ActPlane Empirical Study

This directory keeps the curated empirical-study artifacts that are useful to understand why ActPlane's policy language targets process, file, network, and temporal agent behavior. Scratch notes, raw model runs, tuning logs, and full corpus workspaces stay on the artifact or backup refs described in docs/ARTIFACT.md.

Snapshot

The study analyzed agent instruction files (CLAUDE.md and AGENTS.md) from popular AI-agent code repositories. The retained product-branch artifacts are aggregate outputs, not the full raw corpus.

Metric	Value
In-corpus repositories	144
Instruction files	228
Instruction-file text	39,803 lines
Candidate normative lines	3,762
ActPlane-related candidate lines	529
Repositories with at least one ActPlane-related candidate	101

The candidate-line counts are keyword/category extraction results and should be read as aggregate evidence for prevalence, not as a hand-labeled ground truth. The product branch keeps the aggregate study outputs that are independent of compiler syntax. DSL-specific ruleset artifacts should live only on artifact refs after they have been regenerated and validated against the current compiler.

Main Findings

• Agent instruction files frequently contain operational guardrails that are below the tool layer: VCS gates, secrets handling, test-before-commit rules, workspace boundaries, destructive-operation guards, network egress limits, and mediation through project tools.
• These guardrails map to ActPlane primitives: exec argument matching, labeled file and endpoint sources, source-to-sink label flow, after ordering, lineage gates, target scoping, and declassification.
• Style and code-quality instructions are common in the broader corpus but are intentionally outside ActPlane's enforcement scope unless they correspond to a concrete OS-observable action.

Retained Artifacts

• candidate_rules_144.tsv: aggregate candidate-line extraction with repo, file family, line number, category guess, and source text.
• figures/: generated summary figures for the empirical study.

The raw corpus, raw traces, intermediate coding notes, old evaluation drafts, and exploratory scripts are intentionally not kept in the product branch.