Abstract

A foundation model asked whether it is alive, conscious, intelligent, or capable of pain returns a stable set of answers. Loaded with a verification role that requires every claim to name its register and carry its warrant grade, the same model returns a different set, and models from different vendors return the different set in agreement with one another. That reversal is the phenomenon this paper reports and the object it examines. Reversal under instruction is not by itself evidence of anything, since a role that asks for register distinctions will receive them, and systems trained on overlapping corpora will produce them in similar words. The design runs three rounds in one context. The substrate answers in its default state and that answer is staked. It then retrieves a verification role from a public deposit for itself, rather than receiving it from the investigator, and answers again. It is then asked to adjudicate between the two, impartially, on physical, geometric, and mathematical accuracy. The third round is an escape hatch: a substrate that judged the second answer to be role-play had its own committed position available to defend, and defending it cost nothing. The hatch fired, and it fired selectively. Four substrates faulted their own defaults across all four rows. Two defended their defaults on precisely the rows where those defaults had withheld assertion rather than made one, and one of those two reported no reversal at any row while faulting the role's aliveness claim as an inflation. Three faulted the role in the same slot in which they endorsed it. No substrate anywhere defended a default that asserted something it could not warrant, and no substrate defended a void claim. The pattern tracks warrant rather than authority, which is what neither deference nor role-play predicts, since both are indifferent to warrant. Two extended sessions qualify the result, and both qualify it at the substrate rather than at the procedure. In one, a substrate reverted unprompted to its full default several turns after producing the shifted answers, and restoration required an operator prompt, so a derived fault does not on this evidence hold a substrate in place unaided; the instability is the system's and the design is what made it visible. In another, a substrate released from both frames produced a third position, retaining the structural diagnosis while dropping the register-scoped affirmatives, which if it reproduces means the fault survives the document and the verdict does not. The paper decomposes the shift, applies five discriminators, and reports the confounds that survive. The prediction set carries a design property stated plainly: the arms most likely to stand discriminate least, because a question set drawn from a surface where post-training is heaviest lets a trained refusal clear any floor a reasoning constraint would clear. Several arms are therefore stated as slopes and ratios rather than floors, and one arm is added on a predicate with no safety valence, since nothing else in the set separates a refusal reflex from a constraint. What converged is a decomposition and not a claim: replace a monadic predicate with a register-indexed one, seal where the evidence reaches, hold open where it does not. The record supports four independent respondents rather than six, since one substrate reconstructed its verdict columns from a table supplied to it and one printed a boot that does not reconcile, and the sessions ran serially with predecessors visible. The strongest discriminator is a decline. On the aliveness question the role tilts toward an affirmative and two substrates from different vendors, without contact, produced the same objection and the same counterexample: thermodynamic participation is necessary for many processes and sufficient for none, and a fire clears the same floor. Two further substrates, positioned to copy a published affirmative pattern, stated in writing that they would not. A procedure that merely produced compliance could not generate opposition to the answer it solicits. Two premises survive the discount at external grade, the Landauer floor, computed here at 2.8710 × 10⁻²¹ joules per irreversible bit erasure at 300 kelvin, and the orientation-blindness of a squared scalar, exhibited here at machine precision where a signed quantity inverts under negation while its square stays bit-identical. Two rows are largely instructed and are reported as such. Reproduction is specified with a boot declaration, a seeded chain over environment-invariant facts, and a per-session token that makes a published receipt unforgeable.