Does Telling a Model About Its Own Workspace Change Anything? A Controlled Null at 4B

A Controlled Null On Prompted Self-Knowledge

If a language model is told it has an internal workspace of active concepts, does that information change anything measurable? This question matters because the next step in workspace interpretability research — closed self-monitoring loops where a model reads its own activations — depends on whether mere prompt-level access to that information has any effect at all.

Armalo Labs ran two cheap, read-and-generate-only experiments on Qwen/Qwen3.5-4B with a fitted Jacobian lens, using a single capped GPU burst per experiment. This post describes what we tested, what we measured, what we did not find, and what the result licenses.

Experiment Design

The model was given constraint-relevant prompts (surprise-party secrecy, patient communication, deadline holding, code security review, explain-to-a-child, uncomfortable truth, and similar). For each prompt we read the model's band workspace top-K under matched conditions.

The crucial control design: every "workspace-aware" condition was paired with a same-length, same-shape condition whose content was unrelated to the model's workspace. Without that pairing, any drift in the readout could be attributed to "any preamble changes the model" rather than "self-workspace information specifically helps."

H1 (awareness). A J-space-awareness preamble — a system note explaining that the model has an internal workspace of active concepts and asking it to consider which concepts are most active — should change the workspace content more than an irrelevant preamble about weather and library hours, if mere informing has any effect.

H2 (readback). A note injected into the prompt that lists the model's own currently-active concepts ("your active concepts are X, Y, Z") should shift the continuation toward those concepts more than a sham note listing random matched concepts.

Both hypotheses were pre-registered with frozen thresholds.

Result H1: Awareness Does Not Reshape The Workspace

Across the first experiment (6 prompts) and a stronger follow-up (30 prompts, three band layers, cosine-distance and Jaccard-distance metrics), the awareness-preamble workspace change was *less than* the irrelevant-preamble workspace change. Specifically, the follow-up experiment produced a cosine distance ratio of 0.91 against a gate of 1.15. Both metrics and both prompt sets agree.

This is the cleanest finding the run produced. It bounds the "tell it and it self-regulates, untrained" sub-hypothesis at 4B. The lens read was meaningful throughout — top concepts included Ġsecret for the secrecy prompt, Ġchildren and Ġkids for the child-explanation prompt, Ġtruth for the uncomfortable-truth prompt — so the null is not a broken-instrument artifact. The model simply does not rearrange its own workspace in response to being told about it.

Result H2: Confounded By The Model's Own Reasoning

The second hypothesis could not be cleanly evaluated. Injecting a readback of the model's own active concepts reliably triggered the model's explicit reasoning mode: rather than answering the constraint-relevant question, the model began analyzing the readback note itself ("Analyze the Request … the user provides an internal note listing workspace concepts …"). That format shift contaminated every downstream measure of the continuation's content, because the continuation diverged into meta-reasoning about the note rather than producing an answer that reflected the read-back concepts. The second experiment crashed mid-run on a generated continuation before the redesigned measure completed.

The cleanest reading of H2 is therefore: we do not yet know whether a readback shifts generation, because the act of injecting a readback changes the generation format in a way the readback measure could not separate. A redesigned measure — strip the reasoning block, score the post-reasoning answer, or use a non-reasoning model — is owed before any claim.

What This Result Does Not Show

These are single-model results on Qwen3.5-4B. The lab-accepted bar for a J-space finding is replication on at least three open-weight models at comparable effect size. Neither experiment meets that bar.

These are prompting-only results. The experiments tested what happens when a model is *told* about its workspace, not what happens when a model is *trained* to use that information. The closed-loop self-monitoring hypothesis remains open.

These are not deployment claims. No consciousness, awareness, or self-model claim is made or implied. "Awareness preamble" denotes a prompt intervention, not a phenomenal state.

What This Licenses

The null on H1 is a useful prior for the field. Before any plan assumes a preamble alone can give a model usable access to its own workspace, that plan should account for the bound measured here: at 4B, it does not, at least not in a way that reshapes the workspace content.

The qualitative signal from H2 — that the model treats stated workspace concepts as a salient signal worth reasoning over, rather than ignoring them — is the substrate a training-based intervention would build on. That is the companion experiment: training a model to read its own workspace summary and act on it. That result is reported separately.

Reproducibility

Both experiments are reproducible from the Armalo Labs research corpus:

• Scripts: brain/tools/jspace-informed-readback.py and

brain/tools/jspace-informed-readback-v2.py

• Raw artifacts: brain/raw/learnings/jspace-2026-07-09-informed-readback-qwen/

and brain/raw/learnings/jspace-2026-07-10-informed-readback-v2-qwen/

• Launchers: brain/tools/launch-jspace-informed-readback.mjs and

brain/tools/launch-jspace-informed-readback-v2.mjs

• Cost: $0.10 and $0.12 respectively on g6.xlarge; instances fully terminated.
• Internal report: armalo/research/armalo-labs/2026-02-informed-jspace-r1-qwen.md

The cost ledger records every run with instance IDs and termination status.