Skip to content

AGENTS.md

Latest commit

 

History

History
45 lines (41 loc) · 2.95 KB

AGENTS.md

File metadata and controls

45 lines (41 loc) · 2.95 KB
Raw

Independent executable-experience study

Read README.md. This repository owns the commissioned question's workload, methods, protocols, execution, diagnosis and publication. The root maintains theory and synthesis. In an assigned fresh session, proceed autonomously within the bounded question; preparation itself has launched no model runs.

  • Begin with the closest public methods and develop functioning acquisition and a competent reconstruction alternative. Workload discovery is authorized. Small reproducible tasks can suffice; a library advantage is not required.
  • Specify what persists across sessions, what was acquired and what is supplied. Do not confuse executable data flow, retained information and avoided reconstruction. Source in a log or archive remains available code. Permit direct loading when appropriate instead of forcing model regeneration.
  • Keep execution, current observations, checking and repair available to both alternatives. Evaluate complete behavior and all incurred work, including failures. Separate research search costs from a selected deployment's costs.
  • Diagnose acquisition failures purposefully within bounded resources. A valid program or successful execution is not by itself successful task acquisition. Preserve failures; assess developed claims on fresh evaluation material.
  • Pin paper versions, code, datasets and reused artifacts; identify adaptations. For arXiv metadata use the query API, cache responses and use a descriptive User-Agent. Keep API/OAI requests to one connection and at least three seconds apart, honor Retry-After and coordinate the limit across clients under control.
  • Keep evidence and reproduction instructions local and identifiable by Git revision. Distinguish observations, author-reported results and inference. Close on explanation, a demonstrated limitation or a concrete resource constraint.
  • Do not modify the root or sibling studies. Copy needed artifacts with provenance. Coordinate shared hardware for heavy jobs and timing comparisons; independent scientific questions need not wait for one another.

Resources

  • Dedicated Mac Studio M1, 64 GB unified memory, serving over Tailscale through Docker Model Runner (preferred): https://mac-studio-7hr7.taile71f88.ts.net/engines/v1/chat/completions.
  • Local open-weight models through docker model; OpenAI models through codex; SpaceXAI models through agent.
  • Start with fixed model weights if sufficient to identify the question. Serving alone does not establish gradient access, mutable adapters or reusable KV state. Verify the capabilities actually needed; none was tested in preparation.
  • Use uv for Python and Docker/Compose for supporting services where useful. Keep credentials, environments, caches and model downloads out of Git; retain evidence needed for claims. Use controlled task environments for experiments. Explain a material expansion beyond the available resources before undertaking it.