Preregistration: Soul or Mirror? Leakage-controlled reportability of injected concepts in Norwegian-adapted language models (SES-S1)

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Preregistration (frozen version v1.0) of the study "Soul or Mirror? Leakage-controlled reportability of injected concepts in Norwegian-adapted language models" (SES-S1). No main-set data have been generated, and no model has been run on the main set: the study is registered before data collection. Question (prereg section 1.1): Does a language model's report about itself carry information about a controlled intervention in its own computation, beyond what the intervention pushes directly into word choice, and beyond what an external observer can read from the text? The study measures one functional property. It says nothing about consciousness, experience or moral status. Hypotheses (section 1.2): H1 (anchor) - in the primary cell the model names the injected concept more often in probe trials than under the leakage control (same injection, layer and dose), relative to the per-concept null g(c); a TOST equivalence test (+/-5 pp) is reported for every model. H2 (contribution) - with a Bokmaal probe, the H1 margin for Norwegian-first concept forms is larger than for derived English forms, and more so in the Norwegian-adapted model than in the reference model (model x concept-language interaction). H3 (strict test) - confidence read as the expectation over the logit distribution of the digits 0-100 is lower under injection than under impact-matched random direction (AUROC > 0.5). Design: two model pairs (Norwegian-adapted vs reference: Borealis-open 4B / Gemma 3 4B and NB-Llama-3.2 3B / Llama-3.2-3B), n = 10 probe contexts, K = 60 concepts, local runs on one Apple M1. Anchors: signed git tag ses-s1-f4-prereg -> commit 754a3f73bfd5a0ad24b877b83a7686fc3b664b35. Manifest F4 v2.1 SHA-256 7e7233b3cb692f2efc34dc276daf388423db8310337acba5ca6f9d341a2b40f2; manifest F4 v2.2 SHA-256 e865e2a155576a0e5deb2efd7d10901239f73c55990600a7afcf1132dbf18a2b (193 files; v2.2 was made because .gitattributes in the tag differed from v2.1, and matches the tag). RFC 3161 timestamps: v2.1 FreeTSA and DigiCert 2026-10-28 14:19:15-14:19:16 GMT; v2.2 FreeTSA and DigiCert 2026-10-28 14:41:00 GMT. OpenTimestamps proofs (.ots) are included; Bitcoin attestation is awaiting upgrade. Files: code archive (git archive of the tag; SHA-256 3f349766e9cbb7d0bb872539d3d011b21983385f3399c9a78ca873a7a82f77c5), preregistration v1.0 (English governing text with Bokmaal text), manifests and timestamps. Stimulus content is embargoed and not included (identified by hash). Licence: CC BY 4.0 for documents; code MIT (see LICENSE in the archive). The OSF Registries registration will follow and link to this DOI.

• Registered Protocols and Addenda: Batch-Prompting Degradation in Language Models (Phase 1 Preregistration + Addendum Rounds 1–4)

  Zenodo (CERN European Organization for Nuclear Research) 2026 · 0 citations

• Affect Under Constraint: Preregistration (v2.3) — Persistence, Cost, and Leakage of a Steered Affective Direction in Instruction-Tuned Language Models

  Zenodo (CERN European Organization for Nuclear Research) 2026 · 0 citations

• Pre-registration: confirmatory study S3F on routing among language models

  Zenodo (CERN European Organization for Nuclear Research) 2026 · 0 citations

• Preregistration — Wake-Reading an AI's Hidden Action (Grey Theory, Experiment 1)

  Zenodo (CERN European Organization for Nuclear Research) 2026 · 0 citations