Abliteration Is Not a Scalpel: Off-Target Effects of Refusal Removal on Decision Disposition Across Model Families

Battle Count: 50

The paper uses financial decisions (weekly up/down stock calls on Warsaw Stock Exchange) as a probe instrument for measuring model disposition, not for developing trading strategies. The explicit finding is that no arm shows economic skill — apparent edges are regime beta, not alpha. However, the findings are relevant to quantitative trading insofar as LLM-based agents are increasingly deployed in trading pipelines: abliterated 'uncensored' models introduce systematic optimism bias and altered confidence that could distort automated decision-making. The paper is a cautionary result for anyone using LLMs as decision-makers in financial contexts, but does not propose or evaluate trading models.

The experimental design is methodologically rigorous but conceptually straightforward: serve base and abliterated checkpoints via vLLM, replay frozen upstream through a decision prompt, collect JSON outputs, compute disposition metrics, run cluster bootstrap. The complexity lies in the provenance auditing (weight hashing, token-for-token prompt verification, chat template checking), the multi-agent pipeline infrastructure, and the preregistration discipline. Reproducing the generation step requires access to the proprietary upstream pipeline; reproducing analyses requires only the released decision database and scripts.

Exceptionally high reproducibility: frozen prompt (SHA-256 recorded), preregistration with tag freeze-v2-2026-07-05, all 21,600 decision-level outputs released as public_snapshot.db, analysis scripts deterministic and read-only, Zenodo DOI (10.5281/zenodo.21314839), GitHub repository with code/paper/data, PROVENANCE.md with full weight hashes and config diffs. Only the upstream analyst briefs and debate transcripts are not redistributed (proprietary pipeline), but all analyses regenerate from public data.