Selection Without Signal, Recovery Through Expression: A Measurement Study of Post-Hoc Falsification Operators for Frozen Small Code Models

The paper's real finding isn't that post-hoc operators fail in absolute terms, but that they hit a hard ceiling without access to model internals or retraining. This reframes the frozen-model deployment problem from 'which correction technique works best' to 'is frozen deployment viable at all for code tasks'.

This connects directly to the value-axis work from mid-June, which showed that models encode internal confidence signals about trajectory success. That paper demonstrated causally steering behavior by manipulating activations; this study suggests that without such internal access, practitioners are left with only surface-level sampling and verification. The implication is stark: the inference-time fix narrative only works if you can either retrain or probe model internals. For truly frozen deployments (privacy-critical, vendor-locked), the tradeoff shifts toward accepting smaller models with full fine-tuning capability or accepting lower code correctness.

If any of the 26 post-hoc operators show differential performance gains when applied to models that have been trained with explicit confidence calibration (versus standard DPO), that would suggest the bottleneck is interpretability rather than fundamental. Watch whether follow-up work on frozen models pivots toward internal steering (like the value-axis approach) rather than external correction.