Imported from Etha-Sun/Verus-Skill-Learning (AGENTS.md). Install upstream with npx skills add Etha-Sun/Verus-Skill-Learning. Copyright stays with the author.

Repository Agent Instructions

This repository is the only active development workspace. The legacy verusys-result tree is read-only data and historical-run storage.

Startup Context

• If .agent-context.local.md exists, read it for machine-local paths. It is local-only context and must not be committed.
• For research tasks, read these files in order before acting:
  1. research_memory/BOOTSTRAP.md
  2. research_memory/CURRENT.md
  3. research_memory/INDEX.md
  4. research_memory/projects/verus_self_evolving/PROJECT.md
• ./research_memory is the canonical research memory for this repository. This repository rule overrides any global skill or instruction that still points research-memory writes at the legacy verusys-result directory.

Local Skills

• Optional project-local skills are installed below .agents/skills/ from an external ARIS checkout. Both the links and their machine-local manifest are ignored and must not be committed.
• Do not edit symlinked skills in place. Update the external checkout and run its project installer with --reconcile --no-doc.

Working Style

• Think before coding. State material assumptions, surface ambiguity, and ask before choosing among interpretations that would materially change scope.
• Prefer the simplest implementation that fully solves the requested problem. Do not add speculative features, abstractions, or configurability.
• Make surgical changes. Match the existing style, preserve unrelated work, and remove only unused code created by your own change.
• Work toward explicit, verifiable success criteria. For multi-step work, give a brief plan that pairs each step with its verification.

Data And Output Safety

• Treat all raw datasets and sealed data below VERUS_SKILL_DATA_ROOT as read-only. Never modify, move, rename, or commit them.
• Write generated experiment outputs only below VERUS_SKILL_RUN_ROOT. Keep only reviewed compact summaries, contracts, and pointers in this repository.
• Do not commit personal absolute paths, secrets, API credentials, raw traces, sealed data, meeting transcripts, token tables, or complete run directories.
• The reviewed fixed-claude-stratified-80-seed20260814/ benchmark fixture is an explicit exception: its fixed train/val/test sources, manifests, and historical split-stratification token/time fields remain tracked so every method uses the same unchanged test-20. This exception does not cover run outputs, provider ledgers, prompts, responses, or other token tables.
• All newly created files and directories must use English ASCII names.

Research Evidence Boundaries

• Information gain is a secondary offline proxy, not the primary endpoint.
• R040 is complete, R041 prompt distillation is next, and the R042 frontier experiment is not complete.
• Do not claim that distilled knowledge improves solved rate or token efficiency until leakage-safe live evaluation directly establishes it.

Research Closeout

Before finishing non-trivial research work:

1. Update research_memory/CURRENT.md with the active result, caveat, and next action.
2. Run python3 research_memory/scripts/mem.py index.
3. Report the durable memory/artifact paths and confirm raw-data safety.