Practical steering

Practical steering is the runtime system the whole ladder is built toward: a prompt arrives, the model reads its own routing to guess the task (a calibration-free routing readout, S3.0), and that guess picks a soft-bias direction (S2.1) — all with no per-prompt setup and no retraining. The two halves are the task-agnostic substrate (computed once, shared) and the task-conditioned overlay (thin, per-prompt).

Calibration — planned, not measured. This page describes the intended end state. S3.0 and S4.0 are pre-registered and not yet run; S4.0 is additionally gated on S2.1 and S3.0 shipping first. Nothing here is a demonstrated result — it is the program’s target configuration with decision rules written in advance.

Question: can we steer at runtime, with no per-prompt calibration, preserving quality?

Practical steering = task-agnostic substrate + lightweight task-conditioned overlay. Builds on causal probes (S1.0, S2.0) and soft control (S2.1) and is intended to ship the deployable runtime configuration.

S3.0 — Routing-readout task detector

Per-prompt forward → top-k expert distribution per layer → cosine similarity vs cached task priors → predicted task. No calibration data per prompt, no probe training, sub-second overhead.

If S3.0 ships: cheap inference-time task inference enables prompt-conditional soft-bias direction without forcing the user to label the prompt.

Prereg: cross-check/preregistry/s3_0_routing_readout_task_detector_2026-05-04/.

S4.0 — Runtime composition (S3.0 ⨂ S2.1)

prompt arrives
  ↓
S3.0 fingerprint forward pass
  ↓
predicted task (humaneval-aligned vs wikitext-aligned vs ...)
  ↓
S2.1 selects bias direction at small ε (from S0.0 ε_safe atlas)
  ↓
soft-steered generation

Zero training. Reusable substrate. Lightweight task overlay. Composable.

Prereg: deferred until S2.1 SHIP and S3.0 SHIP land.

What practical steering gives

• Runtime control without per-domain calibration
• Coherence-preserving task adaptation
• Cheap to deploy (no probe training, no SAE infrastructure)

What is required before practical steering ships

• S0.0 universal safety map: identify ε_safe per (component, layer) — substrate
• S2.1 soft-bias dose-response: confirm soft control regime exists
• S3.0 routing-readout: confirm task detection at sub-second overhead
• S4.0 composition: confirm composing S3.0 ⨂ S2.1 ⨂ S0.0 produces equivalent selectivity to oracle-task S2.1 within 1.2× ratio

Each gate is pre-registered. Practical steering claim depends on all four passing.

Status

• S0.0 prereg: cross-check/preregistry/s0_0_universal_safety_map_2026-05-04/
• S2.1 prereg: cross-check/preregistry/s2_1_soft_bias_router_steering_2026-05-04/
• S3.0 prereg: cross-check/preregistry/s3_0_routing_readout_task_detector_2026-05-04/
• S4.0: deferred, gated