● Published
2026
DOI: 10.5281/zenodo.20531347
GPL-3.0 · CC-BY-SA-4.0
Proprioceptive Channels: a Native Behavioral Control Axis for Language Models
Mario Gutierrez · Celiums Research Labs
We inject six external cognitive-state channels — memory, affect, time, ethics, identity, and continuity —
into a frozen Gemma 4 E2B model through a small gated cross-attention adapter, and show the model learns
to use them across all six capabilities. Under direct conflict, where the text prompt and the channel state
disagree, the channel overrides the prompt 264 of 265 times: a measurable, perpendicular axis of behavioral
control that is orthogonal to in-context instruction. We further reproduce the effect from scratch and report
the honest negatives along the way.
proprioception
conditioning
Gemma
interpretability
● Published · preprint
2026
DOI: 10.5281/zenodo.20436643
Apache-2.0 · CC-BY-4.0
Hyphae v2: Hash-Chained Verbatim Quotation — a Verifiable Provenance Layer for Grounded Retrieval
Mario Gutierrez · Celiums Research Labs
Hyphae is a cognitive substrate that answers grounded queries by verbatim quotation of hash-chained memory
fragments — with no LLM in the cognition path, and a cryptographic audit relation between every output and
its sources. Faithfulness is true by construction: every answer is a tamper-evident quote of a stored
fragment, not a generated paraphrase. Version 2 implements this provenance layer and measures it against
retrieval baselines.
verifiable provenance
grounded retrieval
hash chain
machine memory