Weft is the Layer 2 model and weight-space layer of the Siliconcordat stack. It compiles public written rules and domain conventions into small, verifiable weight adapters for local open-weights models. It is for teams and engineers who want repeatable, inspectable model adaptation instead of hidden fine-tunes.
Weft sits between written human rules (Plane 1) and local model inference runtimes. Instead of dumping tens of kilobytes of repeated instructions into every prompt, a public compiler turns approved text rules into bounded low-rank adapters that load behind strict verification gates.
- Deterministic compilation: A public CPU
float64compiler builds null-space projected adapters using history-aware covariance tracking (BetaEdit) and emits exact quantized factors. Running the compiler twice on the same inputs gives identical bytes. - Exact checkpoint and lineage binding: Every adapter binds to five checks: tensor layout, full checkpoint weights, config and tokenizer files, source rule digests, and runtime serving flags. Any one-byte drift causes rejection before GPU load.
- Division-free integer certificates: Post-quantization norm, spectral, leakage, and safety-direction bounds are stored as scaled integers checked without floating-point division.
- Binary GGUF LoRA and
weft-slab/v0.1wire containers: Adapters export directly to F16 GGUF LoRA files for simultaneous multi-adapter serving inllama.cpp, or to compactweft-slab/v0.1asymmetricINT4/INT8/F16containers with embedded certificates, L5zone_manifest_idpins, and 32-byte SHA-256 trailers. - Verifier-owned behavioral gates: Expected canary outputs live with the verifier, never inside untrusted adapter headers. Promotion requires both visible canaries and private holdouts.
Weft enforces five hard vetoes inherited from Model Domain Expansion research:
- No hidden weight deltas. Every Plane 2 adapter must be re-derivable from public Plane 1 inputs with a public, simple compiler.
- No automatic resolution of conflicting human rules. Contested rules stop compilation and return to the owner.
- No closed-weight injection. Adapters run only on self-hosted open-weights runtimes with load-time verification, audit logs, and an instant kill switch.
- No premature protocol standard. Wire extensions stay vendor-namespaced until outside users exercise them.
- Human authority stays in plain text (Plane 1). Every weight adapter is a disposable cache.
| Layer | Repository | Role |
|---|---|---|
| L5 | siliconcordat/mde |
Universal structured agent protocol and read-before-write locks |
| L4 | siliconcordat/exo |
Continuous harness, human approval gates, causal memory, zone routing |
| L3 | siliconcordat/warp |
Operating system boundary, sandboxes, and fail-closed load verifier |
| L2 | siliconcordat/weft |
Deterministic weight compiler, integer certificates, multi-adapter composition |
| L1 | siliconcordat/loom |
Hardware theory and verification physics research |
- ARCHITECTURE.md: Two-class adapter policy, three-pack composition, verification primitive, and hardware layout.
- PLAN.md: Phased build milestones, test suite, runnable experiments, and kill criteria.
- experiments/RESULTS.md: Verbatim measured outputs from Experiments E1 to E8, including failure boundaries, live
llama-serverresults, Latent Lattice MDE probes, and multi-layer SwiGLU/RMSNorm tangent projection.
python3 -m venv .venv
.venv/bin/pip install -e ".[dev]"
.venv/bin/pytest -v
.venv/bin/python experiments/e1_shared_token_composition.py
.venv/bin/python experiments/e2_anti_goodhart_gate.py
.venv/bin/python experiments/e3_real_checkpoint_smoke.py
.venv/bin/python experiments/e4_asymmetric_quant_packing.py
.venv/bin/python experiments/e5_noise_k0_scaling_and_readout.py
# Opt-in live 1.5B llama-server multi-LoRA check:
WEFT_LIVE_BASE_GGUF=/path/to/qwen2.5-1.5b-instruct.gguf .venv/bin/python experiments/e6_live_llama_server_lora.py
.venv/bin/python experiments/e7_latent_lattice_mde.py
.venv/bin/python experiments/e8_swiglu_rmsnorm_tangent_coupling.pyStatus: v0.1 core compiler (with asymmetric INT4/INT8/F16 packing, weft-slab/v0.1 containers, and multi-layer symmetric TangentBetaEdit), llama.cpp-verified GGUF v3 F16 LoRA codec, five-layer binding, anti-Goodhart anchor harness, 58 tests, and Experiments E1 to E8 are verified locally. For E8 that covers Parts A and B only, and Part C is an unverified source claim.
Experiment E6 verifies simultaneous multi-LoRA serving and 0.99 ms bit-exact rollback inside live llama-server on Qwen2.5-1.5B-Instruct, Experiment E7 verifies Latent Lattice capability banks and atomic multi-bank rollback, and Experiment E8 verifies exact SwiGLU/RMSNorm finite-perturbation coupling and multi-layer symmetric TangentBetaEdit on a synthetic CPU float64 stack.
The live Qwen3.5-4B-Q8_0 E8 Part C numbers are an unverified source claim, because no re-runnable script or base-model hash is committed. Claims carry explicit labels (verified, source claim, or proposed).
Code is licensed under Apache-2.0 (see LICENSE). Documentation is licensed under CC BY 4.0.