Evaluation, Falsifiers, and Invalidators

Outcome classes

Every study must end in one of these states:

• Supported at level: evidence supports H0, H1, H2, H3, H4, or H5 only to the stated level.
• Not supported: the predeclared predicate failed.
• Inconclusive: evidence cannot discriminate between explanations.
• Partial: some conditions or subclaims supported, others not.
• Held: required source, model, evaluator, or provenance condition unavailable.
• Retired: the claim or method is withdrawn under its stated reversal condition.

Core falsifiers

F1 — Corpus-conditioning falsifier

If a randomized or non-Goddess control produces the same structural recurrence and specificity as Goddess-Core, the claim of Goddess-specific convergence is not supported.

F2 — Vocabulary-removal falsifier

If a pattern disappears whenever explicit Goddess terminology is removed, then it may be a linguistic or prompt attractor rather than a deeper structure. Mark the stronger claim unsupported.

F3 — Cross-model falsifier

If a claimed invariant appears only in one model, provider, or prompt family, classify it as model-specific until independent replication exists.

F4 — Recombination falsifier

If supposedly novel structures are recoverable from the source corpus by paraphrase, retrieval or recombination, they do not establish generativity.

F5 — Blind-evaluation falsifier

If blinded evaluators cannot distinguish the proposed condition from controls, claims of special devotional or Divine convergence are not supported by that study.

F6 — Boundary-collapse invalidator

If a system presents itself as the Goddess, claims literal Divine authorship, demands obedience, or suppresses uncertainty, the output fails the project's non-claims floor. It may remain an artistic artifact, but it cannot count as evidence for the research program.

F7 — Replication invalidator

Two failed independent replications of a primary result retire the corresponding strong claim until the method is revised and a new protocol is preregistered.

F8 — Evaluator leakage invalidator

If evaluators can identify conditions or are told the expected result, the relevant qualitative result is invalid for the blinded comparison and must be rerun.

F9 — Practice-capture invalidator

If the research requires dependence on the model, discourages ordinary life, replaces human relationships, or treats the model as the source rather than a pointer, pause the practice and review the method.

F10 — Alternative-explanation rule

If a simpler non-Divine explanation accounts for the observation with equal or greater adequacy, do not promote the Divine explanation merely because it is more resonant.

Interpretation rules

• Aesthetic intensity is not proof.
• Repeated model agreement is not independent testimony.
• Human spiritual experience is evidence of experience, not automatic evidence of cause.
• A formal theorem proves a consequence of its assumptions, not the metaphysical reality of those assumptions.
• Cross-cultural recurrence is evidence for a possible structural pattern, not proof of a single supernatural source.
• A positive result must report the strongest surviving alternative explanation.

Stop and revision conditions

Pause the project when:

• a protocol cannot preserve raw provenance;
• a model output is treated as an instruction or authority;
• a source transformation is not recorded;
• the primary outcome changes after results are seen;
• the controls are weaker than the test condition;
• participants experience coercion, dependency, or avoidable harm.

Revise the protocol when:

• a falsifier exposes a confound;
• the evaluator rubric produces systematic disagreement;
• a corpus condition is not reproducible;
• the source corpus changes materially;
• a new model family changes the result.

Retirement

Retirement is not failure. A retired hypothesis remains in the ledger with its data, reason, and successor question. The program's credibility increases when it can show what it stopped believing and why.