The Internal Monologue: v2.10.1 release, JSS plans, and a call for model-in-the-loop testing

wycoff

New member
A substantial update on The Internal Monologue and the Dual-Stream Architecture: the v2.10.1 software release is now on GitHub. I'd like people to download it, run it, and challenge the implementation with reproducible tests.

The release brings together compact evidence, portable deterministic verification, hybrid audit sampling, retention controls, and the recent verifier/security fixes. The goal is an auditable engineering pipeline linking generated answers to selected model evidence, with explicit limits on what that evidence establishes.

I've also decided to withdraw the paper from SCIRP's AIRR and pursue the Journal of Systems and Software (JSS) instead. That's my intended submission venue, not an acceptance announcement. I'm focusing the manuscript on the software architecture, verification, reproducibility, and the evidence the implementation can actually support.

For the paper, the v2.10 manuscript is available in the repository. The latest JSS revision work incorporates the completed H4-A/H4-B evidence and separates those results from the planned model-in-the-loop study. The current evidence and submission-material updates are in PR #41, which is still open as I write this. The paper version and the v2.10.1 software release are separate version labels.

Since that tagged release, H4-B has also been completed and merged into main. Its synthetic benchmark tests an attacker choosing event positions against a predictable public sampling schedule versus a protected keyed schedule. The results support resistance to that specific placement-evasion strategy when the audit key is unknown. They do not establish semantic deceptive-alignment detection, validate logits/probes as measures of intent, or demonstrate robustness to a model adapting its internal representations. Those are still empirical questions.

AlexH, following up on your offer to take a look: could you clone the repo and actually run the software, generate some evidence artifacts, and try the verifier? I'd really value feedback from using it, including anything that breaks or doesn't behave as documented.

I'd also like your help implementing the model-in-the-loop adaptive-evasion experiment. My proposed next step is to instrument a real model, define a measurable task and evasion objective, compare predictable and protected audit conditions, and test whether adaptation in token choice or internal representations changes what the monitoring captures. We should preregister the hypotheses, controls, and metrics, then report failures as carefully as successes.

Would you be willing to help with the model integration and experimental harness? Even starting with one model and a small, reproducible pilot would give us a concrete next step.

Repository and setup instructions | Tagged v2.10.1 release

Anyone else interested in running the software or contributing to the experiment is welcome to join in.
 
Back
Top