Skip to content

AGENTS.md

Latest commit

 

History

History
46 lines (35 loc) · 2.39 KB

AGENTS.md

File metadata and controls

46 lines (35 loc) · 2.39 KB
Raw

Instructions for AI agents

This repository is a reproducible research project, not a generic chat application. Preserve the distinction between development evidence and confirmatory evidence.

Start here

Read these files in order before changing code:

  1. README.md
  2. docs/results-ecp6.md
  3. docs/protocol-specification-ecp6.md
  4. docs/research-design-ecp6.md
  5. docs/AI_AGENT_START.md

Then run:

python3.12 -m venv .venv
.venv/bin/pip install -e '.[dev]'
.venv/bin/pytest
.venv/bin/ecp6 --config config/ecp6.yaml validate

Non-negotiable research rules

  • Never tune a model, split, threshold, metric or seed after reading its confirmatory test result.
  • Never edit a frozen config/ecpN.yaml in place. Create config/ecpN+1-development.yaml for new work.
  • Use smoke and develop while the new test split is sealed. Only a frozen config may use experiment --unseal-test.
  • Record the config SHA-256, split SHA-256 and freeze time before unsealing a new test.
  • Keep train, validation and confirmatory test meanings disjoint and deterministic.
  • Preserve channel isolation: receivers may receive symbol matrices only, never meaning IDs, sender parameters or hidden state.
  • Preserve hard-symbol evaluation. Continuous relaxations may be used for training only.
  • Report negative and partial results. Do not silently discard failed seeds or variants.
  • Do not commit generated runs/*/ directories, checkpoints or episode logs. Put compact, reviewed evidence under evidence/.
  • Do not claim that ECP is a general replacement for human language. Every conclusion must state the synthetic-world and architectural limits.

Change discipline

  • Add or update tests for every invariant or behavior change.
  • Run the full test suite and the relevant config validation before committing.
  • Keep existing ECP-0 through ECP-6 configurations reproducible.
  • Document new experiments in three stages: preregistration, sealed development log, confirmatory results.
  • Write new configurations, documentation, reports and user-facing output in English.
  • Preserve the completed ECP-8 Batch 1 capacity isolation and its sealed confirmatory split. Any Batch 2 must keep the 16-bit arm fixed and preregister exactly one structure-induction mechanism.

ECP-8 Batch 1 found that surplus capacity nearly removes collisions but does not reach injectivity or compositional validation. Do not continue a channel-size search.