Skip to content

AGENTS.md

Latest commit

 

History

History
130 lines (115 loc) · 7.97 KB

AGENTS.md

File metadata and controls

130 lines (115 loc) · 7.97 KB
Raw

Trinite agent instructions

Follow the current user's scope. Read README, operating instructions, invariants, architecture, modules, roadmap, and the affected topic contracts before editing. Inspect any more specific AGENTS.md in the paths you change.

Implementation

  • Keep one authoritative numerical updater, scorer, admission operation and safe tensor reader. Use small explicit modules; no registry or workflow framework.
  • CPU with the hash-locked dependencies is the canonical native numerical lane. Never substitute a backend, download model code implicitly, or use pretrained/ teacher inputs in an existing native run. The explicitly owner-approved teacher-origin formal lane is a separate versioned exception for verified numeric operands only, never raw prose or answers. The isolated external NCM reference below is not a native model or training input. Geometry loss stays zero until its own reviewed protocol.
  • Keep narrative documentation under docs/ and README concise. This root file is the agent entrypoint; detailed instructions remain in docs/INSTRUCTIONS.md.
  • Preserve historical protocols, fixtures, producer snapshots and negative results. New numerical or evidence semantics need a new version and source-bound request. Do not refresh fixtures, reseal original evidence or weaken guards to pass checks.
  • Preserve the flat v1 maths/retention modules and their source closure. New v2 helpers belong in src/trinite/retention_v2/; include their complete runtime dependencies and fixed launcher in the new receipt and boundary tests.
  • Freeze and publish the producer/protocol before target experiment outcomes. Reused test families remain exploratory; never tune a ratio on their results.
  • Keep unique training diagnostics separate from scheduled repeated visits. Test/validation examples cannot enter updates or training diagnostics.

Current workflow entry points

Read the linked contract and actual CLI before using a lane; these checkout-only pilots do not extend the installed native API or unlock later research phases.

  • Fixed-code scale-refit training: scripts/run_scale_refit_training.py owns freeze/run/full training verification. Preserve the original four-cell report and its environment. Slightly lower validation loss did not improve complete-answer accuracy.
  • Frozen-model decoding: use scripts/run_frozen_decoding_v2.py for new requests, runs and verification. The v1 launcher and evidence remain frozen for historical reproduction only. Syntax-valid constant answers are not input-sensitive reasoning; inference replay is not optimizer replay, and equal token caps are not equal compute.
  • External NCM reference pilot: scripts/run_ncm_reference.py is a separate pinned pretrained Sudoku feasibility lane under scripts/ncm_reference_v1/, not a Trinite capability result.

External NCM execution boundary

  • Require a reviewed, green main commit and explicit local-run authorization before GPU execution. Use a clean checkout of that exact commit and record its Git SHA separately from the launcher's actual source closure. Freeze a request before outcomes; keep its SHA-256 independently. CPU contract tests use software doubles and do not demonstrate real-checkpoint compatibility, GPU execution or solve rate.
  • Follow the documented separate Python 3.11/CUDA environment and exact pinned artifacts; do not alter requirements.lock, drivers or system settings, use paid compute, or fall back to CPU/another GPU. The admitted target is one RTX 5060 Ti 16 GB-class device with native sm_120 support. Preserve the frozen time, allocator-memory and workload bounds; the allocator cap is not a whole-device VRAM cap.
  • Run check-assets before loading the two pinned local checkpoints. Keep weights_only=True, finite/shape checks and upstream import/hash checks intact. No unrestricted pickle, custom loading allowlist, remote model-code execution, upstream dataset, implicit download or unsafe fallback. Permissive upstream licensing alone does not satisfy native data admission.
  • Use fresh destinations for request, run and replay artifacts. Retain complete attempt directories, partial progress, failures and both archives on mismatch; never overwrite historical fixtures or reseal evidence. Keep each printed ZIP SHA-256 independently and supply it to verify/replay as documented.
  • verify is read-only offline archive/source/metric verification, not a fresh GPU run or hardware attestation. replay actually reruns GPU inference and compares deterministic result/environment bytes, not observed timings or peak memory. Never weaken source/asset pins, budgets, equality or tolerances to pass.
  • Report clue-only raw generation separately from gold-codec diagnostics. No solver, clue overwrite, repair, retry or answer-conditioned generation. Four correlated visible formal boards are a smoke test, not a held-out benchmark, paper reproduction, speed comparison or native learning/release evidence.

Verification and reporting

Name affected TRI requirements and compatibility. Use actual admission and upstream verification in evidence tests. Process fakes are suitable for testing controller failures, not for certifying custody or numerical outputs.

PYTHONPATH=src python -m unittest discover -s tests -v
PYTHONPATH=src python -m unittest discover -s tests/upstream -v
# Standard-library external-reference contracts, without model loading:
PYTHONPATH=scripts python -B -m unittest discover -s tests/ncm_reference_v1 -v
# With requirements.lock installed in the chosen CPU environment:
PYTHONPATH=src python scripts/check_research_cpu.py
python -m pip check

The historical CPU aggregate covers model, training, geometry and its declared research suites, not every later checkout-only pilot. Also run the affected topic suites/workflows linked above; do not treat one aggregate as complete coverage. Never silently skip missing numerical dependencies. Check changed-doc links/assets, conflicting status claims and diff whitespace. Never cache verifier, admission or test success instead of running the current check.

Retention v2, budget and task-balance source maps also bind AGENTS.md. The retained budget producer snapshot preserves the original instructions. Historical replay requires the exact producer sources and matching environment for that run, not an evolved checkout. Preservation tests check archived instructions but still pin every runtime file. Those live run/verify source maps continue to bind current instructions; changed guidance requires a fresh request, never a substituted historical hash.

Verification receipts must distinguish closed custody, source/input binding, safe structure, scheduled visits/targets, final predictions/loss and complete optimizer replay. Full replay starts from the declared native seed and compares model/AdamW tensors and every numerical history/diagnostic against externally bound expected artifacts. Do not describe final inference as update replay. Foreign environment rejection is intentional: retain the mismatch and use a new request for a separately identified run, without patching the old fingerprint.

Retain all expected cells, failure/timeout logs and failed comparison receipts. Read-only verification must leave all run bytes unchanged. Report completed checks, their environment and limits; custody is not learning adequacy. Original fold/composed learning, format selection, scaling, Phase 5 and release remain blocked until their own gates are demonstrated. Open a reviewable PR by default; a direct-main edit requires the user's explicit scope. Merging, tagging and releases require the user's corresponding task. Report CI for the exact published commit, distinguishing passed, failed, pending and unrun checks.