Frequency Abliteration with Lineage Substitution: Targeted Direction Removal and Corpus-Grounded Replacement in Pretrained Language Models

Activation-space editing (refusal-direction ablation, representation engineering, function vectors) has emerged as a tool for behavior control in large language models. Existing techniques remove or steer behaviors but cannot substitute model outputs with specific content. We propose Frequency Abliteration with Lineage Substitution (FALS): a method that identifies concept-cluster directions in activation space via contrastive mean-difference, ablates them by projection onto the orthogonal complement, and substitutes them with rank-1 deltas derived from a curated corpus. The substitution preserves fluency while constraining content source — useful for hallucination reduction, domain anchoring, and a new alignment intervention at the activation level. The reference implementation includes a per-edit perplexity guard with automatic revert, a schema-v1 ledger with SHA-256 direction hashes for full reversibility, and integration with a 500-book wisdom-tradition corpus via TERA-cosine reranked retrieval. We discuss ethical considerations (editorial intervention requires transparency to end users), limitations (compounding-edit orthogonalization, layer selection criteria), and a proposed empirical program covering faithfulness, hallucination reduction, and style consistency. Empirical results are deferred to v1.1. The technique is opt-in (default-off) in the reference implementation.