A twelve-step pipeline across eleven archives, source identification, transcript collection, multi-pass AI analysis, two-model cross-validation, and automated quality assurance. Applied to first-hand accounts.
Identifying long-form interview archives that feature first-hand awakening and NDE accounts. Each source was evaluated on interview depth (typically 60–180 minutes), back-catalog size, caption availability on YouTube, and whether the content focuses on first-person experiential accounts rather than commentary or teaching. Eleven archives met the bar: Buddha at the Gas Pump, Conscious TV, Shaman Oaks, Simply Always Awake, a curated YouTube Mixed collection, the Kundalini Conversations series from Brent Spirit, a hand-curated selection from the Guru Viking channel, long-form NDE documentaries from Anthony Chene Production, first-hand spiritual emergency accounts from Integrative Mental Health University, and first-hand spiritually transformative experience accounts from Spiritual Awakenings International.
The complete video library for each archive is compiled automatically, then filtered by title to remove known non-qualifying content, panel discussions, book reviews, educational how-tos, and multi-guest roundtables. The same filter rules apply consistently at both the library and import stages, so nothing slips through from one step to the next. Output is a list of qualifying video URLs for each source.
Revised 31 July 2026, and applied from the next round of collection onward. Episodes with more than one guest are no longer removed here. They are kept, read, and separated by speaker, so that each person's account is studied on its own rather than merged into one record or dropped altogether. The figures on this page describe the collection assembled under the earlier rule.
Captions are retrieved from YouTube for each qualifying video, using auto-generated or manually-uploaded transcripts depending on what each video has available. Raw transcripts are saved as structured data, one file per source archive, ready for import.
Each transcript is imported and normalized into a structured record, with a second filtering pass that checks for minimum length and drops duplicates. Every episode record is a container that accumulates tags from each subsequent pipeline pass, all analysis layers are stored together in one place per account.
Each transcript receives a quality rating (high, medium, or low) during the first AI pass. Episodes rated low, insufficient experiential content, transcripts too short for analysis, apparent third-party summaries, or interviews where the guest's own experience is barely touched, are removed before the full analysis pipeline runs. The accounts in the corpus all cleared this threshold; the raw import count was higher.
Claude reads each full transcript and assigns tags across 17 experience categories. The interpretive standard is inclusive: any mention of a phenomenon counts, whether explicitly stated, implied through description, or referenced in passing. The broad pass also identifies trigger types, extracts divine presence descriptions, and captures guest metadata. These are the upper-bound prevalence figures seen throughout the data.
A second AI pass applies a higher evidentiary standard to the same transcripts. Only phenomena with explicit first-person description and sufficient phenomenological detail are tagged, implied references, single-word mentions, and metaphorical language do not qualify. The experiencer must clearly describe the phenomenon as something they directly underwent. These are the lower-bound figures: the most conservative reading of what was explicitly reported.
Two dedicated passes extract additional data layers: somatic phenomena (energy surges, involuntary movement, physical collapse, sleep changes, visual and auditory shifts) and integration challenges (social isolation, identity disorientation, dark night of the soul, difficulty communicating the experience). Each layer is extracted at both broad and strict levels, following the same interpretive standards as the primary passes.
All episodes are independently analyzed by a second AI model, GPT-4o, using an equivalent prompt, with no access to the first model's results. The two sets of tags are then compared for every episode across every category. Disagreements, cases where one model tagged a feature and the other did not, are flagged for adjudication. This step surfaces tags that are genuinely on the margin and forces closer reading of ambiguous passages.
For each disagreement between the two models, a dedicated third pass re-reads the full transcript alongside both tag sets and issues a final determination with per-tag reasoning. This three-pass pipeline, broad analysis → independent cross-check → adjudication, is the quality-assurance backbone of the corpus. It reduces but does not eliminate tagging error; the broad/strict range holds that residual uncertainty explicitly rather than papering over it.
An automated validation script checks the post-tagging corpus for structural integrity, including a check that every strict-tagged feature is also broad-tagged, required field presence, word count thresholds, episode counts against a saved baseline, and vocabulary consistency. Any episode failing validation is flagged for inspection before statistics are published.
A generation script computes corpus-wide prevalence figures across all experience categories, trigger types, and sub-corpus groupings (NDE, by source). Results are published to the public site automatically. All figures reflect the post-filter, post-adjudication corpus of episodes.
The broad read captures any mention of a phenomenon, whether it was stated explicitly, implied through description, or referenced in passing. If an account described "feeling like everything was connected" without using the word "unity," unity was tagged broadly.
Broad figures represent the upper bound: the scope of the corpus. They answer the question, across all accounts, how often does this phenomenon appear in any form? Use broad figures when you want to understand the reach of a given experience type.
Limitation: broad figures may include marginal mentions that fall short of explicit first-person confirmation. They should not be read as precise event counts.
The strict read requires an explicit, first-person description with sufficient phenomenological detail. Vague mentions, metaphors, and implications did not qualify. The experiencer had to clearly describe the phenomenon as something they underwent, not something they believe in generally or heard about from others.
Strict figures represent the evidentiary lower bound: what was explicitly confirmed. They answer the question, how often was this phenomenon clearly, unambiguously described in first-person terms? Use strict figures when precision matters.
How to read the gap: a large gap between broad and strict for a given feature suggests that the phenomenon is commonly referenced but less commonly described in explicit phenomenological detail. A narrow gap suggests consistent, direct description.
"If a subject says 'I felt like everything was one' during a discussion of their general worldview, the broad pass might tag unity-oneness; the strict pass would not, because it lacks the explicit first-person phenomenological description of the experience itself. If the same subject then described 'a moment where the boundary between me and everything else simply vanished,' both passes would tag it."
The most common reaction to AI-powered analysis is reasonable skepticism: AI makes mistakes, AI hallucinates, how do you know the tags are right? These are fair questions and the methodology was designed to address them directly.
The interviews in this collection average roughly 90 minutes each, about 1,100 hours of recorded testimony and 10 million words of transcript. 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 is being coded by hand specifically to test whether the AI judgments can be trusted, because a study that never checks its instrument against a person has no way of knowing.
The honest summary: AI made it possible to read 10 million 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 operates upstream of all prevalence figures presented on this site. 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.