ChronoMoEv4

Bob: consequence-accumulating control plane for Mixture-of-Experts models.

Bob observes MoE routing decisions, accumulates consequences, and learns when to take the cheap path. The model does the thinking. Bob decides how much thinking is necessary.

Version: v4.0.0-alpha1 License: PolyForm Noncommercial 1.0.0 (see bob_core/LICENSE) Backend: Qwen1.5-MoE-A2.7B on Apple Silicon via MLX

---

Structure

bob_core/          Control plane (15 modules)
backends/          Adapter layer (MLX only)
experiments/       Experiment scripts + shared prompt corpus
tests/             70 tests (triad monitors, relational graph, association basins, B1 equivalence)
designs/           Specifications (triad monitor spec)

Quick Start

# Install dependencies
pip install -r requirements.txt

# Run tests (no model needed)
python3 -m pytest tests/ -v

# Verify imports
python3 -c "from bob_core import BobSubstrate, BOB_CORE_VERSION; print(BOB_CORE_VERSION)"

# Run governed experiment (requires Qwen model + MLX)
source qwen_moe_mlx/bin/activate
python3 experiments/qwen_governed.py --smoke --seed 42

# Run memory bias validation (Phase 8c)
python3 experiments/memory_bias.py --smoke --seed 42

What Bob Does

Component	What it does
Motif store + compound gate	Recognises repeated routing patterns, offers cheap-path shortcut
Three clocks (fast/medium/slow)	Detect instability at different timescales, modulate intervention
Governor + ledgers	Authorise or block cheap-path commits based on scar/cost/commitment history
Triad monitors	Detect angel/devil/maniac expert pathologies per layer
Conflict register	Track angel-devil co-occurrence, mode switching
Relational graph	Store typed entity triples, alias lookup. Inert -- does NOT touch routing
Association basins	Pre-softmax logit bias injection wiring. Plumbing only -- see Phase 8c

Phase 8c: Memory Bias Validation

5-condition A/B experiment testing whether pre-softmax logit bias from entity-specific routing signatures does anything useful.

Gate	What it checks
Gate 1	B1 (scale=0.0) identical to baseline -- plumbing clean
Gate 2	Toggle A (graph only) identical to baseline -- graph inert
Gate 3	Bias audibility in routing units
Gate 4	Geometry stability (Neff, entropy, scars)
Gate 5	Primary success metric (synthetic entity loss)

Results: experiments/phase8c_logs/

Conclusion: Memory bias is audible but does not improve synthetic entity loss at tested scales. Graph-only is the correct architecture at this point.

Tests

70 passed in 0.5s

tests/test_triad_monitors.py      17 tests  Angel/Devil/Maniac monitors + conflict register
tests/test_memory_graph.py         25 tests  Relational graph CRUD, serialization, assertion detection
tests/test_memory_basins.py        24 tests  Association basins, entity linking, bias computation
tests/test_b1_equivalence.py        4 tests  Deterministic B1=C0 guarantee, cross-run determinism