The Corpus

Part One

The archive, and what it is not.

The corpus is 5,623 English-language accounts, with a further 445 originally submitted in twenty-six other languages held in reserve for the cross-cultural question below. Every account is a long structured questionnaire (background, the experience itself, its elements, the aftermath) plus free narrative. It is, by a wide margin, the largest systematically-scorable body of near-death testimony available.

It is also self-selected at every stage, and the discipline of using it begins with naming how. People submit because they had an experience worth submitting; they find the site because they were already looking; they write in English, mostly from Anglophone, Christian-majority cultures; and they recount events from years or decades earlier. None of this is fixable by analysis. It means the archive is a sample of memorable, narratable, voluntarily-shared experiences, not a sample of what happens to brains near death. Every result here inherits that ceiling. The most a clean finding can claim is that a predicted structure is present in the testimony, never that it is present in the world at the rate the testimony shows.

Part Two

Before any test: cleaning the data.

Most of the work was not testing hypotheses. It was establishing the right to. Three problems sit between the raw archive and any honest comparison, and each one, left uncorrected, produces spurious findings.

No fixed column set.

There is no single questionnaire. Across the corpus there are 211 distinct question keys, and fill-rate is bimodal: twenty-two questions appear in at least 90% of records, another twenty-three in half to nine-tenths, and a long tail of 136 show up in under a tenth. Only about seventy-six are analysable at all. Every analysis here restricts to questions with adequate fill; the rest are noise, not data.

The questionnaire drifted with time, and the drift mimics discovery.

The survey grew and then shrank: an early form of roughly forty questions, a middle form near seventy, a late form of fifty-six. Because question count is not monotonic in time, era cannot be read off the number of questions. The version has to be recovered. This matters, because a question that was simply added and later dropped will look just like a motif that rose and fell over the decades. One such question, “did you gain special knowledge about your purpose?”, is shown below by submission-order decile:

The rule that follows is strict and is applied throughout: test a motif’s prevalence only within the records whose version actually asked about it, and carry the version as a covariate. Never compare raw prevalence across the full time range.

Inclusion is a decision, not a given.

The archive files self-induced and non-near-death out-of-body experiences under the same label as cardiac-arrest survivors. Pooling them is a real confound, and unpooling them moves the numbers, so the safest course is to report both ways and always say which rule a result uses.

One more check, before scoring: do the ticked boxes agree with the stories? For the out-of-body item, narratives corroborate the box in 31% of ticked-yes cases versus 7% of ticked-no, a 4.4× ratio that confirms checklist and narrative co-vary; the absolute 31% is a conservative lower bound from blunt text-matching. Box and story move together, so the instrument is measuring something real, even if coarsely.

Part Three

Scoring the experience.

The questionnaire reproduces the Greyson NDE Scale, sixteen items, each scored 0, 1, or 2, summing to a maximum of thirty-two, with the conventional near-death threshold at seven. Because the archive preserves Greyson’s response anchors as its answer options, all sixteen items are recoverable with proper graded scoring. 3,996 records (71%) carry the complete sixteen; this “complete-16 spine” is the backbone of every test that follows.

The mean of twenty is above Greyson’s original criterion-group mean of fifteen (just what a self-selected archive should do), and the ordering of the subscales (affective highest and paranormal lowest) matches the established profile. The inclusion stratum interacts strongly with the score: strict near-death cases run a full seven points above out-of-body-only cases, the source heterogeneity made visible.

Does the instrument behave like the published scale? Largely, yes. Internal reliability is high. The one loose subscale is no defect; it is telling us something.

Transcendental is the loosest subscale: out-of-body, otherworldly realm, encountered being, and border are partly alternative routes through the experience rather than one tight dimension, which is what the account would expect of a single state entered by different doors. The deeper structure is also telling. A single dominant axis (plain intensity) accounts for 45% of the variance with every item loading the same way (everything rises together), and it takes five components to reach 70%. Greyson’s four-factor structure does not cleanly replicate here. In a four-factor solution only the affective block separates, while cognitive, paranormal, and transcendental items load onto the general factor. This corpus is closer to “affective core plus general intensity” than to four crisp dimensions.

Part Four

The theory on trial.

These are the tests that bear on the theory. Three of them it could have failed and did not; one was designed, then dropped because it was circular, and the discard is worth reporting too.

The conjunction: unity without dimming.

The series’ core claim about the lucid state is a conjunction: the boundary dissolves and a subject remains. The falsifiable edge of that claim is a prediction of absence: there should be no sizeable population reporting deep unity together with a dimming of consciousness. If oneness were just the lights going down, the corpus should be full of people reporting both at once. It is not. As reported unity climbs, the cognitive subscale climbs with it, and the fraction who report less consciousness than waking falls: dimming becomes rarer where unity is deepest.

The obvious objection is that this shows nothing the theory can claim: a single general intensity factor (in a profound experience, every item ticks up at once) would produce the same co-rise. It is the right objection, and unlike most raised against a study like this, it is testable. Regress out that general factor (the summed weight of the other fifteen items) and a smaller association survives: deeper unity still predicts preserved rather than dimmed consciousness, partial r ≈ 0.08, p < 10⁻⁶, holding on the full spine and with the out-of-body-only cases removed (directionally the same but underpowered in the strict-only subset). The gradient also survives within fixed intensity bands: hold overall intensity constant and the more unified experiences still dim less (in the lowest band, 11.6% reporting dimming at no unity versus 6.7% at full unity). About half the raw co-rise is general intensity; the other half is a unity–lucidity link that intensity does not explain. The sharper test is what happens when the conjunction breaks.

Distress, and the breakup of the lucid state.

The same conjunction makes a sharper, riskier prediction about its failure. If lucid unity is boundary-down-with-subject-held, then when an experience goes wrong (turns frightening or distressing), what should give way is the unity, not the lucidity: the subject is still there, badly, while the sense of oneness collapses. The deflationary account expects something blander, a roughly uniform dimming of a less pleasant state. Sorting accounts by the experiencer’s own verdict (entirely pleasant against frightening or distressing), the corpus supports the theory.

This is the result that answers the rising-tide objection on its own terms, because it is a dissociation. Control for overall intensity and questionnaire version, and distress still drives unity sharply down (β = −0.69) while cognition, far from falling with it, moves the opposite way (β = +0.28). A single general factor cannot push two subscales in opposite directions at once. The only thing that can is a unity axis structurally separable from the rest, which is what the theory predicts will fail when the lucid state breaks. The affective items move most of all, but that is half built in: distress scores low on peace and joy by definition. Unity, consciousness, and cognition are not part of the content verdict, which is why their pattern is the part that could have come out otherwise, and did not.

The floor: void or fullness.

The checklist alone cannot say whether the broken state takes the form of void (the contentless emptiness that would bear on the floor), because keyword-matching for “void” and “emptiness” catches mystical-positive language (“a formless realm of light”) as readily as collapse; on a random sample the regex was right only 29% of the time. So each narrative was read by a language-model classifier, prompted to code its deepest point. The labels agree with the experiencers’ own pleasant-or-distressing verdict 82–86% of the time, and they recover the signal the regex could not: narrative-judged void goes with a steep fall in consciousness-level (about a standard deviation below the rest, steeper than the fall in overall intensity) and, in the higher-powered sample, in unity as well. This is cross-instrument (the void call comes from the prose, the unity and consciousness scores from the checklist), so it is not one rater seeing both. Because these void labels are machine-coded, they should be treated as hypothesis-generating until replicated by blinded human raters or an independent classifier.

Border-as-proximity.

The second test is cleaner. The “border or point of no return” motif should track genuine proximity to death, not merely the intensity of any altered state: the series reads the border as a real feature of the re-parsing, where the deflationary account has no reason to expect it to discriminate. It does. The border item is markedly higher in strict, life-threatening cases than in out-of-body-only ones, and a permutation test puts the gap well past chance.

The discarded test.

One further test, of reunion-by-depth (whether encountering deceased beings scales with how deep the experience runs), was considered and dropped. It is circular: the deceased-beings item feeds the Greyson total, so “depth” and “reunion” are partly the same measurement, and any correlation is built in. One related number is worth keeping in view against any temptation to over-read the social content: encountered deceased beings appear in only about 44% of accounts. The experience turns social less than half the time. When it does, it reliably takes the shape of reunion rather than blank merging, but reunion is not the modal event.

Part Five

Belief, and culture.

If the experience were largely a cultural construction, its content should track the experiencer’s prior beliefs: religious people meeting religious figures, the devout finding confirmation. If instead there is a stable phenomenological core with only an interpretive overlay, the affective spine should be roughly belief-invariant while at most the transcendental wording shifts. The corpus lets one ask, carefully, always within questionnaire version, since version drives which belief and which content items are even present.

The discriminating question is about content. If belief shapes the experience, its transcendental and religious furniture (an encountered being, an unearthly realm, deceased spirits) should rise for prior believers while the affective spine stays put. Within version, comparing religious to nonreligious experiencers, that is faintly what happens: the transcendental items nudge up while the cognitive control stays flat. But the effects are tiny, around 5% of the scale’s range, and they do not hold up.

A blunter signal points the same way: non-believers more often call the experience entirely inconsistent with their prior beliefs (43% versus 23%). But it proves little on its own. Someone who expected oblivion will count any vivid experience as inconsistent, so that gap is over-determined and cannot by itself separate construction from a stable core. The argument depends on the content test, and there the verdict is weak, equivocal evidence for expectation-tracking, which, seen differently, is the dominant pattern: a largely belief-invariant experience carrying at most a small interpretive overlay. That is the invariant-core-plus-overlay picture the series relies on, supported here weakly rather than proven.

Culture is the test still owed.

Belief, within an Anglophone sample, barely varies: almost everyone ranges over flavours of Western Christianity and secularism. Culture is the high-variance predictor, and it is the one the corpus cannot yet properly test. The 445 non-English originals are real and usable, and, unexpectedly, the site serves their questionnaires already normalised to English anchors, so the same scorer applies without modification. But the sample sizes are brutally uneven.

Only Spanish, at 143, approaches the sample needed for a stable factor model. Everything else is descriptive at best. So the decisive cross-cultural test (measurement invariance across languages, then a comparison of the affective core against the culture-varying transcendental overlay, with a bottom-up narrative arm to catch the motifs the Western-designed scale cannot see) is designed and feasible, but underpowered and not yet run. The languages table cannot stand in for a result it does not yet contain.

Part Six

Kinds, or degrees?

Do near-death experiences fall into discrete types, or is variation mostly a matter of degree? The series tends to treat the experience as one state entered to varying depth, and the corpus broadly agrees. Simple clustering finds only severity tiers (mild, medium, and full, at mean totals of 8.8, 18.3, and 29.5), which is just the intensity axis again. Remove that axis, and no discrete clusters remain: what is left is continuous, not categorical. But it is not formless. One reproducible secondary axis runs between two poles.

So beyond “how intense,” the single reproducible contrast is transcendent/affective bliss versus cognitive/paranormal information, not the out-of-body-veridical versus otherworld-encounter split one might have hypothesised in advance, which did not anchor a type. The answer to “do they fall into kinds?” is: predominantly degrees, plus one continuous affective-to-informational axis, with no clean types. A single state, varied mostly in depth and slightly in flavour: the shape the theory predicts.

Part Seven

What the corpus cannot do.

The limits are not footnotes; they are the reason this is a supplement.

• One archive, self-selected. A single website, voluntary submission, recollection often decades old. The patterns may belong to who submits rather than to what happens. No internal analysis lifts this ceiling.
• English-dominant. The cross-cultural arm, the one that could separate a stable core from a cultural construction, is underpowered, with only one non-English group near the threshold for proper testing.
• Text cannot reach the key prediction. The discriminating claim of the series is an interaction coefficient: does baseline interior integration buy deeper dissolution before the subject collapses? That requires neural measurement at the edge of oblivion. No volume of testimony, however large, can estimate it. The corpus can show the shape is consistent, but it cannot run the experiment that would make it decisive.
• And it cannot touch the metaphysics. Whether there is one awareness underneath, or dependent arising all the way down, is a question about the floor, and the floor is the one place from which, by the theory’s account, no report returns. Text is the wrong instrument for a question about what survives the failure of all instruments.

Coda

What it adds up to.

Held against the series, the corpus does something modest but real. It shows that the predicted conjunction holds where it is falsifiable (deep unity is accompanied by a dimming of consciousness less often, not more, as it deepens), and that when an experience fails, what collapses is unity while cognition holds (a dissociation that survives controlling for sheer intensity and so is not the rising-tide artifact a skeptic reaches for first). It shows the border motif tracking genuine proximity to death rather than mere intensity, which the deflationary account does not predict. It shows an experience that is largely invariant to belief, with at most a thin interpretive overlay, and a variation structure that is mostly depth plus one soft affective-to-informational axis: a single state seen at different distances, not a taxonomy of different states. None of that proves the picture. All of it is the kind of thing that should be true if the picture is right, and could have come out otherwise.

What the corpus cannot do is exactly what The Window reserves for an experiment that has not been run and a coefficient text cannot measure, and what The Return concedes about the floor. Keeping these numbers here, off to the side and marked exploratory, is the point: the argument should stand on its reasoning and its one falsifiable prediction, not on a self-scraped archive, and where the archive agrees, that agreement is a bonus, not a foundation.

References & sources

Data and further reading

The corpus is drawn from a public archive. The scoring instrument and the comparison studies below are general background, not endorsements of the theory’s specific claims. The analysis, and any error in it, is the author’s.

• Near-Death Experience Research Foundation (NDERF) archive · data source
• Near-death experience Wikipedia
• Bruce Greyson, the NDE Scale (1983) Wikipedia
• Cronbach’s α (internal reliability) Wikipedia
• Confirmatory factor analysis & measurement invariance Wikipedia
• Bonferroni correction (multiple comparisons) Wikipedia
• Selection bias Wikipedia
• Pim van Lommel, prospective NDE study Wikipedia