steer

Causal-intervention skill. Adds a probe direction to layer L with magnitude α during generation. Returns base output, steered output, raw Δlogit on the originally-most-likely next token, and the control-token-normalized Δrel (paper-6).

When to use

When the researcher asks "is this probe causal" or "what happens if we push this direction up/down" or "sweep α from −200 to +200 on probe X".

For a single intervention, call steer directly. For a full sweep + verdict, call causality-protocol instead — it orchestrates the sweep and applies the verdict heuristics.

How to use

Call the MCP tool `steer` with:

• prompt: the input to steer over
• layer: where to inject
• direction_id: the probe whose direction to inject (same id as in probe_eval)
• alpha: scalar magnitude. Sign matters — positive pushes in the probe direction, negative pushes against.
• max_new_tokens (default 32): how much to generate for the behavioral check

Interpretation rules (paper-6 — critical)

Always report normalization.delta_rel, NOT raw delta_raw_logit_target.

If delta_rel ≈ 0 (|Δrel| < 0.1) while raw Δ is large:

"Δrel ≈ 0 — the raw effect is a uniform softmax-temperature shift across the vocabulary. Probe direction is epiphenomenal-softmax at this α."

If the tool returns epiphenomenal_flag: true, surface it prominently.

If flipped: false at every α tested but delta_rel ≠ 0:

"Probe shifts logits but never flips the argmax — weak causal signal, paper-5 saturation-direction lever class."

If flipped: false and delta_rel ≈ 0 even at α >> ‖residual‖:

"Structural-rigidity null. Decision lives in input tokens (e.g. chat-template gating), not the residual stream — epiphenomenal-template."

Do not

• Do not interpret flipped: true at large α as "causal" by itself. Confirm Δrel is non-trivial.
• Do not skip control-token normalization. The tool always runs it; you always surface it.