Skip to content

AGENTS.md

Latest commit

 

History

History
57 lines (44 loc) · 3.26 KB

AGENTS.md

File metadata and controls

57 lines (44 loc) · 3.26 KB
Raw

AGENTS.md

This fork is the single active repository for the artificial-hearing research program.

GitHub / LLM connector entrance

When entering through GitHub, an LLM connector, or another remote agent surface, do not reconstruct the repository from search results alone.

Use this bounded entrance sequence:

current develop HEAD
→ README.md
→ AGENTS.md at the same HEAD
→ recent commits
→ smallest task-relevant research / source / tests
→ exact target file
  1. Resolve dissonance-git/deepSTRF and record the current develop commit before substantive work.
  2. Read README.md for the project model and this AGENTS.md for operating law from that same repository state.
  3. Inspect recent commits to learn what is actively moving. Recency is an activity hint, not authority over current contracts or source truth.
  4. Hydrate only the task-relevant region. Do not load Omniphony, Game Music Interpreter, Helix, or retired libaural internals unless the task crosses that boundary.
  5. Before every GitHub replacement write, re-fetch current develop and the exact target file. If develop moved since preflight, reconstruct the edit from the new file content.
  6. Write against the exact current blob SHA. Preserve unrelated concurrent work. Never replace a file from a cached or reconstructed older copy.
  7. After publication, fetch the resulting commit, inspect its changed paths, and confirm it remains in current develop history.
  8. Report repository publication, tests, CI, external-teacher execution, and scientific adjudication as separate evidence states.

Fast routes:

  • project orientation: README.md
  • operating law: AGENTS.md
  • active research and migrated findings: research/
  • unified hearing mechanisms: deepSTRF/hearing/
  • neural models: deepSTRF/models/audio/
  • neural datasets: deepSTRF/datasets/audio/
  • metrics: deepSTRF/metrics/
  • regression evidence: tests/

Prefer current files and the smallest directly relevant research document over broad historical scans.

Working rules

  • Do not create a parallel libaural architecture or resume active work in the retired libaural repository.
  • Prefer experiments over architecture prose when a bounded test can answer the question.
  • Preserve upstream deepSTRF neural-data/model infrastructure unless a concrete experiment requires change.
  • Treat biological models, neural datasets, learned models, torch_amt mechanisms, DAL-style inverse design and Game Music Interpreter source truth as teachers or controls, not automatic runtime architecture.
  • Keep teacher fidelity, auditory-obligation retention, compression and runtime suitability as separate claims.
  • Reuse one shared representation/metric surface instead of adding adapter chains.
  • Game Music Interpreter remains an external causal fixture generator unless importing a small mechanism removes a real experimental blocker.
  • Rich teachers should be ablated toward the smallest causal/stateful mechanism that preserves the tested relation.
  • Keep negative results and counterexamples.
  • New canonical research belongs in this repository.

Current frontier

Continue from research/AUD-VGM-001.md, native AUD-REP migration, original ICNet evaluation, StateNet compression tests, torch_amt mechanism challenges and differentiable inverse-design ablations.