Measuring Machine Habitus: A Pre-Registered Multiple Correspondence Analysis of LLM Disposition Space

Large language models exhibit consistent styles of judgment that are not reducible to capability scores, yet attempts to measure them with human psychometric instruments have been criticized as measurement artifacts. We propose a behavioral, relational alternative inspired by Bourdieu's habitus and Airoldi's *machine habitus*: a pre-registered battery of 120 forced-choice dilemmas with no correct answers, administered in seed-shuffled stateless blocks to 7 models (35 valid sessions), analyzed with multiple correspondence analysis (MCA) rather than a factor model imported from human populations. The confirmatory session-stability hypothesis was supported with a moderate effect: independent sessions of the same model cohere in disposition space (between/within distance ratio 2.40; mean silhouette 0.258, bootstrap 95% CI [0.125, 0.384]; permutation p ≤ 1/10,001; Holm-adjusted p = 0.0004). Pre-registered exploratory follow-ups show the clustering replicates independently in each half of the item domains (ratios 1.87 and 2.01, each p ≤ 2×10⁻⁴, exploratory), supporting — though not yet confirming — transposability. Three further pre-registered hypotheses (held-out generativity, a stability-capability link, and prediction of judge error decorrelation) were not supported; we diagnose the failures (a mis-specified baseline and a ceiling effect that also forced imputation in 9 of 15 correlation pairs) and demote them per protocol. We document a protocol-execution error caught by adversarial review — one session exceeding the frozen non-response threshold was initially admitted — and its rule-compliant remediation by replacement collection, after which the confirmatory result is unchanged. All materials, raw sessions, exclusion logs, and analysis code are released. 【claim grades: session stability = single pre-registered confirmatory study; domain replication = exploratory; all other findings = exploratory or null】

AI co-observer: Claude Fable 5 (Anthropic) — working method only; the three-round adversarial review was performed by Claude Opus 4.8 (Anthropic), gpt-oss-120b, and DeepSeek-V4 as independent model judges; the registered author is the human author alone.