Soft control

Where a causal probe switches a component fully off, soft control turns a dial. Instead of masking experts, it gently tilts the router’s internal preference scores (“logits”) toward a task and away from the rest, by a small tunable amount — soft router-logit biasing. The size of the nudge, written ε, is the steering knob.

Calibration — planned, not measured. Everything on this page describes a pre-registered experiment that has not been run. The conditions, gates, and decision rule below are design intent, locked in advance. There are no results yet, and the numbers (ε values, condition counts) are the planned sweep, not findings.

Question: can we move smoothly from “no perturbation” to “task-shifted output” while preserving competence?

Soft control = bias-magnitude knob, dose-response, coherence preserved across the transition. Hard ablation answers causal questions; soft control answers steering questions. This is the intended deployable primitive.

S2.1 — Soft router-logit bias

The deployable steering primitive. Same calibration source as S2.0 (humaneval-p99 / wikitext-p99 top-13/bottom-13 per layer), but the application is:

router_logits[task_top_experts]    += ε
router_logits[task_bottom_experts] -= ε

Top-k still operates normally. No expert is hard-skipped. ε is the steering knob.

Conditions (13)

• Baseline ε=0
• Humaneval-direction: ε ∈ {0.5, 1.0, 2.0, 4.0}
• Anti-direction: ε ∈ {1.0, 2.0}
• Cross-corpus (wikitext-direction): ε ∈ {1.0, 2.0}
• Random-direction control at every primary ε (budget-matched per-ε)

Decision gates

• G1 dose-monotonicity (principled humaneval-direction)
• G2 smallest-ε task-selective at quality metric (not just drift)
• G3 coherence floor (no degeneracy collapse at the working ε)
• G4 cross-corpus reversal at matched ε
• G5 anti-direction symmetry
• G6 per-ε random-bias trap
• G7 ε=0 sanity

SHIP = soft-control regime exists with quality preserved.

What soft control gives

• Continuous knob, not binary on/off
• Coherence-preserving steering (selection of small ε)
• Composability with task detector (S3.0) for runtime adaptive policy

Why this is the deployable primitive

Per 2026-05-04 strategy note: hard task masks are expensive and blunt. Soft bias amortizes calibration across domains and gives explicit dose-response control. Enables runtime task-conditional policy without per-domain calibration burden.

Status

• S2.1 prereg: cross-check/preregistry/s2_1_soft_bias_router_steering_2026-05-04/
• Run: gated on S2.0 completion (provides hard-mask ceiling reference)