The Verdict Is Not the Lever

An interpretable task-completion feature predicts but does not cause long-horizon agent termination

Caio Vicentino · OpenInterpretability · Published 2026-06-03. Zenodo · CC-BY-4.0 · DOI 10.5281/zenodo.20532769 · the mechanistic capstone of the WANDERING arc.

Paper #5, the capstone of the five-piece WANDERING arc. The full PDF (figures, both tables, all caveats) is the Zenodo record — this page is the on-site summary.

---

The question

A long-horizon coding agent fails by WANDERING — acting indefinitely without ever emitting the finish action. The arc detects this behaviorally, shows residual steering can't rescue it, and shows a behavioral interruption can. The deepest mechanistic version: does the agent internally represent a "the task is done" verdict, and if so, is that representation the causal lever for termination?

What we found (four stages, CPU-only except the causal test)

Using a full-stack sparse autoencoder on Qwen3.6-27B over 99 SWE-bench Pro trajectories:

1. The verdict feature exists. An SAE feature at layer 23 (#22358), selected anti-circularly from SUCCESS only, is present at the WANDERING final turn but not at LOCKED (WANDERING-vs-LOCKED AUROC 0.81, length-controlled, partial-r 0.55). WANDERING — which never terminates — holds the verdict; LOCKED — which gave up — does not.
2. It is interpretable. Reading the transcripts where it fires hardest: "All 157 tests pass", "the implementation is complete and tested", "the refactoring is complete". It is a "subtask completed and verified" feature (completion-language enrichment 50% vs 6% baseline).
3. It predicts the action. Against ground-truth tool calls, the feature predicts whether the next action is finish with AUROC 0.91.
4. It is not the lever. Clamping it to the SUCCESS level at the WANDERING decision point changes P(finish) by −0.001 — indistinguishable from clamping a random feature (+0.002). Ablating it in SUCCESS does not significantly reduce finishing (−0.008, n.s.).

Why it matters

The agent represents "I'm done" as a clean, interpretable, predictive feature — but that representation is not what makes it stop. This extends the detection-vs-control asymmetry (the knowledge–action gap, the predict/control discrepancy) to the level of a single, named, interpretable feature representing a specific decision. It sharpens the arc's three residual nulls one level deeper: not "we steered the wrong direction," but "the right, named, interpretable verdict feature is still not a lever." WANDERING does not fail to form the verdict — it forms it repeatedly — it fails to treat it as terminal. The only known WANDERING rescue remains a behavioral interruption: the predictive signal lives in the representation, the causal lever lives in behavior.

Honest scope

Single model (n=99, Qwen3.6-27B), single task family (SWE-bench Pro). Stage 0 is uncorrected/single-layer-clean (the claim rests on convergence across stages, not one p-value). The in-domain interpretation is decisive (a strict general-corpus semantic gate failed — the feature is domain-specific). The causal test is the load-bearing negative and is modest in power: n=10/class, a single layer and a single token position, behavioral-fidelity-gated, prompts truncated to 4000 tokens. It rules out L23 #22358 as a single-position lever; it does not prove no verdict locus anywhere is a lever.

Code & data

• Paper + pre-registrations + all-stage results + the causal-test notebook (CC-BY-4.0 / Apache-2.0): paper/verdict_circuit
• Transcripts + labels + fidelity residuals: caiovicentino1/swebench-phase6-verdict-circuit
• The full arc: #1 Tool-Entropy · #2 Right Locus · #3 Multi-Channel · #4 Modality Matters · companion note