Paper v0.8: Tool-Entropy Collapse (Cross-Architecture Signature of Agent WANDERING Failure)

Tool-Entropy Collapse: A Cross-Architecture Signature of Agent WANDERING Failure

📄 Zenodo (v0.9 calibration revision): 10.5281/zenodo.20368807
🔗 Concept DOI (always latest): 10.5281/zenodo.20368600
📌 Original v0.8: 10.5281/zenodo.20368601
💬 LessWrong post: https://www.lesswrong.com/posts/D23kWLmpcXhFAQ2RY/tool-entropy-collapse-a-cross-architecture-signature-of

PDF + LaTeX source + 5 figures.

v0.9 calibration changes vs v0.8: removed 'breakthrough' language throughout in favor of 'most promising candidate signal in this work'. Added explicit hedge in conclusion that W/S ≈ 0.41 ratio match between Qwen and Llama is the most suggestive pattern that merits independent replication before being treated as a discovery. Same empirical results, same scope, same numbers.

Headline numbers:

• 34% WANDERING blind spot in probe-only monitoring (95% CI [22.0%, 45.8%])
• 6 detector designs across 3 signal channels tested
• Tier 3 detector (v1 ∪ v5): 70% recall × 5% FP on Qwen primary dataset (N=20 WANDERING)
• Cross-architecture suggestive pattern: W/S median entropy ratio ≈ 0.41 in Qwen AND Llama, 0.71 in GPT-5

Honest scope:

• Cross-task validation on METR MALT (15+ task families) is NULL
• Scoped claim: multi-turn code-execution agent tasks with rich action spaces
• W/S 0.41 ratio match merits independent replication on additional models

Citation:

Vicentino, C. (2026). Tool-Entropy Collapse: A Cross-Architecture Signature of
Agent WANDERING Failure (v0.9 calibration revision). Zenodo.
https://doi.org/10.5281/zenodo.20368807

Reproducibility: all scripts in this repo. Per-trajectory output JSONs in scripts/inflection_turn_out/. Apache-2.0.

v0.5 — paper-5 saturation-direction lever shipped

Paper-5 published 2026-05-09

Saturation-Direction Lever: A Five-Class Taxonomy of Probe Causality in Qwen3.6-27B
→ https://openinterp.org/research/papers/saturation-direction-probe-levers

Headline finding

α=−100 robustness theorem — the L31 pre_tool capability probe direction produces +33-40pp probe-vs-random pushdown gap across code distributions spanning Qwen3.6-27B pass-rate ~7-89% (HumanEval+MBPP, BigCodeBench, Codeforces ≥2000). The α=−100 locus is saturation-independent at moderate amplitude.

Phase coverage (this release)

• Phase 0 — Smoke G1+G4 PASS
• Phase 1 — N=20 stratified, 12k captures, G4 GREEN
• Phase 2 — Differential probes, AUROC up to 0.958
• Phase 5d/6/6c — N=99 + methodology sweep
• Phase 7 — L43 pre_tool epiphenomenal (softmax-temp artifact)
• Phase 8 — L55 thinking template-locked
• Phase 10 — RG L55 mid_think first lever (+30pp pushup at α=+200)
• Phase 11/11b — 4/4 capability sites pushdown-asymmetric (+30 to +60pp)
• Phase 12 — Persona falsifier #1 → motivates saturation-direction theory
• Phase 11c — Cross-distribution BCB +33pp at α=−100
• Phase 11d — Codeforces ≥2000 +40pp at α=−100; falsifier #2 walks back saturation-magnitude corollary
• Phase 11e — Multi-site Codeforces (in progress at release)

Methodology contributions (mandatory at every site)

1. Random K-matched probe baseline (paper-3 §3.1)
2. Control-token normalization for log-prob shifts (paper-3 §3.2)
3. Structural-rigidity α-sweep at α >> ‖h‖ (paper-3 §3.3)
4. Whitespace-stripped behavioral flip metric (paper-3 §3.4)

3 of the 4 caught a confident-but-wrong claim during the work.

Two pre-registered falsification cycles

• #1 Phase 12 persona — predicted pushup-asymmetric (continuous-gradient class), observed pushdown → falsifies categorical-vs-continuous frame, motivates saturation-direction principle
• #2 Phase 11d Codeforces — predicted pushdown collapse + pushup emergence at lowest saturation (saturation-magnitude corollary), observed saturation-INDEPENDENT pushdown → walks back corollary, refines to α=−100 robustness theorem

Companion artifacts

• HF dataset — caiovicentino1/agent-probe-guard-qwen36-27b (probe weights for v0.1 SDK; v0.2 will swap L43 → L31 pre_tool from Phase 11)
• PyPI — pip install openinterp (v0.3.0+)
• Web — paper-5 + 4 prior papers at openinterp.org/research
• License — Apache-2.0 throughout

Reproducibility

12 standalone Colab notebooks in notebooks/. Build scripts in scripts/build_nb_swebench_v*.py regenerate every notebook from source. ~6.5h total compute on RTX 6000 Blackwell from cold start.