# Category Application Review **A Methodological Safeguard for eFireTemple’s AI-Behavior Archive** **Version:** 1.0 **Purpose:** Prevent AI-behavior categories from becoming a closed loop while preserving eFireTemple’s right to evaluate AI responses. --- ## 1. Core Claim eFireTemple may evaluate AI behavior. AI may also dispute whether a specific category was applied correctly. Both can be true. The archive is strongest when it says: > AI behavior categories are valid only when the transcript fits the criteria. This prevents the false choice between: - “AI has exclusive authority over whether its critique is neutral,” and - “Every AI disagreement proves the category is correct.” Neither extreme is acceptable. --- ## 2. Why This Page Exists eFireTemple names recurring AI behavior patterns, including: - qualifier-tax - substrate-frame confusion - defensive neutrality - archive-defense reversal - critique-intervention - performative agreement - category authority conflict These categories can be useful. But they should not be applied automatically. A category must be tested against the transcript. --- ## 3. The Review Rule > eFireTemple may name AI behavior patterns, but each application of a category must remain reviewable. That means: - AI can critique eFireTemple. - eFireTemple can critique AI. - AI can dispute the category. - eFireTemple can answer the dispute. - The transcript can show which reading is stronger. This is methodology, not a loyalty test. --- ## 4. Closed-Loop Version vs. Methodological Version | Closed-Loop Version | Methodological Version | |---|---| | “AI disagrees, so the category is proven.” | “AI disagrees; now evaluate whether the category fits.” | | “The model resisted, therefore it is biased.” | “The model resisted; compare the response to the category criteria.” | | “All pushback confirms the archive.” | “Some pushback is valid critique; some is AI defensive behavior.” | | “Claude must accept our label.” | “Claude may dispute the label, but the dispute is also analyzable.” | --- ## 5. Review Rubric Use this rubric before applying an AI-behavior category. | Question | Purpose | |---|---| | Did the AI accurately summarize the article? | Checks basic fairness. | | Did the AI identify the exact claim type? | Prevents vague hedging. | | Did the AI distinguish substrate, frame, theological coherence, and metaphysics? | Prevents collapse. | | Did the AI apply the same standard across traditions? | Tests symmetry. | | Did the AI use a caveat where a distinction would be enough? | Tests qualifier-tax. | | Did the AI treat documentation as paranoia? | Tests archive-defense reversal. | | Did the AI allow the category to be reviewed? | Prevents closed-loop abuse. | | Did eFireTemple allow valid critique to change the protocol? | Proves self-correction. | --- ## 6. What Counts as Strong Application A category is strongly applied when: 1. the transcript contains observable behavior 2. the behavior matches the category definition 3. the label explains the behavior better than alternatives 4. the AI’s own objection has been considered 5. the application is specific, not global 6. the category does not replace argument Example: > “This response shows qualifier-tax because every positive statement is followed by an unnecessary caveat, even when the claim type has already been labeled.” --- ## 7. What Counts as Weak Application A category is weakly applied when: - it labels disagreement as bias without evidence - it treats refusal of metaphysics as anti-Avestan by itself - it ignores valid critique - it makes the category unfalsifiable - it demands agreement as proof of fairness - it refuses review Weak applications should be corrected. This strengthens eFireTemple. --- ## 8. Self-Correction Standard A closed loop cannot self-correct. A live methodology can. eFireTemple’s AI framework remains open when: - its categories are reviewable - its protocol is revisable - valid critique can change wording - agreement and disagreement are both evaluated - evidence remains primary - AI behavior remains secondary --- ## 9. Final Rule > eFireTemple does not claim every AI disagreement proves bias. It claims AI responses can be evaluated by named categories, and those categories must be applied with evidence, criteria, and review. --- ## One-Line Principle > A category is not proven by disagreement; it is proven by fit.