R / 01 — RESEARCH RECORD
Proprioceptive channels as a perpendicular control axis
A bounded experiment asking whether cognitive self-state injected outside the token stream can exert causal control over generation.
01 / BOUNDED RESULT
What happened
Under direct conflict, the channel determined output in 264 of 265 frozen-base cases; the effect replicated from scratch on 88.8% of 455 counterfactual pairs.
02 / METHOD
How it was tested
The study tests six channels—memory, affect, time, ethics, identity, and continuity—read at every layer through ReZero-gated cross-attention.
Two tracks separate mechanism from training: an adapter on a frozen Gemma 4 E2B base and a native 110M-parameter replication trained from scratch on 1B tokens. Evaluation uses direct conflict and held-out counterfactual pairs.
03 / NEGATIVES AND LIMITS
What the result does not support
Toy scale, in-distribution held-out data, and one final native replication. The work does not demonstrate general capability, out-of-distribution transfer, or validity at scale.
The injection is Flamingo-lineage, not claimed as a new primitive. Teacher-generated corpus text is not redistributed; the record publishes the method and prompts for regeneration.
The evidence supports a causal property within the tested scope. It does not establish consciousness, subjectivity, real identity, or that channels improve any model or task.
04 / ARTIFACTS
Read, verify, and cite
Gutierrez, Mario. “Proprioceptive Channels: Cognitive Self-State as a Perpendicular Control Axis in Language Models.” v1.2.0, 2026. doi:10.5281/zenodo.20531347.
Back to all research