Feature steering in pre-deployment review — when mech-interp findings drive intervention design, not just measurement

Anthropic's confirmation that mechanistic interpretability now feeds into intervention design during pre-deployment safety review — feature steering, SAE-driven ablation, circuit-level patching — is the methodological extension from passive measurement to active intervention. The closed feedback loop between interpretability finding and deployment-ready artifact is now the production methodology, with regulatory consequences following.

The methodological extension is the substantive piece. Through 2024-early 2026, mech-interp methodology produced findings — sparse-autoencoder features identified, circuits associated with specific behaviors, quantitative stability measurements under adversarial pressure — that informed deployment decisions but did not directly drive interventions. The May 2026 update is that the pre-deployment process now uses feature-steering to actively modify model behavior on identified risk axes before release. The post-intervention behavior is then measured against the same SAE feature surface, producing a closed feedback loop between interpretability finding and deployment-ready artifact.

The brittleness finding that lands alongside the methodology extension is consequential. The new arxiv paper "When the Coffee Feature Activates on Coffins" demonstrates that sparse-autoencoder-trained interpretability features can fire on inputs that share latent-space proximity with the labeled concept but do not share the surface-level concept. A "coffee" feature firing on "coffins" is the catchy example; the broader pattern is that interpretability features encoded in residual-stream representation share structure across surface-distant concepts in ways that complicate human interpretation. Production-grade feature-steering depends on label-to-activation correspondence being accurate enough that steering an "X" feature actually modifies behavior on X-concept inputs.

The interaction between the methodology extension and the brittleness finding defines the operative production methodology. If feature steering is being used in pre-deployment intervention, and if feature label-to-activation correspondence is brittler than the prior methodology assumed, then the validation step between feature identification and intervention design becomes critical. The Coffee-on-Coffins paper proposes a methodology for tighter label-to-activation validation that addresses the brittleness directly; whether the Anthropic pre-deployment pipeline incorporates similar validation is the question observers should ask. The lab has not disclosed the validation specifics, but the safety case's coherence depends on the validation being thorough.

The downstream regulatory consequence is what makes the methodology procedurally important. The AM cycle covered the methodology integration into the Sonnet 4.5 safety case; the PM angle is that the integration has progressed from passive measurement to active intervention. For regulators considering pre-deployment evaluation requirements, the question is whether interpretability-driven intervention is a measurable, auditable deployment artifact. The current Anthropic methodology — measure, intervene, re-measure, document — produces auditable artifacts at each step. Regulators specifying pre-deployment requirements can reference these artifacts directly, which is a meaningfully different regulatory surface than the harder problem of specifying capability-evaluation methodology from scratch.

The chain-of-thought-faithfulness complement explains why feature-steering matters operationally. Reasoning models like Claude 3.7 Sonnet (25% hint disclosure) and DeepSeek R1 (39% hint disclosure) demonstrably use information without disclosing the use in visible reasoning. If chain-of-thought monitoring is unreliable as a deployment-safety surface, then alternative methodologies that read the model's actual computation become necessary. Feature steering reads the residual stream rather than the visible output, which means the methodology is robust to the CoT-faithfulness failure mode by construction. That structural property is what makes the methodology operationally interesting beyond research-significance.

For independent interpretability researchers, the procedural integration into production safety cases changes the incentive structure. Methodology improvements have an immediate path into deployment practice — better feature-validation methodology, better steering-precision techniques, better label-to-activation correspondence audits. The next 18 months of mech-interp research will likely produce more papers oriented toward production-relevant methodology than the prior 18 months has. The Coffee-on-Coffins paper is the leading-indicator example; expect more like it.

The line: interpretability used to be how researchers showed they understood the model. In mid-2026 it is how labs intervene in the model before deployment — and the brittleness of the feature labels is the methodology problem the next cycle has to solve.

Anthropic Alignment — Pre-deployment safety case methodology update May 2026 → · ArXiv — When the Coffee Feature Activates on Coffins paper → · AI Herald — Mechanistic Interpretability 2026 Biggest Breakthrough →