Five tier crates ported. 589 tests pass across Phase 1-4, zero
failures, zero warnings.
jouleclaw-liquid
Ported from pattern-lang/liquid (~1123 LOC, 5 files).
L3 tier — Closed-form Continuous-time (CfC) recurrent language
models. The SSM class with O(L) state evolution vs transformer
O(L²) attention. This is the v0.3 "SSM bet" from the architecture
proposal — already shipping in the JouleClaw reference impl on
day one.
jouleclaw-prism
Ported from pattern-lang/prism (~3066 LOC, 6 files).
L3 tier — 1-bit / ternary BitNet-class kernels. Ternary transformer
forward pass. The lowest-joule local-model tier. Provides the
TernaryDecoder backbone that jouleclaw-lmm builds on.
jouleclaw-mrl
Ported from pattern-lang/mrl (~1464 LOC, 7 files).
L2 tier — Matryoshka Representation Learning nested embeddings.
Sub-millijoule retrieval tier. Embedder + GGUF embedder +
matryoshka truncation + nearest-neighbor picker + retrieval.
jouleclaw-lmm
Ported from pattern-lang/lmm (~1819 LOC, 6 files).
L3 tier — Large Multimodal Model. Vision + text + audio pipeline
on top of jouleclaw-prism's TernaryDecoder backbone. Real VL path
(LFM2.5-VL preprocessing via the `image` crate, JPEG/PNG only —
keeps the dep footprint minimal). The internal `prism::` references
rewritten to `jouleclaw_prism::` during port.
jouleclaw-ebm
Ported from pattern-lang/ebm (~1738 LOC, 7 files).
Energy-based constraint-satisfaction tier — SAT, n-queens, sudoku,
graph coloring, generic solver. Deterministic compute priced in
joules, sits at the L1/L3 boundary depending on the problem class.
Bare crate-name references in integration tests fixed:
`use liquid::...` → `use jouleclaw_liquid::...` (and same for the
other 4 ported tiers). Initial sed pass only caught `joule_*`
patterns; this second pass caught the bare-name ones.
gemma4 deferred. It depends on `joule-l1` and `pattern-core`
(OpenIE IP that doesn't port). Will land in Phase 6 once the
thin `LawfulTier` trait surface ships in jouleclaw-cascade or a
new jouleclaw-l1 crate so consumers can plug their own lawful
lexicon in.
Phase 5 next: omni-modal generation — diffusion (40+ samplers from
efficient-genai's ALGORITHM_CATALOG), Stable Diffusion / SDXL / SD3 /
Flux paths, MusicGen, Whisper, Gaussian3D, video synthesis,
cross-modal fusion. Source codebase: efficient-genai.
Co-Authored-By: Claude Opus 4.7 <[email protected]>
05dd2e0c4c