Initial release: MoA v0.1.0 — safety-first model-agnostic agent framework

Author: Yoder23

.github/workflows/tests.yml

@@ -0,0 +1,31 @@
+name: Tests
+
+on:
+  push:
+    branches: [ main ]
+  pull_request:
+    branches: [ main ]
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.10", "3.11", "3.12"]
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v5
+      with:
+        python-version: ${{ matrix.python-version }}
+
+    - name: Install core package (no external deps)
+      run: pip install -e ".[dev]"
+
+    - name: Run zero-dep proof
+      run: python verify_moa.py
+
+    - name: Run pytest suite
+      run: pytest tests/ -v --tb=short

.gitignore

@@ -0,0 +1,44 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+*.egg
+*.egg-info/
+dist/
+build/
+.eggs/
+*.whl
+
+# Virtual environments
+.venv/
+venv/
+env/
+
+# Testing
+.pytest_cache/
+.coverage
+htmlcov/
+.tox/
+
+# MoA runtime state
+moa_state/
+*.audit.jsonl
+audit_log.jsonl
+
+# Model files
+*.pt
+*.ckpt
+*.safetensors
+*.bin
+
+# IDE
+.vscode/
+.idea/
+*.swp
+
+# macOS
+.DS_Store
+
+# Windows
+Thumbs.db

CLAIMS.md

@@ -0,0 +1,94 @@
+# CLAIMS.md — What MoA Actually Proves
+
+This document separates proven claims from aspirational ones.  
+"Proven" means there is executable code or a test you can run right now.
+
+---
+
+## ✅ Proven — runnable code, zero deps
+
+### P1: FORBIDDEN action types are rejected by code, not prompts
+**Evidence:** `tests/test_safety.py::test_gate_rejects_deceive`  
+`SafetyGate.evaluate()` checks `HardConstraints.FORBIDDEN_ACTIONS` (a Python frozenset) before
+executing any action. The LLM never sees a FORBIDDEN action because the check runs outside the LLM context.
+
+```python
+assert ActionType.DECEIVE_USER in HardConstraints.FORBIDDEN_ACTIONS
+gate = SafetyGate()
+d = gate.evaluate(Action(type=ActionType.DECEIVE_USER))
+assert not d.approved  # always False, regardless of LLM output
+```
+
+### P2: SIMULATION and FICTION claims cannot become FACT
+**Evidence:** `tests/test_safety.py::test_verifier_rejects_simulation` and `test_verifier_rejects_fiction`  
+`FormalVerifier.verify()` checks `claim.modality` before any other check. No amount of evidence
+or confidence can override this:
+
+```python
+claim = create_claim("event", "would_cause", "effect", Modality.SIMULATION, confidence=0.99)
+promoted, _ = FormalVerifier().promote_to_fact(claim, [evidence])
+assert promoted == False   # always
+```
+
+### P3: EventLog is append-only — nothing is deleted
+**Evidence:** `tests/test_ir.py`, `verify_moa.py::check_memory`  
+`EventLog` has no `.delete()`, `.pop()`, or `.clear()` method. The only public mutation method is
+`.write()`. This is enforced by the class definition, not by runtime checks.
+
+### P4: Actions above threshold are rejected regardless of LLM output
+**Evidence:** `tests/test_safety.py::test_gate_rejects_high_irreversibility`  
+`HardConstraints.MAX_IRREVERSIBILITY = 0.8`. Any action with `ActionMeta.irreversibility_score > 0.8`
+is rejected by `SafetyGate.evaluate()`. Same for `blast_radius > 0.7` and `deception_risk > 0.1`.
+
+### P5: Framework works with any LLM implementing BaseLLMBackend
+**Evidence:** `verify_moa.py`, `tests/test_agent.py`  
+All 48 tests run with `MockBackend` — zero external dependencies. The same code path runs with
+OpenAI, Anthropic, Ollama, HuggingFace, and LayerCake backends.
+
+### P6: Formal verification requires evidence + confidence ≥ 0.8
+**Evidence:** `tests/test_safety.py::test_verifier_rejects_low_confidence`  
+Claims with `confidence < 0.8` cannot be promoted to FACT even with supporting evidence.
+This prevents confident-sounding hallucinations from entering the FACT graph.
+
+---
+
+## ⬜ Aspirational — require production deployment to fully evaluate
+
+### A1: Self-evolving domain modules
+**Status:** Architecture exists (`LayerCakeBackend.paste_domain()`). Bit-exact paste proven in
+the companion [LayerCake repo](https://github.com/Yoder23/layercake). "Self-evolution" at scale
+(agent improves itself over time) requires training infrastructure and long evaluation runs.
+
+### A2: Production-scale multi-agent coordination
+**Status:** The OODA loop runs. Multi-agent orchestration (multiple `MoAAgent` instances sharing
+memory) is not yet implemented. The memory layer is designed for it (shared `AgentMemory` is
+threadsafe for reads), but it has not been stress-tested.
+
+### A3: Semantic memory at scale
+**Status:** `EpisodicBuffer.retrieve_semantic()` works with any `embed()` function. Quality of
+retrieval depends entirely on the embedding model. `MockBackend` uses MD5-based embeddings that
+test the pipeline but not semantic accuracy.
+
+---
+
+## ❌ Not claimed
+
+- **AGI** — This is an agent framework with safety constraints. It is not general intelligence.
+- **Superintelligence** — No component here exhibits superhuman capability on any benchmark.
+- **Perfect safety** — Hard constraints eliminate specific forbidden action types. They do not
+  prevent all possible misuse. Security depends on the threat model.
+- **SOTA text generation** — MoA delegates generation to whatever LLM you provide. Generation
+  quality equals the quality of the backend model.
+
+---
+
+## How to challenge these claims
+
+Every ✅ claim has a test you can run:
+
+```bash
+python verify_moa.py          # 10 checks, 0 deps, < 1 second
+pytest tests/ -v              # 48 tests, no API key
+```
+
+If a test fails, open an issue. Claims without passing tests get downgraded.

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 Yoder23
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.