What a language model does when you poke it: the clinical map of the Perturbational Complexity Index, pre-registered and reproduced in Pythia-410m

Abstract

The Perturbational Complexity Index (PCI) decides whether a human is conscious without asking them: a brief transcranial pulse, and the EEG tells whether the response is long and complex (wakefulness) or short and simple (anaesthesia, deep sleep). This record applies the same logic to an open language model that nobody built for the purpose. Pythia-410m is run in a closed loop over its own text (soft feedback: the expectation over words is fed back — an input the model never saw in training, so the apparatus is a system of its own and says nothing about deployed models; the sampled loop of ordinary generation was measured separately, and its calibration against the weight null models failed — its stage is open); a single pulse is injected into the residual stream at one block; under a running noise field the apparatus records, at six blocks, how many units still deviate from the unperturbed run at each step. The duration τ of the response (in words) and its form (uniformity across noise repetitions, noise sensitivity of the unperturbed run) were calibrated in advance against the null models chaos and freezing at three model sizes and then never changed; the healthy value was reproduced fifteen times. Interventions were translated by biological function — damping of all weights (anaesthesia), amplification (psilocybin), unit loss (neurodegeneration), context masking (amnesia), activation threshold (GABA), gain adaptation (fatigue), preamble (state), truncated prompt (sensory deprivation), rounding (substrate change), a bridge between two models (conjoined twins), common firing of all units (seizure) — in 36 experiments, each pre-registered before the first number with predictions, probabilities and tipping conditions. Of 190 bound predictions 93 were met and 25 tipping conditions were triggered, each with a pre-stated meaning. The clinical map appears in the model in the same order as in humans. A second readout, the fingerprint (Spearman of τ-profiles over 40 prompts), identifies the individual: repeat 0.99, siblings of identical architecture 0.32, states 0.4–0.7; in three unrelated interventions the person is lost before the stage. A third readout, the boundary: a transformer is one stage, and whether two models are one is decided by the speed of the return path, not the strength of the bridge. What the model cannot do — split-brain, half-brain sleep, deep sleep by a per-unit switch, hysteresis after a seizure — has a named, retrofittable architectural cause. Thirty-four predictions for human TMS-EEG follow. All failures are reported. Files: the measurement paper (version 1.0), the reproduction package (all scripts, batch files, seeds, 379 expected values and a checker; MIT licence), the pre-registration documents with their result sections appended unchanged (CC BY 4.0), all logs and result files (CC BY 4.0), the consistency report v7 (assignment to the theory Process-Identity Theory, PIT), the Phase 1 and Phase 2 reports, and a statement on what this work contributes to the question of model welfare — and what it does not. Hardware: one consumer GPU (RTX 4070 Super, 12 GB); a full re-run takes about 60 GPU-hours. AI involvement: Claude (Anthropic), a large language model, wrote the apparatus code, the pre-registration documents, the evaluations by wording and the drafts of all reports under the author's decisions; this is declared in every document, and no self-report of the model enters any result. Models: Claude Opus 5.5 (Anthropic) for the model-network programme, the calibrated re-measurements and the first language-model branch (September 2026), and for two external readings of the final drafts; Claude Fable 5.1 (Anthropic) for the pre-registered language-model experiments, the measurement paper, the theory paper and the deposited packages (October 2026). Theory: Process-Identity Theory (PIT): theory paper, canon document and the model-network programme are a separate record, doi:10.5281/zenodo.23280480.