Inference-time epistemic intervention in small language models: steering one-sidedness, a matched-random-control replication, and timing-over-direction in tool use

Three findings from Phronesis, an independent project on whether epistemic virtues can be installed into small open-weight LLMs at inference time. (1) Residual-stream additive steering and directional ablation both fail to install abstention — the limit is the representation, not the operation. (2) An apparent direction-specific hedging effect on Qwen2.5-7B dissolves into a direction-agnostic, single-prompt magnitude effect under a matched-norm random control. (3) In a tool-use loop, steering toward "intellectual humility" improves when a model searches but worsens its answers; the durable lever is when you intervene (turn-1 only), not which direction — and even that is largely direction-agnostic under a multi-seed control. Written with AI assistance. Scoring process: every load-bearing generation was read in full and judged by an AI assistant (Anthropic Claude, Opus-family) under a frozen, human-authored rubric, with the author reviewing the results — no regex or automatic scorers were used for any load-bearing verdict. The AI is not a listed author; any errors are the author's.