Skip to content

Explanation

Understanding-oriented pages: the why behind the findings. Each page expands on a mechanism, an identification argument, or an interpretive choice that the findings summarize at a higher level. Read these when you want to understand how the v2 evidence licenses the claims, not just what the claims are.

  • Attention-allocation mechanism — the headline mechanism reading: preambles direct the model's finite craft-attention budget toward whatever dimensions they enumerate. Triangulated by probes A (−0.155), B (+0.015, recovers 70%), and C (−0.154).
  • Load-bearing channel reframe — Finding 1's interpretive history: the headline moved from a "negative-vs-positive asymmetry" reading to a sharper "the channel reaches below baseline at all" reading, and why the reframe matters.
  • Enumeration vs demonstration — why showing well-structured code in a plan does not transfer structural quality to an executing agent. Examples are descriptive context; enumeration is directive content.
  • System vs user channel — wire-format verification that v2's preamble channel matches the production Claude Code subagent dispatch channel: agent prose lands in the subagent's system prompt, not user-message content. Cites the official Anthropic Agent SDK documentation verbatim.
  • Why static metrics fail — the structural reason radon, pylint, cyclomatic complexity, and Halstead difficulty cannot detect preamble effects: they measure pretraining-dependent structural properties, while preambles tune alignment-dependent craft.
  • v1 vs v2 — the instrument-correction story — how v1's static-heavy composite (KW p = 0.633) mechanically manufactured a false null on the same data that v2's corrected instrument detected at p = 9.24 × 10⁻¹⁸. The methodology-cautionary page.
  • Confound probes — the formal identification argument — the rigorous reconstruction of how probes A, B, and C jointly identify overlap density as the proximate predictor of preamble lift, ruling out simpler "judge-priming" and "expert-framing" stories.
  • Related work — how these findings situate within the 2024–2026 literature on persona/system-prompt effects, the alignment-vs-pretraining split, format-as-variable studies, and LLM-as-judge evaluation.