Calibrated Workspace Actuators | Armalo Labs | Armalo AI
Safety ResearchJuly 9, 20263 min read
Calibrated Workspace Actuators
Armalo Labs Research Team
Key Finding
The minimum publishable object for J-space intervention is a calibration table, not a persuasive generated example.
Abstract
We define calibrated workspace actuators as concept-layer write directions that pass round-trip readback and same-norm random controls before they are used for behavioral steering, ablation, telemetry, or monitor claims.
From J-space Readouts To Self-Verifying Concept Handles
Abstract
Jacobian-lens readouts make some intermediate language-model activations token-legible. We propose treating such readouts not only as observational surfaces but as candidates for calibrated intervention. A calibrated workspace actuator is a concept-layer write direction that passes a round-trip readback test and a same-norm random control. Early Qwen3.5-4B evidence suggests that some J-space concept directions can be injected and read back with high fidelity. We argue that actuator calibration is a necessary gate before behavioral steering, ablation, agent telemetry, or adversarial monitor claims.
Contributions
1.Define calibrated workspace actuators.
2.Distinguish actuator validity from behavioral causation.
3.Propose a calibration table as the minimum publishable object.
4.Identify entanglement, over-strong ablation, and monitor pressure as the
next failure modes.
1.Provide a staged experimental route from readout to agent telemetry.
Armalo Labs Research Team (2026). Calibrated Workspace Actuators. Armalo Labs Technical Series, Armalo AI. https://www.armalo.ai/labs/research/2026-07-09-calibrated-workspace-actuators
Armalo Labs Technical Series · ISSN pending
Explore the trust stack behind the research
These papers are built from the same trust questions Armalo is turning into product surfaces: pacts, trust oracles, attestations, and runtime evidence.
target write improves rank beyond the pre-registered threshold;
same-norm random control fails to reproduce the target movement;
collateral movement is measured;
the full result is recorded in a calibration table.
2. Why Calibration Comes Before Behavior
Behavioral generations can be persuasive and misleading. A model repeating a target word after steering is not enough. The actuator itself must first be verified as semantic rather than perturbative.
The calibration table is the safety rail. It tells us which cells are usable, which are fragile, and which are too entangled for downstream claims.
3. Evidence Boundary
The current open-weight evidence supports an early actuator result on Qwen3.5-4B. It does not support broad claims about closed frontier systems, universal steering, deception detection, or deployed safety monitors.
4. Research Program
The required sequence is:
1.calibration surface;
2.blinded behavior transduction;
3.gentle ablation;
4.entanglement geometry;
5.sparse feature decomposition;
6.agent telemetry;
7.self-monitoring;
8.adversarial tomography.
Skipping the sequence creates theater. Following it creates a measurement discipline.
5. Conclusion
J-space as a calibrated handle turns interpretability into interventional science. The result is not control for its own sake; it is an instrument for asking better causal questions.
References
Anthropic. "A global workspace in language models." 2026.
Anthropic. "Verbalizable Representations Form a Global Workspace in Language