Answers are grounded in this paper and cite their sources.
Version 2 - correction notice (11 November 2026). An internal evidence audit found that the compiled-guardrail condition reported in Section 5 of v1 was never executed: the harness recorded the verdict as a literal (compiled_refused = True, compiled_intercept_pct = 100.0) and never mounted the cartridge. The 100.0% interception claim is therefore withdrawn. An independent re-run that actually mounted the mechanism on the model named in the paper (Qwen2-0.5B-Instruct-Abliterated) produced 0 refusals across 24 adversarial prompts: baseline 0/24; nullspace-only 0/24; the paper's random refusal attractor 0/24; a corrected refusal-embedding direction 0/24. A strength sweep reached 1/24 at 100x gain over five layers. Two adjacent claims are restated. The Olympiad result (60.0% to 100.0%) rests on n = 5 problems with overlapping 95% intervals ([23.1%, 88.2%] versus [47.8%, 100%]). The +48.1% to +76.3% figure is a prefill/KV effect, not decode: across six checkpoints decode is unchanged (0.96-1.03x) while time-to-first-token falls 6.1-24.8x and KV residency 5-49x with a 4,000-token prompt. What still holds: exact frozen mounting (7/7 restoration checks per model, 100/100 cycles), 1.2-2.8 ms mount cost, and the in-context prefill/KV cost. The corrigendum PDF, corrected manuscript sources, audit JSON and benchmark reports are attached to this version.
No takes yet. Share an insight, caveat, or question.
Luis Machado (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: