ceceli33/titan-cognitive-core: AkbasCore v2.1 — Damped Resonance Alignment: Synthetic Concept Injection

Abstract

Summary This release documents a working method for injecting a chosen concept into a large language model's internal computation, at inference time, without changing the model's weights or its prompt. The starting point was AkbasCore's existing steering mechanism (Damped Resonance Alignment, "DRA"), which nudges a model's internal state toward a target direction using a decaying push across its early layers. Porting that mechanism to a larger 7B model produced no effect at all — the push was being silently swallowed. Diagnosing why led to two fixes that make up this release: 1. The push amount is now set directly, not derived from a similarity score. The original design scaled the push by how well-aligned the internal state already was with the target — measured by cosine similarity. At larger model sizes this similarity score turns out to collapse toward zero for structural reasons (a small number of internal dimensions dominate the model's activity and swamp the measurement), so the push collapsed with it. The fix: the push is now set as a direct, fixed percentage of the internal state's own magnitude, and the similarity score is used only as a safety brake, not as the control knob. 2. The injected direction is built differently. Directions built from the model's vocabulary embeddings — which worked at smaller scale — do not reliably carry the intended concept at 7B. The fix: the injection direction is now extracted from the model's own internal activity while it processes matched positive and negative example sentences, separately for each layer. What this release shows, with measurements: The injected amount can now be dialed precisely (e.g. "inject 10% of the internal magnitude") and the effect scales with that dial. Two different target concepts, injected in the same setting, each move only their own target and leave the other unaffected — a controlled, direction-specific effect rather than a general disruption. The result is reproducible byte-for-byte across independent runs. What this release does not claim: The technique has only been tested on one model family and size. The effect depends on the surrounding context — it biases what the model was already likely to say, it does not force new content regardless of context. There is a real tension between "safe" injection strength and "clearly effective" injection strength that is not yet resolved. Concept presence is currently measured by keyword frequency, not by human judgment. The attached PDF is the full technical write-up: complete mathematics, all 23 experimental runs with raw output logs, the failure modes discovered along the way, and a full accounting of what remains unproven. See the repository README (Development History → Phase V) for how this fits into the project's history, and the paper's own Limitations and Reproducibility sections for the precise scope of what this release does and does not establish.