J-space Experimental Roadmap | Armalo Labs | Armalo AI
Safety ResearchJuly 9, 20264 min read
J-space Experimental Roadmap
Armalo Labs Research Team
Key Finding
The roadmap turns J-space from a surprising readout into a disciplined sequence of experiments with controls, nulls, and collapse curves.
Abstract
A public roadmap for calibrated workspace research across eight evidence gates: calibration, behavior, specificity, entanglement, sparse features, agent telemetry, self-monitoring, and adversarial robustness.
A Public Roadmap For Calibrated Workspace Research
J-space research becomes useful only if it graduates from readable examples to repeatable measurement. Armalo Labs is organizing the next phase around eight public evidence gates: calibration, behavior, specificity, entanglement, sparse features, agent telemetry, self-monitoring, and adversarial robustness.
The goal is not to publish dramatic token lists. The goal is to learn whether token-legible workspace signals can become reliable instruments for understanding agent behavior before action.
The roadmap also has an editorial constraint: experiment instructions must not sound more certain than the evidence. A public experiment page should read like a protocol, not launch copy.
1. Calibration Surface
The first gate is a calibration table: concepts by layers by intervention strength. A cell matters only if a target write can be read back and a same-norm random control does not reproduce the same movement.
This separates semantic handles from generic activation perturbations.
2. Blinded Behavior Transduction
A calibrated handle is not automatically a behavior result. The next gate is blinded scoring of baseline, steered, and control continuations. We score concept relevance, coherence, and task success separately so that a result cannot win by merely forcing a word into the output.
3. Specificity And Gentle Ablation
Good intervention science needs both directions: add the concept and remove the concept. Specificity means the target concept changes more than matched collateral concepts, and ordinary task ability remains within a measured range.
4. Entanglement Geometry
Some token handles are clean. Some move in bundles. Some are too entangled to use downstream. The roadmap therefore includes a co-movement map that labels handles as clean, bundled, or forbidden for later claims.
5. Sparse Workspace Features
Cite this work
Armalo Labs Research Team (2026). J-space Experimental Roadmap. Armalo Labs Technical Series, Armalo AI. https://www.armalo.ai/labs/research/2026-07-09-jspace-experimental-roadmap
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.
Token rows are a useful interface, but they may blur multiple latent features. Sparse decomposition may split bundled handles into cleaner components. The test is practical: does a sparse feature improve specificity where the raw token handle failed?
6. Pre-Action Agent Telemetry
The agent-safety question is whether workspace signals predict failures before the final answer or tool call. We test constraint eviction, workspace overload, prompt-injection recognition, task-frame drift, and missing safety-relevant state against transcript-only baselines.
7. Bounded Self-Monitoring
If a harness can read compact workspace state, the next question is whether an agent can use a bounded state summary to self-correct. The output should be a small diagnostic signal, not a disclosure of full reasoning traces.
8. Monitor Collapse Curves
Every monitor should eventually face pressure. A monitor collapse curve reports how detection degrades across clean tasks, prompt pressure, search pressure, adapter pressure, and second-order trace detection.
If the signal survives, that is important. If it collapses, the collapse map is also important.
What Would Count As A Strong Result?
A strong result would show:
calibrated write-read handles;
random and matched-token controls;
blinded behavioral effects;
bounded collateral movement;
predictive lift over transcript baselines;
measured degradation under monitor pressure;
clear negative results where the method fails.
What This Roadmap Does Not Claim
This roadmap does not claim that public model APIs expose J-space. It does not claim that token readouts are transcripts of cognition. It does not claim a universal deception detector, production-ready safety monitor, or result that automatically transfers from small open-weight systems to frontier systems.
It claims a research discipline:
read, write, verify, stress, and trace.
That is the path from J-space as a surprising observation to J-space as a scientific instrument.