Related research — not part of INSPEKTORIUM
ASCR
A preregistered hypothesis: several known single-behaviour directions in language models — refusal, abstention, uncertainty reporting, evaluation awareness — may be special cases of one more general family of task-induced control representations.
Framing
ASCR does not claim or attempt to establish subjective experience, feeling, consciousness or sentience. “Appraisal” names a set of situation variables, and “control” names an influence on how a model chooses its response strategy — both are used as technical terms.
ASCR is a working abbreviation, not an established field term. It is a separate research project by the same author and not a module or capability of INSPEKTORIUM.
The research programme
A hypothesis, registered before the data.
Research question
Do variables such as signed relevance, uncertainty, controllability, norm tension, goal congruence and response pressure form shared latent control representations that partially govern how a model chooses its response strategy across domains?
Hypothesis
The key test separates concept tracking — a prompt that merely names a concept — from task induction, where the model’s own task really contains ambiguity, conflicting constraints or low controllability. The design is a 2 × 2 factorial: actual task state × concept mention.
Method
A planned feasibility pilot, not a powered study: three axes (uncertainty, norm tension, controllability) across four task domains on an open-weight 7–9B instruction-tuned model, with linear probes and later causal interventions against matched-norm random and PCA controls.
Current version
v0.1.2 — a pre-data methodological correction, released on 1 September 2026.
Current evidence
None yet. The first study (Mini-0) has not run, and no model-run data, activations, statistics or results exist. A technical smoke test on a small 0.5B model checked the pipeline only and is excluded from any scientific analysis.
Tests / preregistration
Hypotheses, falsification criteria, analysis plan and controls were published before any data. Version 0.1.2 repaired a decision statistic before data collection; earlier versions remain unchanged.
Known limitations
The novelty is a combination, not new parts: each component exists in prior work, and the contribution is a unifying, falsifiable synthesis. The exact model revision and run plan are not yet frozen.