# An Alignment Claim Without a Reference Frame Is Unfalsifiable ## Claim Calling a model "aligned" says nothing testable until it specifies aligned to which values, whose values, and measured by which test — without those three, the claim cannot be wrong and so cannot be evidence. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Aimed at vendor marketing and informal usage, where "aligned" appears with no stated target. - Does not apply to technical alignment research that already names its target (a published constitution, a reward model, a documented value set) — those claims are falsifiable against that target even when the target itself is contested. - The fix is cheap: naming the reference frame converts an unfalsifiable claim into a narrow but checkable one. - Same structure as any unqualified comparative — "faster", "safer", "more accurate" — that omits the baseline. ## Evidence From the WordPress Accessibility Day 2025 session on AI ethics: the argument is that alignment scores are quoted without ever stating the value set being aligned to. ## Links - [[A Ceiling Score Reported by the Scored Party Measures Test Fit]] — the companion defect — that note attacks who administered the test, this one attacks whether the test names what it measures - [[A Threat Taxonomy Is a Policy Decision Not a Neutral Description]] — the same move one level up: the category scheme a safety claim is measured against is chosen, not given - [[3. Resources/MOCs/Evidence & Claims Literacy MOC]] — the claims-literacy thread on missing baselines, which is what a reference frame is. - [[3. Resources/Permanent Notes/A Variance Test That Always Finds a Difference Cannot Show a Group Deviates]] — has evidence from: In a 9:50 excerpt, two speakers describe a method for judging religious groups, which they call finding the variance: compare what a group teaches to a baseline and ask which way the difference points.