Methodology

For episodes added by the current pipeline (October 2026 onward), the summary, key takeaways, Fringe Score, topics and glossary terms come from one request to OpenAI's gpt-4o-mini model, sent through the OpenAI Batch API with a fixed instruction prompt that asks for a structured (JSON) answer. Earlier episodes were processed by older scripts that used other OpenAI models and inputs that were not recorded, and some of their topics were assigned afterwards from the title and summary alone.

What the model reads. For podcast episodes it reads the episode description published in the show's RSS feed (the show notes), up to the first 15,000 characters. It does not listen to the audio and does not read a transcript. For YouTube videos it reads the video's captions when they are available (first 15,000 characters), and otherwise the video description (first 5,000 characters). A short summary can therefore only be as informative as the notes the publisher wrote.

What it is asked to produce. A three-sentence summary, and four to six key takeaways described in the prompt as “specific claims, names, book titles, or historical connections mentioned”. Takeaways list what the notes say is discussed. They are not checked against the episode and are not endorsed.

Episodes that were imported before the current pipeline existed may have been produced with earlier versions of these instructions.

The Fringe Score (shown as FS) is a whole number from 1 to 10 that the same model assigns to each episode in the same request. The prompt defines the scale like this:

• FS 1-2Mainstream science and history
• FS 3-4Alternative but grounded
• FS 5-6Conspiracy adjacent
• FS 7-8Deep conspiracy
• FS 9-10Interdimensional or paranormal extremes

What it means. It estimates how unconventional the subject matter of the description is. It is a filter for finding grounded or speculative material first. It is not a measure of whether any claim is true or false, a rating of a host's credibility, or a safety or reliability score. Government documents are not scored.

How reliable it is. Scores are machine-generated and have not been checked against independent human raters. The model gives one answer per episode and is not asked to explain it, so the same description could receive a different score on another run or with a different model. Treat a difference of a point or two as noise. If the model returns no score, the system stores 5, so a 5 can also mean “no usable answer”.

Topics come from a fixed list of twelve names. The model picks one to three for each episode, and any name not on the list is discarded in code, so the list below is complete:

• UFOs & UAP
• Aliens & Abduction
• Disclosure & Government Secrecy
• Alternative History & Ancient Civilizations
• Consciousness & Psi
• Paranormal & Hauntings
• Cryptids
• Occult & Esoterica
• Conspiracies & Cover-ups
• Fringe Science & Technology
• Mysticism & Spirituality
• High Strangeness

Tags describe the primary subject as the model reads it from the description. They can be broad, and an episode may carry a tag its description only touches on.

For each episode the model is asked to list three to ten terms from fringe domains (UFOs and UAP, paranormal phenomena, ancient mysteries, consciousness research, occult traditions, cryptids, conspiracy theories with specific claims) and to write a one or two sentence definition of each. The prompt tells it to leave out mainstream entertainment, finance, generic technology, sports, mainstream politics and generic wellness, and to return no terms for mainstream episodes. A small pattern filter in code removes a few obvious mainstream terms the model still returns. At most 15 terms are stored per episode.

A term with the same name (ignoring capitals) as an existing glossary entry is linked to that entry instead of creating a new one. The first definition written is kept.

Two things to know. Definitions are written from the model's general knowledge, not from a cited source, and can be wrong or out of date. And a term's episode count is the number of episodes whose description led the model to list that term. It is not how often the word is spoken, so a common term can show a small count.

The title, document ID, agency, dates, location, official description, redaction flag and file links on a PURSUE file page are taken from the catalog published at war.gov/ufo and reproduced as published. “Flagged as redacted” means the releasing agency's own catalog says so; Fringe Index does not inspect the files.

The AI summary on a record page is a two to three sentence plain-language summary from the same model, given only the record's official title, agency, incident details and official description, with the instruction not to speculate beyond that description. Topics use the same twelve-name list as episodes. Documents do not receive a Fringe Score. The AI summary is labeled on each page; the official description above it is the authoritative text.

Search is semantic. Your query is converted to an embedding with OpenAI's text-embedding-3-small model and compared with the stored embeddings of each episode (built from its title and AI summary) and each PURSUE record (document ID, title and summary) in a vector database. Search therefore covers titles and summaries. It does not search inside transcripts.

Results below a relevance floor are dropped. For short name-like queries such as “Roswell” or “AARO”, a weaker match must also contain the query words in its text, because embeddings of rare names are unreliable. When nothing clears the floor, the search page says so instead of showing unrelated items.

“Related episodes” and “Possibly discussed in” use the same embeddings and a similarity cutoff. They show similar content, which is not the same as a citation. An episode listed under a document has not been verified to discuss it, and the cutoff for documents has not yet been validated by manual grading.

Where a transcript is available it comes from the publisher (a transcript link in the feed), YouTube captions, or automatic speech recognition (Whisper models, run by Fringe Index or by a hosted speech-to-text service). Full transcripts are not published. Episode pages show a short excerpt of about 300 words with attribution and a link to the original.

Before an excerpt is cut, speaker markers and bracketed sound labels are stripped and a leading sponsor message is skipped using simple phrase rules. Those rules are conservative and can miss, so an excerpt may still start with an ad or contain speech-recognition errors such as misspelled names. Podcast transcripts are not used to generate the summaries, scores or topics described above; YouTube captions are the one exception (see Summaries and takeaways).

• Episodes: the public RSS feeds and YouTube channels of the shows on the Sources page. Each episode page links to the original so you can listen at the source.
• Documents: the US government's PURSUE release catalog at war.gov/ufo.
• Models: OpenAI (summaries, scores, topics, glossary, embeddings) and Whisper (speech recognition).

Inclusion of a show or document is not an endorsement, and the claims made in indexed episodes are not verified here.

The archive is not real time. Episodes are added in batches after they are published, so a new episode can take days to appear. The “Archive last updated” date on the home page is the date the newest episode was added to the archive, not the date it was published.

PURSUE releases are fetched by hand from war.gov, which blocks automated downloads, so a new release can appear a few days after it is announced. New items are also available as an RSS feed of episodes and an RSS feed of PURSUE Files.

• Summaries, takeaways, tags, glossary definitions and scores are machine-generated, can be incomplete, wrong or misleading, and are not reviewed by a person.
• They are based on show notes or captions, not on the full audio, so an episode can cover more, less or something different from what the notes say.
• Fringe Scores describe how unconventional a subject is, not whether it is true. They are unvalidated and can be inconsistent.
• Related-item links are similarity matches, not citations.
• Search looks at titles, summaries and key takeaways (for documents, also the description, agency and location), not at full transcripts. A name mentioned once inside an episode will not be found unless it appears in one of those fields.
• Coverage is limited to the listed sources and may lag the originals.

Always check the original episode or document before quoting or relying on anything here.

Spotted a wrong summary, score or definition, or are a rights holder with a request? A formal corrections and takedown process is still being set up. Until then, contact us via the Contact page, and see the Legal page for the site's disclaimer.