Repository navigation
Publication
The current publication is the author-approved ten-page revision 14 (PDF, Word source). It replaces the previous generated manuscript. The evidence is still best framed as an exploratory workshop or short-paper case study, not a strong main-track conference claim. This is an assessment, not an acceptance prediction.
- A clearly specified classifier-plus-vector-plus-guard pipeline.
- A training-only search followed by recorded evaluation on two splits.
- Source-pinned observations, independent audits and a compact GPU configuration report.
- Honest accounting of abstention, answer-order dependence and the distinction between score checks and behavioral reliability.
Only four evaluation scenarios improve, all SELF cases. Zero wrong-way flips partly follow from the guard's definition. Without a probe-budget-matched search baseline, random-direction control and ablations, the result cannot isolate the learned vector's added value. There is no fresh independent confirmation, broad model comparison, live feature-extraction benchmark, or free-text/action evaluation.
- Compare with matched random directions and equally budgeted output/probe selection.
- Test on newly authored, independently labeled scenarios with policy settings frozen beforehand.
- Ablate the classifier gate and each guard condition; report coverage and probing cost.
- Evaluate real generated answers or task outcomes, not only conditional A/B preferences.
Venue choice should follow the narrow empirical contribution. Before submission, choose the exact template/page limit, confirm authors and affiliations, and check venue-specific policies. The generic PDF is not a venue-formatted camera-ready submission. The paper has not been submitted or accepted.