Retrieval is the baseline, not the enemy

A residual control vector can tilt a model toward a domain, style, or semantic posture. It is a constant direction applied broadly; it does not naturally bind a query to one new arbitrary fact. At larger amplitudes it can simply destabilize generation.

The early program repeatedly crossed the same boundary: cvecs changed the space of likely answers, while prefixes and logit interventions supplied the answer-bearing content. That difference now anchors the experimental vocabulary.

Encoding a prefix into KV state can remove visible tokens from the prompt and still be functionally equivalent to the prefix. That is an implementation win—latency, prompt budget, or caching may improve—but it does not prove the model learned from experience.

The primary claim requires fixed-shape learned state whose effect depends on the taught rule, survives deletion, transfers to unseen operands, and disappears under zero/swap interventions. A cached correction-produced KV trace is explicitly forbidden in Research 20.

The subtractive and additive matrices asked every organ to earn its place on held-out behavior. The scope-slot reranker produced the only decisive additive effect and was retained as RETRIEVAL-BASELINE. Answer-time hippocampal retrieval added exactly 0.0 in the additive test; DSI was unsupported at the tested power and net-harmful in the full stack; critic, identity, immune, and autoencoder paths did not earn KEEP.

Research 15 therefore became vacuous: there were no non-retrieval survivors to rewire from strings into tensors. That is a useful closure, not missing implementation.

Scope-slot reranker

RETRIEVAL-BASELINEAdditive S0 +0.667 and S4 +0.250 with zero seed variance.

Answer-time hippocampus

ARCHIVEAdditive delta exactly 0.0.

DSI

ARCHIVE / APPEAL OPENUnsupported in additive v2 test and −0.060 in the full stack.

Tensor wiring

VACUOUSNo non-retrieval component earned KEEP.

Future cortex experiments keep three different questions on the same page. Vanilla asks what the frozen specialist already knows. Retrieval asks how well a byte-matched content store performs. Oracle asks whether the specialist can express the behavior when fully informed.

The primary cortex condition must beat its no-update and random-update controls, depend on semantically correct feedback, survive trace deletion, and depend causally on the right cortex state. A retrieval win is reported as a retrieval win; it cannot accept the metabolism hypothesis.