| Filename | Latest commit message | Latest commit date |
|---|---|---|
950634f200
All checks were successful
standards workspaces build + test / which workspaces this change touches (pull_request) Successful in 7s
jouleclaw fabrics / cuda (linux, RTX 4050 + Vulkan loader) (pull_request) Successful in 16s
jouleclaw fabrics / vulkan (linux, RTX 4050 + Vulkan loader) (pull_request) Successful in 20s
standards workspaces build + test / mesh-rs (mesh/mesh-rs, -p mesh-engine-gguf -p mesh-openai --features mesh-engine-gguf/gguf,mesh-openai/server, mesh-rs, linux-arm64) (pull_request) Successful in 1m39s
repository checks / lockfiles, tests in CI, workflow hygiene, spec and section references, public text (pull_request) Successful in 1m58s
standards workspaces build + test / jouleclaw-rs (-nodebug, jouleclaw/jouleclaw-rs, -p jouleclaw-energy --features rapl,nvml, jouleclaw-rs, linux-arm64, -C target-feature=+fp16) (pull_request) Successful in 4m46s
jouleclaw public surface must stay documented / Public-surface audit (jouleclaw-rs) (pull_request) Successful in 1m2s
jouleclaw fabrics / metal (macOS, Apple GPU) (pull_request) Successful in 1m12s
standards workspaces build + test / macOS energy backends (pull_request) Successful in 9m13s
standards workspaces build + test / jouledesk-rs (jouledesk/jouledesk-rs, jouledesk-rs, macos) (pull_request) Successful in 2m54s
`DecodeBuffers::new` built the KV cache with
let max_seq = 4096usize;
let num_attn_layers = 6; // Nemotron: 6 attention layers
while `encode_decode_step` indexes it with the model's real attention-layer
ordinal and the real position:
let attn_idx = layers.iter().take(layer.idx + 1)
.filter(|l| l.layer_type == LayerType::Attention).count() - 1;
let dst_offset = _position as u64 * kv_row_f16;
Two consequences, neither of which reported anything:
- A model with a seventh attention layer got `attn_idx == 6`, the
`if attn_idx < bufs.kv_k_cache.len()` guard was false, and that layer's
K and V were simply never written. Attention then read a zero-filled
cache: wrong output, no error.
- `_position` is never bounded anywhere in the function, so a position at
or past 4096 gave `set_buffer` an offset past the end of the buffer.
Both lengths now come from the model. `init` resolves the layer types and
passes the attention-layer count; `max_seq` comes from
`config.max_seq_len`, falling back only when it is 0, which would
otherwise allocate a zero-length cache that fails every write. The
capacity is recorded on `DecodeBuffers` so the decode path can check
rather than assume, and `encode_decode_step` now checks both once, up
front, where an error can propagate -- the per-site `if idx < len` guards
are deep inside nested closures where a `return Err` cannot.
tests/ggml_kv_cache.rs pins the two lengths: one buffer per attention
layer for counts of 1, 6, 7 and 11 (6 being the literal that hid this),
buffer length following max_seq_len for 512, 4096 and 8192, and a
zero-length config still producing a usable cache.
Scope: this path is behind the GGML_GRAPH environment variable, so it is
opt-in rather than the default decode path.
Still assumed, and worth separate attention: `GRAPH` and `LAYERS_CACHE`
are process-global `OnceLock`s, so a second model loaded in the same
process reuses the first one's buffers and layer pattern.
Co-Authored-By: Claude Opus 5 <[email protected]>
|
||
| .. | ||
| conformance/v1 | ||
| jouleclaw-rs | ||
| scales | ||
| vocab | ||
| LICENSE | ||
| LICENSE-SPEC | ||
| README.md | ||
| ROADMAP.md | ||
| SPEC.md | ||
JouleClaw — Energy-Optimized AI Runtime
The substrate for AI that costs Joules, not tokens.
Capability per Joule, not capability per parameter.
What JouleClaw is
JouleClaw is a pure-Rust, omni-modal AI runtime + harness + agentic layer governed by Transaction Science as an open standard (5th in the open-standards quartet, alongside OpenPay, Smart Byte, EOC, and WAI).
Every operation JouleClaw performs walks an explicit energy-tiered cascade, top to bottom, stopping at the first stage that closes the query:
| Tier | Name | Cost class | What runs here |
|---|---|---|---|
| L0 | Cache | picojoules | Content-addressed cache hits |
| L1 | Lawful | nanojoules | Deterministic, pre-compiled primitive (text/code only) |
| L2 | Embed | sub-millijoules | Matryoshka embeddings + nearest-neighbor + hybrid search |
| L3 | Model | joules | Local SSM / ternary / multimodal / diffusion / etc. |
| L4 | Wire | tens of joules | Remote frontier RPC (escape hatch only) |
L1 (Lawful) is only meaningful for text and code modalities. For image, audio, video and 3D generation, the cascade collapses to L0 / L2 / L3 / L4 — the energy savings come from caching plus dispatch to the cheapest model that meets the quality bar.
Why JouleClaw exists
The mainstream "agentic AI" ecosystem treats inference as the entry point. Every state shift, every tool call, every routing decision runs through a 70-billion-parameter stochastic engine, burning kilojoules to do work that deterministic compute or a hot cache could resolve for microjoules.
JouleClaw inverts the default. Inference is the last resort, not the first. Frozen model weights are less trustworthy than current world state; fresh, provenance-stamped retrieval beats synthesis from training cuts. Every operation is metered in microjoules where the hardware permits it, and in millijoules where it doesn't — the spec is honest about the floor on every platform.
The thermodynamic circuit breaker is the safety mechanism. Set an energy budget; when measured consumption exceeds the budget, the breaker trips and the operation halts. Hallucination loops cannot survive an honest joule ledger.
What's here
| Path | Contents |
|---|---|
SPEC.md |
JouleClaw v1 spec — cascade tiers, energy contract, receipt |
CHARTER.md |
Stewardship pattern (shared across open-standards) |
jouleclaw-rs/ |
Apache-2.0 Rust reference implementation |
spec/ |
Tier specifications, energy provenance, receipt format |
conformance/v1/ |
Signed wire vectors for cross-implementation verification |
Reference implementation
jouleclaw-rs/ is the Apache-2.0 reference implementation in Rust. It
runs on Apple Silicon (M3/M4/M5 with UMA + Metal 4 Tensor APIs), AMD Strix
Halo / Ryzen AI Max, Intel Lunar Lake, NVIDIA Jetson, and discrete
NVIDIA / AMD GPUs.
It is omni-modal: text, code, vision, audio, video, image diffusion, 3D Gaussian splatting, and cross-modal fusion all dispatch through the same energy-tiered cascade.
The MCP tool surface is a JCP consumer: jouleclaw-jcp wraps any
jouleclaw-mcp metered tool so every dispatch is decided by the JCP runtime
against a signed, metered Grant before a byte reaches the tool — the tool's
identity is the content-addressed resource, the call's energy estimate is
debited from the grant's joule budget, and every decision (allow and deny
alike) seals as a JCR-1 receipt. An injected instruction can name an ungranted
tool; it cannot make the gate dispatch it.
Status
v0.1.0 reference implementation in active development. The spec is draft status. Cross-implementation conformance vectors will land with v1.0.0.
License
- Code: Apache-2.0 (see
LICENSE) - Spec text: CC-BY-4.0 (see
LICENSE-SPEC)