No Better Than Behavioral

A Residual Velocity-Freezing Fingerprint Predicts Agent WANDERING No Better Than the Cheap Tool-Entropy Detector

Caio Vicentino · OpenInterpretability · Published 2026-06-01. Zenodo · CC-BY-4.0 · DOI 10.5281/zenodo.20500053 · companion note to the WANDERING arc.

A pre-registered negative attached to the four-paper WANDERING arc. The full note (with both result tables and the pre-registration) is the Zenodo record (pre-registration + code on GitHub) — this page is the on-site summary.

---

The question

The WANDERING arc detects a long-horizon agent failure — an agent that keeps acting but never finalizes — from a probe-free behavioral signal (tool-call entropy collapse), and shows that residual steering cannot rescue it while a behavioral interruption can. A skeptic's next question, sharpened by the context-rot literature (long-context degradation is representational, not retrieval — arXiv:2510.05381): does the residual stream carry an earlier or better detector than the cheap behavioral one? Does the geometry rot before the behavior does? If yes, activation capture earns its cost. If no, the cheap signal is not just convenient — it's sufficient. We pre-registered the test before looking.

What we found

Stage 1 — a real but weak fingerprint exists. On the same 99 Qwen3.6-27B SWE-bench Pro trajectories, raw residual geometry (no SAE) surfaces one signal: representational velocity-freezing. Trajectories that will WANDER show a smaller per-turn change in their residual state early on — they settle toward an attractor sooner. The direction is consistent across all five layers (4/5 raw p<0.05, length-controlled), and one mid-network layer (L31) clears a pre-registered trend-and-divergence conjunction (p=0.015). But nothing survives multiple-comparison correction over the 4×5 metric–layer grid.

Stage 2 — the fingerprint adds nothing over the cheap detector. This is the decisive, pre-registered test:

• Early-window velocity at L31 reaches AUROC 0.695 — statistically indistinguishable from the fair early behavioral baseline tool_entropy_first10 (0.688; paired bootstrap Δ=+0.008, 95% CI [−0.170, +0.211], straddles zero).
• It loses to the deployed late detector tool_entropy_last10 (0.888) and to trajectory length (1.000, postdictive).
• As a sharp alarm calibrated to ≤5% false-positive on successes, velocity catches only 1–3 of 20 WANDERING trajectories — far fewer than the deployed detector, with too few overlapping detections to even measure a lead-time advantage.

Why it matters

The pre-registered gate returns NO-GO on three independent grounds, so we stop — we do not spend SAE compute or causal compute on a signal with no operational value. The contribution is the negative itself: context rot does leave a faint, directionally-coherent residual trace, but at the granularity a cheap deployable monitor would use, that trace carries no predictive information beyond the probe-free behavioral signal it would replace. The residual geometry is downstream-redundant.

This strengthens the arc rather than undermining it: across four papers the deployed signal has been the cheap behavioral one, and here we show that choice is not merely convenient — for this failure mode, watching the behavior is as good as or better than reading the residual stream. "Just watch the geometry" is ruled out as a better detector.

Honest scope

Single model (n=99, Qwen3.6-27B), single task family (SWE-bench Pro), raw geometry only (a richer SAE-feature or attention-level representation was gated off by the predictive null, by design). And this is a statement about prediction and redundancy, not causation — a predictive-redundancy null does not exclude that the frozen geometry lies on the causal path of the behavioral rescue; that would be a separate, re-pre-registered question.

Code & data

• Note PDF + pre-registration + Stage 1/2 results + analysis code (CC-BY-4.0 / Apache-2.0): paper/context_rot
• Arc PDFs mirror (HF dataset): caiovicentino1/wandering-arc-papers
• The four arc papers: #1 Tool-Entropy · #2 Right Locus · #3 Multi-Channel · #4 Modality Matters