FATE-O-SISTERS / MOIRA
This is the Hugging Face landing page for MOIRA, an experimental sequence-model research project. The executable Phase-0 reference and canonical source are maintained in the
FATE-O-SISTERS GitHub repository.
Current implementation
The reference is a deterministic CPU/FP32 PyTorch research system. It has one shared model with three internal execution orientations:
- Clotho uses bidirectional visibility for representation-oriented work.
- Lachesis uses sealed source visibility with causal target visibility for sequence-to-sequence work.
- Atropos uses causal visibility for generation-oriented work.
The model combines local Thread attention, bounded recurrent Spool state, and Knot landmark attention. The repository also includes typed capability capsules, bounded Pulse state, deterministic training and checkpoint tests, and an agent layer with task budgets, request-bound permits, and receipts.
The current source revision replaces Thread's full quadratic attention-score path with a compact local window. It also adds a prompt-only DSPy/GEPA candidate adapter. That adapter can stage bounded prompt artifacts for evaluation; it cannot train model weights, call tools, or promote a candidate.
Release status
No trained checkpoint or inference weights are published in this repository. The proposed 3–4B target, GPU training, precision research, parameter growth, multimodal capability claims, performance figures, and production deployment remain future work that requires separate implementation and measured evaluation.
Use the GitHub repository for the current research code and its test instructions. Future Hub releases will contain versioned artifacts only after their manifests, compatibility checks, provenance, and evaluation evidence are available.
