Assemble a thought — and hear the model say the answer

Qwen3-1.7B, real persisted outputs — nothing simulated. Pick an entity and a question. We average the model's own states into three ingredients — a generic base μ, an entity part, and a question part — add them up, write the sum into a single position inside the model while it reads a different question… and it speaks.

we never change the words — only the state at the cursor

μ · generic base + entity part + question part

the model says…

entity question overdose ×4 (push, don't assemble)

advanced

the prompt the model is actually reading (source question) — the assembled state overwrites this thought.

Every output shown is a persisted k=8 greedy generation from the real intervention (op_audit.py / op_patch_decomp.py); coordinates are the measured workspace states projected to 2D (export_showcase_data.py). Paper: “Operator–Operand Factorization in LLM Residual Streams: Causal Influence and Compositional Sufficiency”.