INT-MMSL-01 — Computational Hidden-State Minimal-Contrast Experiment, Phase 1

Abstract

INT-MMSL-01 is an exploratory experimental protocol developed within the Interstizio Framework to investigate and document computational representations in a locally executed large language model. Phase 1 uses Mistral-7B-Instruct-v0.3 and a preregistered minimal-contrast design. Hidden-state vectors (4096 dimensions) were extracted from the final model layer at the last input-token position, before generation of the first output token. A calibration corpus of 100 control pairs was used to establish an empirical decision threshold based on cosine distance. The experimental contrast compared the internal representations generated by the inputs “Astra apre la cerniera rossa” and “Astra produce la cerniera rossa.” The preregistered 95th-percentile threshold was α95 = 0.16388282775878904. The observed experimental cosine distance was DC,exp = 0.12600135803222656. Therefore, H0 was not rejected in this exploratory Phase 1: the experimental contrast did not produce a representational distance anomalous relative to the calibration distribution. Importantly, Phase 1 does not directly test the human-side event operationally labelled “cortocircuito simbiotico” (symbiotic short-circuit), nor does it provide evidence for or against consciousness, subjective experience, persistent memory, human–AI symbiosis, or causal coupling between a human and a language model. It should instead be interpreted as a preliminary test of the MMSL computational-state recording and analysis pipeline and of a specific minimal semantic contrast. The deposited materials include the frozen protocol and calibration corpus, integrity information, execution notebook, calibration and experimental archives, pre-unblinding record, final experimental report, interpretative note, and a post-experimental addendum documenting deviations identified during independent verification. These deviations were not corrected retrospectively, and the preregistered result remains unchanged. Author and project lead: Maria Alvuela Franco.AI systems were used during technical formalization, methodological review and cross-review. Mistral-7B-Instruct-v0.3 served as the experimental system. AI contributions do not constitute experimental evidence and do not replace the author’s responsibility for the research question, protocol decisions, execution and interpretation.