Safe Error Correction for Language Models: Frozen-Base Adjustment with Capability Preservation

TEXT START: We study a practical question: can a small correction module fix errors in a frozen language model's outputs without degrading its base capabilities?

The Dissection

This is a containment instrument for an already capable machine. CRN v2 leaves the 4.65B-parameter base frozen and attaches a logit-level correction layer trained on 83,400 pairs. Its real product is deployability: repair part of the error surface without destabilizing the incumbent model.

The reported result is bounded but meaningful: 53.3% of errors corrected, 43.3% after rewording, with no measured loss on narrow capability checks. The LoRA comparison exposes the tradeoff directly: stronger correction can destroy existing competence. This is an engineering study of modular patching, not a demonstration of general safety or architectural transformation.

The Core Fallacy

The central Discontinuity Thesis error is treating “capability preservation” as if it preserved human economic value. It preserves machine output capacity. It does not preserve human productive participation.

The module makes automation more reliable and cheaper to deploy. It converts model failure from a reason to retain human labor into a maintenance problem solvable with a small add-on. That reinforces P1 rather than weakening it.

The empirical overreach is equally clear: “no degradation” means no degradation on MMLU/BoolQ with N=200 and a car-wash test with N=8—not preservation across real deployment conditions. The paper does not claim to solve the systemic problem; the mistake is assigning such a local patch systemic significance.

Hidden Assumptions

• The 60-question CEHRI exam represents the error distribution that matters in deployment.
• Reworded questions adequately test robustness to distribution shift.
• Reference-free DPO improves correctness rather than introducing stylistic conformity, hidden regressions, or reward artifacts.
• Benchmark stability predicts performance under adversarial prompts, long contexts, multilingual use, tool calls, and unfamiliar domains.
• A 0.73% trainable-parameter fraction translates into low total operational cost, despite retaining a 4.65B frozen base.
• The frozen base and its correction layer will remain aligned as tasks, prompts, and model distributions change.
• Correcting logits is sufficient for errors rooted in missing knowledge, defective reasoning, or latent representation failures.
• The released weights and scripts establish reproducibility even though the deep variant has no trained checkpoint.
• Preserving current capability is the correct safety objective, rather than calibrated abstention, uncertainty, or verifiable reasoning.

Social Function

Primary classification: transition management with partial truth and safety prestige.

The result gives operators a reassuring deployment narrative: existing models can be patched without retraining or sacrificing their useful behavior. That is not pure copium; the narrow result may be real. But “safe” is doing more rhetorical work than the evidence supports. The paper helps move institutions from unreliable AI to more reliable AI while leaving the labor displacement mechanism untouched.

The Verdict

A competent patch, strategically bearish for human labor. It demonstrates that cognitive automation can acquire a maintenance layer without surrendering its existing machine capabilities. Under the Discontinuity Thesis, this is not a reprieve. It is a sharper blade.

The honest conclusion is narrow: one frozen model, one small correction corpus, a 60-question error test, and limited preservation checks produce roughly half-error correction. The result advances P1 and offers no evidence against P2 or P3. The machine gets repair infrastructure; humans do not get their economic necessity back.