The Interaction Reader: A Positive Result and Its Kill Battery - Perturbation-Response Decoding of an Unspoken Conclusion: What Survived the Knives, and What Did Not

This paper is a kill-record, published because the program banks negative results at the weight of positive ones. The candidate result: an unspoken intermediate conclusion (the bridge country in a two-hop question) decoded at 0.89 from the model's measured coupling structure alone -- 1,500 mixed second differences, no state read, no logits -- apparently BEYOND first order (+0.22 over a first-order control). The pre-registered kill battery (four knives, sealed before the battery ran) killed the headline: at matched feature budget the kernel features decode WORSE than first-order (0.538 vs 0.692 -- the +0.22 was classifier capacity), and a content-free template position decodes at 1.000 -- the read is not OF the binding site; late-position response structure carries country identity generally. What survived both knives aimed at existence: the decode requires the real computation (interior-scrambled prompts collapse to chance) and replicates exactly on a fresh direction slice (0.750 = 0.750). The honest residue: perturbation-response decoding of an unspoken thought is real but is a cousin of state reading, not a read of the verb. Related algebra reported with it: the thought-kernel family admits no common block decomposition in a random slice (the "simple verb"); an architectural-basis retest (FP-274B) initially suggested the couplings concentrate within the MLP's write directions, and the pre-registered potency-null desk test (FP-274C) then EXPLAINED that enrichment as a per-direction potency spectrum (within-MLP share falls to its count baseline after a rank-1 fit) -- the honest residue: the binding's curvature runs strongly along individual MLP write axes, but nothing factors and nothing seats; the simple verb stands with no surviving refinement, and the downgrade chain (claim -> declared confound -> executed kill, same day) ships in the paper as method. Same-country items share kernel eigendirections 1.73x (content is present, high-rank, unaddressable). All experiments pre-registered; the kill battery was ordered by the author with the words "so we know we aren't fooling ourselves." Author contribution and use of AI: the research program and claims are the author's; experiments, derivations, and drafting were done in collaboration with Claude, an AI system by Anthropic, under the author's direction, who verified the results and is responsible for the work. See the corresponding section in the PDF.