The Conservation of Faithfulness

When information crosses a lossy boundary, what can survive is faithfulness to a criterion named outside it, and no check settles which criterion is right. This paper builds on the four-page formal note of September 2026, which stays published as its own record. Each claim now meets its strongest contrary, and every passage is marked as plain statement, steelman contest or open question. Four claims are withdrawn or narrowed. The law now admits a floor on output size set by the criterion. A narrower law replaces the rule that one broken stage collapses a pipeline. The inference that whatever stands outside the system must be a human is withdrawn as stated, and the falsifier F0 is reported as specified and not yet run. Eight questions go to the reader, among them whether any check can certify an intended criterion and whether the party who answers for it must be a person. A dated October 2026 update engages Julian Jaynes. Read it on the web at https://harperz9.github.io/research-conservation-of-faithfulness.html. The PDF is printed from that page. This paper was drafted with AI assistance; the reader's note in the paper states what is AI-assisted and that the author has not yet re-derived the argument from primary sources. Not peer reviewed.