Datasets:

---

caiovicentino1

/

wandering-arc-papers

like 0

I shipped a 6-paper arc (all CC-BY) + a pip-installable tool on a failure I call WANDERING: long-horizon coding agents that stay internally sure they solved the task but never emit the finish action, burning their whole budget — a ~34% blind spot for agent monitoring.

The finding. The decision to stop is decodable mid-stream (an interpretable SAE "task-done" feature predicts finish at AUROC 0.91) yet causally inert there. But it is writable late: patching a task-matched state into the last ~12 of 64 layers makes a stuck agent emit a real finish 42% of the time (p=0.031). The knowledge–action gap on agents is a layer gap — known mid-stream, writable late. Done solo in ~5 weeks on one Colab GPU.

Tool — decision-locator (find & steer the layer where any tool-calling decision is committed, on any open model, runs on a laptop):

pip install git+https://github.com/OpenInterpretability/decision-locator
decision-locator demo --model gpt2

Papers (Zenodo, CC-BY-4.0):

• #1 Tool-Entropy Collapse — https://doi.org/10.5281/zenodo.20368600
• #2 The Right Locus Is Still Not a Rescue Lever — https://doi.org/10.5281/zenodo.20490278
• #3 Multi-Channel Signatures — https://doi.org/10.5281/zenodo.20490284
• #4 Modality Matters — https://doi.org/10.5281/zenodo.20490286
• #5 The Verdict Is Not the Lever — https://doi.org/10.5281/zenodo.20532769
• #6 The Lever Is Late — https://doi.org/10.5281/zenodo.20534219
• companion (No Better Than Behavioral) — https://doi.org/10.5281/zenodo.20500053

All papers: https://openinterp.org/research · Tool: https://github.com/OpenInterpretability/decision-locator · This dataset mirrors the PDFs. Feedback welcome.