Research Transfer Judgment

Core Insight

AI research findings do NOT transfer universally. Whether a finding applies to your use case depends on whether the model already has strong learned behavior for the task.

The Framework

Step 1: Does the model have prior learned behavior for this task?

YES (strong prior) — The model learned this from training data. Examples: writing clean code, grammar, formatting, common design patterns.
- → Positive directives COMPETE with the model's existing behavior. Two optimization targets create interference.
- → Negative constraints COMPLEMENT existing behavior. They remove failure modes without creating conflict.
NO (no prior) — The task involves novel procedures not in training data. Examples: custom file formats, specific workflow phases, proprietary fingerprint schemas, domain-specific protocols.
- → Positive directives are NECESSARY. They fill a vacuum the model can't fill from training.
- → Negative constraints still work everywhere.

Step 2: Context priming (applies regardless)

Any structured text in a system prompt — even random text — shifts the model into a more careful, methodical execution mode. The benefit comes from the pattern of having instructions, not the specific content. This is why random rules help as much as expert-curated ones.

Decision Matrix

Task Type	Positive Directives	Negative Constraints	Random Structure
Model has strong prior (coding, writing)	Harmful — competing optimization	Beneficial — removes failure modes	Beneficial — context priming
Model has no prior (custom workflows)	Necessary — fills vacuum	Beneficial	Beneficial

Worked Examples

From today's signal (CLAUDE.md rules study, April 13, 2026)

679 CLAUDE.md files tested across 5,000+ SWE-bench runs:

Random rules performed equal to expert rules (context priming explains the benefit)
Negative constraints ("don't refactor unrelated code") = only individually beneficial type
Positive directives ("follow code style") = actively degraded performance

Applied to AI Radar's CLAUDE.md

This system's 700-line CLAUDE.md contains both types:

"Use [wikilinks](/wiki/wikilinks) for internal links" — novel procedure, no training prior → KEEP (necessary)
"Generate fingerprint: {YYYY-MM-DD}|{primary-node}|{normalized-key-phrase}" — novel schema → KEEP
"Keep signal descriptions substantive" — overlaps with model's trained writing behavior → CANDIDATE FOR REMOVAL

Applies To

Evaluating any AI research paper's applicability to your specific deployment
Writing instruction files for AI coding agents (CLAUDE.md, .cursorrules, system prompts)
Deciding whether to adopt a technique that worked in a benchmark context
Any situation where you're asking "does this finding apply to my use case?"