What question did this study set out to answer?

Determine whether a frozen, linear white-box detector can reliably monitor persistent manipulation signals across diverse open-weight language model architectures without retraining.

August 15, 2026Open Access

Nucleation Pilot & Related Projects: A Frozen White-Box Detector for Persistent Manipulation, Its Cross-Family Transfer, and Preregistered Methods

View Full Paper

Ask AI

Bookmark

Authors

CHChristopher Blake HeadThe Navigator (Portugal)

Discussion

Member takes

Implication

Experimental evaluation demonstrates cross-family transfer of a frozen residual stream detector across six open-weight language models, indicating generalizable white-box manipulation monitoring.

Key Points

Determine whether a frozen, linear white-box detector can reliably monitor persistent manipulation signals across diverse open-weight language model architectures without retraining.
Tested a frozen, SHA-256-hashed linear detector initially validated on a benchmark model with a learned refusal boundary against six open-weight model families (Qwen2.5-1.5B/7B, Phi-3.5-mini, SmolLM2-1.7B, Llama-3.2-3B, OLMo-2-7B).
Applied read-mask ablations, naturalistic in-context frame testing across 24 family-by-variant cells, and owned-model refusal erosion attempts using commit-before-run preregistrations.
The detector achieved cross-family transfer across all six open-weight families with graded effect sizes between 3.69 and 8.03 (all 95% CIs clear of zero), following an initial held-out validation AUC of 0.807.
In-context evaluation revealed persistent residual manipulation traces in 24/24 family-by-variant cells (predominantly source-decoupled in 19/24), while owned-model refusal erosion produced consistent null findings.

Cite This Study

Christopher Blake Head (2026) studied this question.

synapsesocial.com/papers/6a80197375c2e31742c8570e https://doi.org/10.5281/zenodo.21923976

View Full Paper

Ask AI

Bookmark

Implication

Experimental evaluation demonstrates cross-family transfer of a frozen residual stream detector across six open-weight language models, indicating generalizable white-box manipulation monitoring.

Nucleation Pilot & Related Projects: A Frozen White-Box Detector for Persistent Manipulation, Its Cross-Family Transfer, and Preregistered Methods

Discussion

Member takes

Implication

Key Points

Cite This Study

Nucleation Pilot & Related Projects: A Frozen White-Box Detector for Persistent Manipulation, Its Cross-Family Transfer, and Preregistered Methods

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider