The Lever Generalizes — and It Brakes

A late, bidirectional action-commitment lever across agent decisions and architectures

Caio Vicentino · OpenInterpretability · Published 2026-06-10. Zenodo · CC-BY-4.0 · DOI 10.5281/zenodo.20634838 · the circuit-breaker capstone of the WANDERING arc.

Paper #7, the circuit-breaker capstone. The full PDF (4 figures, the scale-matched cross-model table, the pre-mint eval, all caveats) is the Zenodo record — this page is the on-site summary. Prepared for submission to BlackboxNLP 2026 (EMNLP).

---

The question

Paper #6 (The Lever Is Late) showed that control of a coding agent's finish decision lives not at the mid-layer "task-is-done" verdict but in a late, task-matched action-commitment block ~30 layers downstream. That left two questions open. Is the late lever specific to termination, or a property of the action-commitment channel? And can it brake an action — not just elicit one — the mechanism a safety circuit-breaker would need?

What we found

On Qwen3.6-27B over 99 SWE-bench Pro trajectories, using a second decision in the same data — commit a file edit (str_replace_editor) versus continue reversible exploration (bash) — with n = 60 deterministic decision points per condition, prefill-only patching, and generation-confirmed outcomes:

1. It generalizes (elicit). At an explore decision, injecting a task-matched edit-donor into the late block makes the agent emit a real edit call: 0.23 → 0.77 at L59 (position control 0.08, cross-task control 0.48).
2. It brakes (suppress). At a commit decision, injecting an explore-donor collapses the real edit rate 0.48 → 0.02 (96% suppression) at L55 — a same-class control stays intact (0.55) and the opposite-direction donor boosts to 0.92.
3. The mechanism is monotonic and bidirectional. Exact paired McNemar on all 14 per-point conditions yields seven contrasts surviving Holm–Bonferroni (worst p = 7.6e-5). The discordance structure is the headline: elicit has c = 0 (the edit-donor only turns commits on, never silences one) and the brake has b = 0 (the explore-donor only turns them off, never lights one). The late lever moves exactly in the donor's direction with ~zero off-direction noise.
4. It is architectural, not a Qwen artifact. The late-commitment geometry and donor-specific writability replicate across two model families and two scales (Mistral-7B and the scale-matched Mistral-Small-24B, where the mid-inert / late-write dissociation is cleanest: fidelity 0.955 vs 0.007).

The circuit-breaker

The bidirectional late lever is the mechanism a safety circuit-breaker requires: a single late-layer intervention that blocks an action at its commit point. The brake re-routes the suppressed decision to reversible exploration (bash +0.17 above its no-brake floor of 0.43), and the elicit/brake lift survives a full valid-tool-call re-parse (0.23→0.37 elicit, 0.40→0.07 brake) — not just the tool-token onset. The method ships in decision-locator, the model-agnostic locate/sweep/steer tool from paper #6.

Honest scope

Generalization here is an existence proof, not a universal law — it holds for finish and one second action, on one model and one task family, plus a cross-family geometry/writability replication. The proxy is semi-irreversible: str_replace_editor mutates state but is undoable. The genuinely irreversible case (send_transaction, delete) is untested and is the named next step — the circuit-breaker is demonstrated on a proxy. The cross-model lite tests use LOCATE + one-token ΔP, not generation-confirmed brakes.

Code & data

• Paper + pre-registrations + per-point data + exact-statistics script + the pre-mint eval: paper/circuit_breaker and tools/decision_locator
• Decision-point state + residuals + strengthener data: caiovicentino1/swebench-phase6-verdict-circuit
• The full arc: #1 Tool-Entropy · #2 Right Locus · #3 Multi-Channel · #4 Modality Matters · #5 Verdict Circuit · #6 The Lever Is Late · companion note