Reading a Deployed 1.58-bit Language Model with the Standard Toolkit: A Component-Level IOI Circuit Matching Full Precision

Ternary large language models—weights restricted to {-1,0,+1}, ≈ 1.58 bits per parameter—now match full-precision (FP) models of the same size at a fraction of the memory, and open-weight checkpoints such as microsoft/bitnet-b1.58-2B-4T are being deployed. Yet mechanistic interpretability had reached ternary networks only at toy scale—small from-scratch transformers on modular addition—and it was not obvious that the standard toolkit (sparse autoencoders, attribution patching, integrated-gradient circuits) even applies to a deployed model when weights are {-1,0,+1} and activations pass through round. Our primary contribution is methodological: the toolkit transfers, establishing the deployed ternary LM as a tractable model organism. The straight-through estimator (STE) used to train BitNet makes gradient-based attribution well-posed on the inference graph despite round(·) having zero gradient almost everywhere: integrated-gradient completeness is 1.01±0.11, and EAP-IG agrees with exact activation patching (Spearman 0.70–0.86 across our tasks). Using it, we recover, to our knowledge, the first full causal circuit of a deployed ternary model: an indirect-object-identification (IOI) copy+suppress circuit read out by a sparse set of late attention heads—positive name-movers L22H19 and L27H13 that copy the indirect object, and an S-inhibition head L22H17 that suppresses the subject—hardened over three templates. This structure matches the canonical FP IOI decomposition at the level of which components matter: on a confound-free real-scale twin with a true FP counterpart (Spectra TriLM/FloatLM 2.4B), the component-level importance ranking agrees at Pearson 0.93 (a single-run corroboration, n=50). We are explicit about the boundary of this claim: the deployed 2B has no FP twin, so head-level identity cannot be tested on it directly; on the one twin where it can, the head-level comparison fails its own reliability self-test, and where measurable the specific name-mover and S-inhibition heads do not coincide (head-level Jaccard ≈ 0)—so we claim a component-level, not head-level, match. At the feature level—measured on a matched from-scratch toy twin, and corroborated on the 2B by a cross-model crosscoder—ternarization shares the FP feature inventory along rotated directions, with no private ternary vocabulary. Ternarization changes the parameters, not the algorithm, at the component and feature levels. We report the heterogeneous and negative results (the read-out is not a uniform name-mover population; "zero exclusive features" required a positive control; head-level identity is untested on the deployed model) as part of the contribution.