thing, so the figure can be leaned on. A long gap means they drew the line in different places, and the number should be held loosely. That is why nothing is hidden for having a wide band: the band is doing the warning.
This replaced an earlier pair of numbers, and they are not the same thing. An earlier version of this site showed a strict and a broad figure. Those were one reader given two different rulebooks, so the gap between them measured the rulebooks. These are two readers given one rulebook, so the gap measures genuine disagreement about what the transcript says. The change was made in August 2026 and every figure on the site was recomputed.
The obvious objection to reading transcripts with AI is that AI makes mistakes. That is true, and it is the reason for almost every design decision described above. Here is what was measured, and what was not.
The interviews in this collection average about minutes each, roughly hours of recorded testimony and words of transcript. (Hours are worked out from the word count at ordinary conversational speed, so treat them as an approximation; the word count itself is exact.) Here is what a traditional human-researcher approach would have required, what the AI work actually cost, and, because it matters, how much human time this project still took.
These figures were revised upward in July 2026. The earlier version understated both columns: it costed a manual study that skipped the eligibility and reliability work a rigorous one requires, and it left this project's own human hours off the page entirely.
AI did not remove the researcher from this work, and a page that implied otherwise would be misleading. Every account the two AI systems disagreed about, and every one either was unsure of, was read and decided by a person, without being shown what the AI had concluded, so that those decisions could be compared against it rather than merely confirm it. Defining what each category means, and rewriting those definitions when they turned out to be too loose, was human work throughout. A further sample of 75 to 100 accounts was to be read cold by a human coder and compared against the model. That was cancelled in August 2026, deliberately and with the cost understood: the reading was more work than the project could carry, and the absence of a pure human gold standard is stated here rather than left for a reader to discover. What it would have measured is how often the model MISSES something it was asked to find. Nothing in this study measures that, which is why every figure on this site is a confirmed minimum rather than an estimate.
The honest summary: AI made it possible to read words consistently, which no small team could do. It did not decide what any of it means. Roughly 150 to 250 hours of human judgment sit behind these figures, and that number will grow before the research is published.
Before the full pipeline runs on any episode, each transcript receives a quality rating: high, medium, or low. This rating is assigned by the broad tagging pass as part of its first contact with the transcript.
Low-quality episodes, those with insufficient experiential content, transcripts that are too short to yield meaningful data, accounts that appear to be third-party summaries rather than direct first-person testimony, or interviews where the guest's own experience is barely touched, were removed from the corpus before analysis proceeded further.
The episodes in the final corpus all cleared the quality threshold. Some were borderline, and those were generally retained and allowed to receive a low quality rating rather than deleted, allowing downstream inspection rather than silent removal.
Quality filtering happens before any figure on this site is computed. The accounts are post-filter; the raw import count was higher.
The corpus is a meaningful analytical tool. It is not a comprehensive or representative sample of all human awakening experiences. The following limitations should be kept in mind when interpreting the data.