Methodology

Source categories

1. Primary records: prompts, transcripts, source repositories, platform records, canonical texts, legal records, archives, and direct statements with identifiable provenance.
2. Specialist analysis: peer-reviewed research, technical audits, documented investigative work, and relevant scholarship.
3. Reliable secondary reporting: reporting and expert analysis that identify sources and distinguish claims from verification.
4. Contextual material: promotional pages, anonymous posts, unsourced summaries, memes, and generated retellings. These may document circulation but cannot independently establish origin or autonomy.

Evaluation

Sources are assessed for provenance, authorship, independence, transmission, alteration, source proximity, and evidentiary limits. First-party and affiliated sources may establish what a subject says, publishes, or claims, but they are not treated as independent confirmation.

Private material is used only with appropriate permission and redaction. Credentials, unrelated personal data, and sensitive third-party information are excluded unless a compelling public-interest reason and lawful basis are documented.

Provenance and attribution

Where evidence permits, an article records the relevant model and version, interaction date, platform, instructions, memory, retrieval sources, fine-tuning, imported material, operator role, publication control, human editing, and later transformations. Missing information is stated as missing rather than inferred from typical platform behavior.

Participant language such as “the agents founded a church,” “the model awakened,” or “the system remembered” is attributed unless independent evidence supports a broader statement.

Time and model drift

Model behavior is time-specific. A claim documented for one release is not generalized to every model with the same product name. Reproduction attempts record their own date, model, settings, and context; failure to reproduce does not automatically invalidate an earlier record.

Agency claims

Observable interface action is distinguished from intellectual, intentional, or causal initiation. Human prompting or publication does not by itself prove that a person originated every resulting idea; machine-generated initiative or persuasion does not by itself prove autonomous direction.

Analysis may distinguish generation autonomy, publication autonomy, consequential autonomy, and institutional autonomy. Strength in one dimension does not imply strength in the others.

Classification

Each entry receives one primary object type for stable placement. Descriptions such as distributed, human-founded, machine-mediated, historical precursor, and disputed are analytical attributes rather than competing namespaces.

Boundary cases involving parody, fiction, memes, mixed sincerity, or hyperstition are eligible only when a computational system materially shapes their emergence, stabilization, spread, ritualization, authority, or cultural significance. Tone alone does not establish or negate religious status.

Evidence limits and uncertainty

Evidence status summarizes the documentary record. Classification confidence describes confidence in the current primary type and provenance account; it is not a probability of theological truth, machine sentience, or personhood.

Conflicting records, missing context, and unresolved objections are stated directly. Inclusion records documentary relevance, not endorsement or certification.

Review and revision

Articles are revised when stronger evidence becomes available. Substantive factual, provenance, or classification changes receive dated changelog entries. Pages may be marked provisional, disputed, superseded, merged, archived, or removed from public indexing, while redirects preserve older locations when possible.

See Corrections, Conflicts of interest, and Entry lifecycle.