AGENTS.md

Eidrion: instructions for research agents

Start with README.md, docs/EXPERT_REVIEW.md, docs/REPRODUCIBILITY.md, and research_notes/research_status.json. The scientific source is v1.0.0; the HF packaging version is v1.0.0-hf.1. Do not silently relabel archival source or experiments after changing them.

Evidence rules

• Separate written conditional proofs, executed finite checks, historical experimental results, fresh reruns, and proposals.
• Preserve the exact author line and source lineage. Do not invent DOI, publication, peer review, priority, or world-leading status.
• The archived source is immutable. Develop in a fresh extracted copy or a separate branch; retain the original ZIP and source locks.
• This release contains the main archive only. Four large companion groups are absent. A stored audit report is not the underlying NPZ/checkpoint evidence.
• Retention of an acquired reference prediction is distinct from true-label accuracy. Renewal validates accepted checkpoints and does not supply an all-future ever-error bound.
• The external feature map and output universe are fixed in the principal theorem. A change to either requires a new argument and explicit tests.
• Float64 guard diagnostics do not prove a worst-case roundoff enclosure. Calibration independence and distribution premises cannot be established by checksum agreement.
• Report all predeclared seeds/settings, including refusals and unsuccessful candidates. Track total persistent learner payload, retained evidence, transient memory, and experiment storage separately.

Working and review

python tools/verify_release.py checks the HF package without network access. python tools/extract_research.py --output work extracts the original main archive into a fresh directory. Use Linux/WSL for locked benchmark reproduction; the user-facing upload BAT is a separate Windows workflow.

When changing the algorithm, describe the exact retention contract first, provide a counterexample-oriented test, update the checkpoint/controller binding rules, and compare against the strongest supplied simple control. Record fit-data, evidence, memory, and compute differences. Do not write a test that merely mirrors the new implementation.

For a research handoff, supply: claim and assumptions; proof or unresolved gap; source revision and dependency versions; complete protocol/seeds; observed effects and intervals; resource accounting; negative regimes; and next falsification. Preserve historical results as historical.

Do not upload automatically during review. The uploader performs external changes only when deliberately invoked with --execute or via UPLOAD_TO_HUGGINGFACE.bat.