Project Profound: Methodology Summary

Project Profound is a computational research platform designed to extract, classify, and analyze first-person accounts of near-death experiences (NDEs) and unidentified anomalous phenomena (UAP) contact experiences from publicly available YouTube video testimony. Unlike traditional survey-based NDE research, this project operates on spontaneous, naturalistic first-person narratives—video testimony shared by experiencers in their own words, in their own time, without researcher intervention.

The system employs a multi-pass pipeline of large language model (LLM) analysis to apply validated psychometric scales, extract phenomenological features, and generate structured datasets at a scale previously infeasible in experiential research.

This document describes the data acquisition pipeline, the analytical instruments employed, the computational methodology, the known limitations and error characteristics, and the research team's intentions for future validation and scaling.

The corpus analyzed by Project Profound is fundamentally different from the datasets used in traditional NDE or anomalous experience research. Rather than structured questionnaires administered under controlled conditions (e.g., Greyson, 1983; Ring, 1980; van Lommel et al., 2001), the source material consists of first-person video testimony—voluntarily shared accounts recorded in naturalistic settings such as interviews, podcasts, and personal testimonials published to YouTube.

This distinction carries significant methodological implications:

• No experimenter demand effects.The experiencer is not responding to a researcher's questions or a pre-structured survey instrument. The narrative unfolds according to the experiencer's own priorities and emotional sequencing.
• Rich phenomenological detail. Video accounts frequently contain paralinguistic information (hesitation, emotional inflection, self-correction) that is absent from written questionnaire responses. While the current pipeline operates on text transcripts and does not analyze audio or video features, this richness is preserved in the raw data for future multimodal analysis.
• Self-selection bias. The corpus is composed of individuals who chose to share their experiences publicly. This likely overrepresents dramatic, positive, or culturally validated experiences and underrepresents distressing, fragmentary, or stigmatized accounts. This bias is acknowledged and documented throughout the analysis.
• Uncontrolled provenance. Unlike clinical samples drawn from cardiac arrest units or ICU populations, the trigger conditions, medical histories, and temporal distances from the experience are self-reported within the narrative and cannot be independently verified.

A foundational insight of Project Profound is that near-death experiences and UAP contact experiences, despite their different phenomenological surfaces, share a common substrate: first-person reports of anomalous states of consciousness. Both involve narrative accounts of perceived reality shifts, entity encounters, information acquisition through non-ordinary means, and subsequent psychological transformation.

By building parallel analytical frameworks for NDE and UAP contact testimony, the project enables cross-domain phenomenological comparison at a scale that has never been attempted. The Greyson Scale has its UAP counterpart in the Contact Depth Scale; the NDE Transformation Index mirrors the UAP Contact Transformation Index; and the cvNDE (veridical perception) scale finds its analog in the UAP Evidence Strength Scale. This symmetry is deliberate and is discussed in §7.

The NDE corpus is sourced from YouTube channels that primarily or frequently feature first-person NDE testimony. Channels are identified through manual curation and are stored in a persistent channel registry with enriched metadata including subscriber counts, total video counts, and scanning status. As of this writing, the NDE archive contains analysis of 4,897 videos, and the UAP archive contains 4,151 videos across multiple channels.

A channel-level scanner periodically audits enabled channels for new uploads, discovers candidate videos, and queues them for intake processing. This scanner architecture ensures the corpus grows continuously without manual intervention.

For each video, the system attempts to retrieve English-language captions via the YouTube subtitle API. Both manual (human-authored) and auto-generated (ASR-produced) captions are accepted, with the source type recorded. When captions are unavailable, the video is marked as no_captions and excluded from analysis. Raw caption segments are processed through a transcript processor that concatenates timed segments, applies punctuation restoration, produces a cleaned version for embedding, and chunks the transcript into overlapping segments for semantic search and RAG applications.

The intake pipeline is an automated, multi-stage orchestrator that processes a single YouTube video from URL to fully analyzed database record. The pipeline is designed as a pure function that can be invoked from an admin interface, a scheduled cron job, or a command-line script.

Before running the computationally expensive full analysis suite, a lightweight classification pass screens each video transcript to determine whether it contains a genuine first-person account of a profound experience. This gate uses OpenAI GPT-4o with a focused prompt, examining only the first ~15,000 characters of the transcript at a temperature of 0.1 for maximum consistency.

The NDE classifier outputs:

• Experience type: NDE, OBE, SDE, ADC, STE, or none
• Confidence score: 0–100
• NDE classification: clear_nde (confidence ≥ 70), possible_nde (40–69), not_nde, or insufficient_info(confidence < 20)
• Experiencer name: Extracted via prompt rules that distinguish the experiencer from the interviewer, host, or narrator
• Justification: A 1–2 sentence explanation of the classification decision

Videos classified as "not profound" are persisted in the database with their classification metadata but are not subjected to further analysis, conserving API resources.

Total pipeline execution time for a single video is typically 60–120 seconds, dominated by the parallel LLM analysis calls and sequential embedding insertions.

The UAP pipeline follows a structurally similar architecture but with domain-specific differences:

• Classification uses a UAP-specific classifier with chain-of-thought reasoning and few-shot examples to determine tier, track, content type, source type, and experiencer names.
• Tier 3 Gatereplaces the NDE "not profound" gate — out-of-scope content is rejected.
• Encounter Segmentation— For multi-experiencer videos, an LLM pass segments the transcript into per-encounter blocks, enabling independent analysis of each experiencer's account within a single video.
• Dual Analysis Suite — Program intelligence analysis runs on all Tier 1+2 videos. Encounter-level phenomenological analysis and CET triad scoring run per encounter segment.
• Name Deduplication — ASR-induced misspellings of experiencer names are normalized using fuzzy matching and LLM-assisted deduplication.
• Tier Reconciliation — If encounter segmentation reveals first-person testimony that the classifier missed, the video is automatically promoted from Tier 2 to Tier 1.

Each video that passes the classification gate is subjected to seven parallel analysis passes, each implemented as an independent LLM call with a domain-specific system prompt and structured JSON output schema. All passes use GPT-4o-mini with response_format: {type: "json_object"} at low temperature (0.2) for scoring consistency.

Reference:Greyson, B. (1983). "The Near-Death Experience Scale: Construction, Reliability, and Validity." Journal of Nervous and Mental Disease, 171(6), 369–375.

The validated 16-item Greyson NDE Scale is the most widely used instrument in NDE research. Our implementation scores each of the 16 items (across 4 subscales: Cognitive, Affective, Paranormal, Transcendental) as 0 (not present), 1 (mildly/ambiguously present), or 2 (definitely present), yielding a total score of 0–32.

Classification thresholds: Not NDE (0–6), Mild NDE (7–12), Moderate NDE (13–20), Deep NDE (21–32).

A custom 7-criterion scale (each scored 1–4, total 7–28) evaluating the evidential strength of veridical perception claims. Criteria include: medical state severity, perceptual access impossibility, specificity and precision, unpredictability, self-reported verification quality, verified perception weight, and temporal precedence of perception report.

Scoring levels: Low (7–12), Moderate (13–17), High (18–22), Exceptional (23–28).

A 10-domain scale (each scored 0–5, total 0–50) measuring self-reported transformation: Appreciation for Life, Self-Perception & Identity, Compassion & Concern for Others, Values & Priorities, Spiritual Awareness, Religious Orientation, Attitude Toward Death, Psychic & Expanded Perception, Relationships & Social Dynamics, and Purpose, Meaning & Life Direction.

Each domain also captures a direction indicator (up, down, mixed, shifted, new), evidence summary, and key quote. Aggregate metrics include overall score, transformation breadth (0–10), and depth (1.0–5.0).

Extracts the presence/absence of 15 standard NDE phenomenological elements: out-of-body, tunnel, bright light, deceased relatives, life review, being of light, border/boundary, feelings of peace, cosmic unity, time distortion, enhanced senses, telepathy, otherworldly realm, knowledge download, and choice to return. Each element is scored with a confidence rating (0–100) and supporting transcript quote.

Also outputs: experience type, trigger category, overall tone, intensity rating (1–10), and content safety flags.

Provides fine-grained phenomenological quality assessment (reality comparison, vividness rating, 6 sensory modalities, emotional progression, altered cognition) and detailed entity encounter documentation (identity, type, appearance, communication method, message content, emotional quality).

Reconstructs the chronological sequence of phenomenological events using a 25-element taxonomy organized across 6 phases: Initial Transition (4 elements), Emotional/Sensory States (7), Encounters (5), Realm/Environment (4), Transformative Experiences (5), and Return (5). An element synonym normalization layer handles LLM output variations.

Generates a concise, objective, 80–150 word summary at a Grade 8 reading level, structured as Trigger → Experience → Aftermath. Used for search result cards and accessibility purposes.

The UAP analysis employs a parallel triad framework designed for cross-domain comparison:

The UAP-ESS evaluates evidential quality across 7 criteria (score 7–28). The UAP-CDS measures contact depth across 8 dimensions (score 0–32). The UAP-CTI extends the NDE-TI with 12 domains (score 0–60), adding UAP-specific dimensions like worldview expansion, relationship to secrecy/disclosure, and ecological consciousness.

UAP channels also receive computed aggregate scores including Intelligence Value, Speaker Credibility, Encounter Depth, Impact Score, Archetype Classification, and a 3-letter Channel Personality Code.

• Application: Next.js 14+ (App Router), deployed to Vercel
• Database: Supabase (PostgreSQL) with pgvector extension
• LLM Provider: OpenAI API (GPT-4o-mini for analysis; GPT-4o for UAP classification)
• Embedding Model: OpenAI text-embedding-3-small (1536-dimension vectors)
• Transcript Source: YouTube subtitle API via third-party caption service

All analysis passes execute in parallel using Promise.allSettled(), meaning that the failure of any single pass does not prevent the others from completing. The pipeline records which passes succeeded and which failed, enabling selective re-analysis.

Scoring passes use a temperature of 0.1 for consistency. Element detection and phenomenological analysis use 0.2 for nuanced interpretation. Summary generation uses 0.3 for natural writing.

Transcripts are truncated to manage costs: classification gate uses the first 15K characters, analysis passes use 50K characters (sufficient for most hour-long videos), summary generation uses 30K characters, and full-text embedding uses 8K characters.

The parallel triad design enables direct research questions: Do NDE and UAP contact experiencers report similar phenomenological features? Do these experiences produce similar transformations? What is the evidential quality of claims in each domain?

A 27-dimension numerical vector encodes each NDE: 15 core element presence/absence flags (binary), intensity rating (normalized 0–1), emotional tone (3-dim one-hot), experience type (5-dim one-hot), and trigger category (3-dim one-hot). These fingerprints enable cosine similarity search for phenomenologically similar experiences via pgvector.

Each video generates multiple embedding layers:

The platform includes a conversational AI interface using RAG to ground responses in actual experiencer testimony, maintaining fidelity to source material.

Each analysis prompt underwent extensive iterative development:

1. Initial design based on academic literature and validated instruments.
2. Manual testing against a diverse sample of transcripts spanning high-confidence NDEs, ambiguous cases, non-NDE content, and edge cases.
3. Model comparison across multiple LLM providers and sizes. GPT-4o-mini was selected for its balance of accuracy, cost, and structured output reliability.
4. Error analysis — manual review to identify systematic biases (e.g., score inflation, interviewer/experiencer confusion) and prompt revisions to mitigate them.
5. Schema refinement — iterative adjustment to capture the right level of granularity while minimizing hallucination.

• Explicit scoring rubrics with concrete examples at each score level
• Negative examples — what does not qualify
• Grounding instructions — "Score ONLY what is described or clearly implied"
• Calibration notes — domain-specific expectations to prevent inflation
• Attribution requirements — evidence summaries and quotes for auditability

LLM outputs undergo post-processing: element synonym normalization (e.g., "darkness" → "void_darkness"), score bounds validation, null handling, and deterministic timestamp matching for UAP encounter analysis.

Through manual review, the following error patterns have been observed:

• Slight score inflation — LLMs assign slightly higher scores than human raters, particularly for emotionally evocative narratives. This bias is consistent and approximately uniform.
• Experiencer/interviewer confusion — occasionally attributes interviewer statements to the experiencer. Substantially reduced but not eliminated by prompt rules.
• Hallucinated quotes (< 5% of outputs) — LLM produces paraphrases rather than direct quotes.
• Cultural and linguistic bias — English-language pipeline may differentially recognize cross-cultural phenomenological features.

The LLM analysis produces approximately 5–10% higher error rates than human inter-rater disagreement on the Greyson Scale (where human reliability is typically r = 0.9+). However, the ability to process thousands of accounts in hours rather than years represents a fundamental capability shift for the field.

• Selection bias: overrepresents English-speaking experiencers comfortable appearing on video.
• Platform bias: YouTube's algorithm may preferentially surface certain content types.
• Temporal bias: videos may be deleted or altered after analysis.
• Channel curation bias: initial channel selection was performed manually.

All analyzed content is publicly available on YouTube. The analysis system is designed with explicit respect for experiencer accounts—prompts include instructions to "Be faithful to the experiencer's own words and framing. Do not pathologize, judge, or reinterpret their experience." Content safety flags identify sensitive content for appropriate warnings.

This methodology document is published openly. The analysis instruments, prompt structures, scoring rubrics, and known limitations are fully documented. We do not claim that LLM-mediated analysis is equivalent to human expert assessment—rather, we present it as a complementary approach that enables scale at the cost of some precision.

Now that the analytical pipeline is built and validated, the team intends to: (1) re-analyze the full corpus with more capable models — the current pipeline uses GPT-4o-mini for cost efficiency, and its text and language skills were deemed viable for these tasks; however, re-analysis with GPT-4o, Claude 3.5 Sonnet, or Gemini 3.1 Pro (or cutting-edge frontier models) would likely improve overall accuracy; (2) implement ensemble scoring to reduce model-specific biases.

Continue growing the corpus through channel scanning, and extend the pipeline to additional experiential domains including: psychedelic experiences (DMT, psilocybin, ayahuasca), mystical/contemplative experiences, reincarnation and past-life memory accounts, and high-strangeness contact reports.

Planned studies include: human inter-rater reliability comparisons, test-retest reliability measurement, and convergent validity against traditional self-administered questionnaires.

The team seeks established researchers in consciousness studies and related fields who can improve analytical instruments, contribute to validation studies, utilize the structured dataset for hypothesis testing, and advise on methodological best practices.

• Greyson, B. (1983). The near-death experience scale: Construction, reliability, and validity. Journal of Nervous and Mental Disease, 171(6), 369–375.
• Moody, R. A. (1975). Life After Life. Mockingbird Books.
• Ring, K. (1980). Life at Death: A Scientific Investigation of the Near-Death Experience. Coward, McCann & Geoghegan.
• Ring, K. (1984). Heading Toward Omega: In Search of the Meaning of the Near-Death Experience. William Morrow.
• van Lommel, P., van Wees, R., Meyers, A., & Groeneveld, I. (2001). Near-death experience in survivors of cardiac arrest: A prospective study in the Netherlands. The Lancet, 358(9298), 2039–2045.
• Hynek, J. A. (1972). The UFO Experience: A Scientific Inquiry. Henry Regnery Company.