8da045d94c
Codebase-state numbers (crate count, test count, LOC, "40+ samplers")
drift on every code change and create permanent sync debt with the
website and downstream READMEs. Replaced with shape descriptions
("multi-crate Rust workspace", "the diffusion sampler catalog",
"covered by a test suite"). Identity-class facts (spec versions,
licenses, algorithm names) preserved.
95 lines
4.6 KiB
Markdown
95 lines
4.6 KiB
Markdown
# JouleClaw roadmap
|
|
|
|
Where the standard is going after v0.1.0 (currently committed to
|
|
`main`). The reference implementation in [`jouleclaw-rs/`](jouleclaw-rs/)
|
|
is a multi-crate Rust workspace; all tests pass.
|
|
|
|
## Deferred work, by phase
|
|
|
|
### Near-term (v0.2)
|
|
|
|
- **`gemma4` port.** Pattern-lang's `gemma4` tier registers itself
|
|
against `joule-l1`'s lawful lexicon and `pattern-core`'s synthesizer
|
|
for prompt routing. With Phase 6's `jouleclaw-cascade::lawful`
|
|
trait surface now landed, the port is unblocked — but the
|
|
`pattern-core` synthesizer routing path needs the consumer (or a
|
|
thin `pattern-core`-shaped consumer crate inside JouleClaw) to plug
|
|
through `LawfulRegistry`. Tracked.
|
|
- **`jouleclaw-edge-server`.** Pattern-lang's `joule-edge/src/server.rs`
|
|
was an HTTP wrapper around the demo runtime. The library form
|
|
(`jouleclaw-edge`) shipped; the server form is a thin axum wrapper
|
|
that can land any time. Out-of-scope for v0.1 because the demo
|
|
binary covers the same surface for development.
|
|
- **Conformance vector population.** `conformance/v1/` ships the
|
|
three structural vectors (cache-hit, breaker-trip,
|
|
joule-mcp-CBOR-negotiation). The remaining seven vectors named in
|
|
`conformance/v1/README.md`'s coverage matrix
|
|
(lawful-arithmetic, embed-hybrid-search, fresh-retrieval-with-
|
|
provenance, model-local-ssm, model-local-ternary, model-local-
|
|
diffusion-image, wire-frontier-rpc) await the canonical receipts
|
|
that fall out of running the reference impl with the standard
|
|
reference-hardware (Apple M5 Max baseline). Each is a few hundred
|
|
lines of JSON + a `.jc.toml` sidecar.
|
|
- **`SHA256SUMS` + release signature.** The v1.0.0 conformance
|
|
vector corpus signs at release time. Tracked alongside the v1.0.0
|
|
release tag.
|
|
|
|
### Medium-term (v0.3)
|
|
|
|
- **Split `jouleclaw-omni`.** The monolithic crate ported wholesale
|
|
from `efficient-genai` in Phase 5 preserves the donor's internal
|
|
layout (`core / hal / inference / modalities / orchestrator /
|
|
quality / runtime / server / tensor / weight_store`). Splitting
|
|
into separate crates is multi-week surgery because the internal
|
|
cross-references (`crate::hal::*`, `crate::tensor::*`,
|
|
`crate::core::*`) need to be rewired into proper Cargo path deps.
|
|
The plan when this lands:
|
|
- `jouleclaw-hal` — `Device / Kernel / Memory / Capabilities`
|
|
trait surface
|
|
- `jouleclaw-tensor` — the tensor module
|
|
- `jouleclaw-diffusion` — the diffusion sampler catalog
|
|
- `jouleclaw-musicgen`, `jouleclaw-whisper`, `jouleclaw-gaussian3d`,
|
|
`jouleclaw-video`, `jouleclaw-fusion` — per-engine crates
|
|
- `jouleclaw-modality-{text,image,audio,video,3d}` — modality
|
|
surfaces
|
|
- `jouleclaw-omni-server` — axum server, optional
|
|
- `jouleclaw-omni` becomes a meta-crate that re-exports the above
|
|
behind a `legacy` feature for downstream consumers who don't
|
|
want to re-find all the names.
|
|
|
|
Until then, `jouleclaw-omni` ships as one crate with feature gates
|
|
(`cuda`, `metal`, `rocm`, `vulkan`, `safetensors`, `jit`, `simd`,
|
|
`rayon`, `server`).
|
|
|
|
- **Standalone GPU-backend crates.** `jouleclaw-omni::hal::*` already
|
|
has the CUDA / Metal / ROCm / Vulkan paths feature-gated. Extracting
|
|
them as standalone `jouleclaw-backend-cuda`, `-rocm`, `-vulkan`
|
|
crates is a sub-task of the omni split above and lands in the same
|
|
v0.3 work. Until then, downstream consumers enable the feature on
|
|
`jouleclaw-omni` and the backend is available behind
|
|
`jouleclaw_omni::hal::cuda::*` etc.
|
|
|
|
### Long-term (v1.0)
|
|
|
|
- **Multi-machine cascade.** The current cascade assumes a single
|
|
box. v1.0 will define how a JouleClaw deployment routes work across
|
|
a fleet — content-addressed cache sharing, per-node energy budgets,
|
|
cross-node receipt aggregation. Out-of-scope for v0.x.
|
|
- **Hardware-attested measurement.** Some platforms (Apple Secure
|
|
Enclave, NVIDIA Hopper Confidential Compute) can attest energy
|
|
measurements cryptographically. Receipt schema v2 will carry the
|
|
attestation envelope as an optional field.
|
|
- **A second reference implementation.** A second language (Go or
|
|
Zig) reference impl that round-trips the same conformance vectors
|
|
is the long-term proof that the standard is implementable cleanly.
|
|
|
|
## Not on the roadmap
|
|
|
|
- **Training framework.** JouleClaw is a runtime + harness. Training
|
|
is the model author's concern.
|
|
- **A specific signing key topology.** Bring your own. Smart Byte's
|
|
KERI-based AID rotation is the recommended path; the standard
|
|
doesn't require it.
|
|
- **A specific retrieval API.** Brave / Tavily / Exa / Serper all
|
|
plug in through `jouleclaw-fresh::SearchProvider`. The standard
|
|
doesn't endorse one.
|