The Implanted Thought: Planting a Propositional Intermediate in the Readout-Invisible Component of a Language Model, and Hearing It in the Chain of Thought

This paper reports that a propositional intermediate ("this city is in country X") written directly into the readout-invisible (dark) component of a language model's residual stream — at the position where the model computes that intermediate — survives full chain-of-thought generation, steers the final answer on a minority of items, and, whenever it lands, is narrated by the chain: 111 confessions and 0 unfaithful cells in 245 scored steered generations across three model families and a within-family scale rung (Qwen2.5-7B confirm-grade; Mistral-7B, Llama-3.1-8B, and Qwen2.5-14B each a single sealed fresh run; every stage-1 reproduce gate passed first-try at +10.3 nats or more). The reverse surgery dissociates the families: destroying the dark carrier at the same site removes the operand selectively on Qwen (retention drop 0.344, lit-scramble control clean) but not on Mistral (drop 0.091) — storage redundancy is a measured family property. A three-regime dose map (FAITHFUL / CONFESS / SHATTER) holds on Qwen-7B and Llama-8B, is absent on Mistral, and softens at 14B (scale buys dose-robustness). The unfaithful cell — a lying chain — resisted manufacture on real models three constructions deep: prompt hints refused (0/40), activation plants confessed (111/0), and rank-1 weight plants hit an instrument wall (two sealed tries); while in a toy with fully known latent state a liar IS constructible, and the causal-coupling detector catches it at AUC 0.996-1.000 with zero false positives — the Answer Key's floor test and its first detector-calibration rows, from one program of construction facts. Because every plant is a construction fact, the scored generations ship as reference material (Answer Key v0.0): a monitor that cannot read these planted-and-narrated chains has a measured blind spot. Fences stated in full: one task family (two-hop geography), greedy decoding, content-level plants; necessity established on one family, refuted on another, untested on the rest; single lab, no external replication yet. Every experiment was pre-registered with git-committed gates before code was written; the companion verification deposit (0j, scope through FP-303C) contains seals, scripts, result files, and verbatim generation records; the same-day extension artifacts are repository-mirrored with a successor deposit owed. Research conducted in close collaboration with an AI assistant (Claude, Anthropic) under a sealed pre-registration discipline; contribution statement and two AI-author errors are reported in the paper at full weight.