33f0b89124
The owner's decision, applied: a vendor energy counter (RAPL, NVML, IOReport) is a real reading, and what separates one such reading from another is provenance, not accuracy. This crate gave RAPL and NVML `HwShunt` and gave the IOReport counter `ModelBased`, on the reasoning that Apple's Energy Model is a vendor model. That reasoning grades accuracy. What the counter could not establish was provenance: on macOS 27.0.1 its CPU rails are frozen for an unprivileged reader, and that is now caught by the liveness probe. A counter that has passed it is as much a reading as RAPL is. The grade. `AppleIoReportCounter` declares `HwShunt`, and stamps it on every reading through one constant, `READING_GRADE`, shared by `provenance()` and `read()`. It earns the grade by being live: the counter exists to read only where `admits()` let it through (CPU rail `Live`), and a declined counter produces no reading to carry any grade, so a frozen host still meters through the load-average estimator at `Estimator`. The rule, written once and shared (jouleclaw SPEC §1.3, §9.4; JCR-1 §2.3; the crate docs; vendor-energy PROTOCOL/README): a vendor counter's reading MAY be graded as a hardware reading only while the counter is demonstrably live for the domain the figure claims; the implementation MUST have observed the counter covering that domain advance under load (for IOReport: the CPU rail itself, in at least 5 intervals, 4 of every 5, over at least 300 ms while the probing thread burns CPU). A counter present but not live MUST NOT produce a hardware-graded figure; the implementation falls back to its estimator or reports the work unmetered, and SHOULD record why. A figure MUST NOT claim coverage of a domain its counter does not read; where the format has a coverage field, such a figure is Partial. No partial reading from the GPU rail. Where the CPU rail is frozen but `GPU Energy` still advances (the M5 Max), the counter stays declined. The GPU rail is machine-wide: every GPU client on the host moves it, display compositing included, with nothing of the metered work running. It does not cover part of the work; it covers other work, so no coverage field makes it honest. The rationale is in the module docs and SPEC §1.3. Docs and data that said otherwise: - SPEC §1.3 table and floors, §9.4; JCR-1 §2.3 and the crate README (toy vectors become measurements only off counters that were live). - JCR-1 vector 6's description said "measured on Apple IOReport". Its joules are toy values like every vector's. Only the `description` changed; it sits outside the signed core, which `verify.py` builds from `receipt` alone. Both verifiers pass 10/10 after the edit: `verify.py` (cbor2 6.1.4, cryptography 50.0.2, exit 0) and the independent `verify.go` (exit 0). `git diff` of the file is that one line; a field-by-field JSON diff finds only (6, description). The conformance README's slot 6 line follows. - vendor-energy conformance: README row 27 names `AppleIoReportCounter` and the `HwShunt` grade, with the liveness condition a filled fixture must meet; `apple-silicon-ioreport.json` is an unsigned, empty stub (`signed_by` null, every workload not run, no hash of it anywhere), so its `impl` and `provenance` are corrected in place. - MEASURED_ENERGY.md: a dated note that the 74.1 J / 25.5 W row was committed 2026-06-12, predates macOS 27, and recorded neither chip nor OS. The "reads CPU/GPU/ANE/DRAM with no sudo" claim is replaced by what macOS 27.0.1 does: the counter opens without sudo, but the CPU rails are frozen (all 645 rails on the M3 Ultra), so those hosts are estimated. - local-index / rerank tier SHOULDs, the CLI's `meter` help, meter.rs, apple.rs, energy_probe, Cargo feature comments, harness-econ docs. - l3-jouleclaw README: it claimed `HwShunt` on RAPL/NVML hosts, but the binding builds jouleclaw-energy with `ioreport` only, so off Apple Silicon the receipt is the token estimate. It now says what happens on each host. - mesh-schema's `HwShunt` definition includes vendor counters subject to liveness; mesh-meter, the mesh README and mesh-engine-gguf no longer say real silicon alone yields `HwShunt`. - The Live-branch assertions in mesh-meter, l3-jouleclaw and jouledesk-github expected `ModelBased` for IOReport; they expect `HwShunt`. jouledesk-github's fixtures that paired `apple-ioreport` with `ModelBased` now pair it with `HwShunt`, and keep a `ModelBased` case under a neutral meter name. Tests. `a_live_counters_readings_are_graded_as_hardware_readings` builds a counter on `Live` directly. That opens the real all-channel subscription, which opens here even though the gate declines this host. It reads it twice and asserts each reading is `HwShunt` over `SocTotal`. On this M5 Max (macOS 27.0.1): "an admitted IOReport counter: opens=true, 2 readings, each graded HwShunt", so the read path is covered on a frozen host. With `--features "ioreport rapl"`, `a_live_ioreport_reading_has_the_grade_of_a_rapl_reading` compares the grade with `RaplCounter`'s own. The used/declined tests assert `HwShunt` in their live branch. This host still takes the declined branch: "ioreport declined (CpuRailFrozen): 154 ms of CPU work metered by `apple-silicon` as 6200000 µJ, provenance estimator". Run on this M5 Max: jouleclaw-energy 24+1+1 (default), 76 lib + 1 + 1 (`ioreport`), 79 lib (`ioreport rapl`); jouleclaw-jcr1 50; harness-econ 14; local-index 41; rerank 37; cli 10; mesh-meter (`hw-meter-apple`) 10; mesh-engine-gguf 1; l3-jouleclaw 2 + 4 (2 ignored, need a model); jouledesk-github 12. All pass. Falsified, each mutant restored with `git checkout` and the tree confirmed clean: - `READING_GRADE` set back to `ModelBased`: the new test fails at the declared grade (left: ModelBased, right: HwShunt), 75 passed / 1 failed. - Only `read()`'s stamp set back to `ModelBased` (the declared grade left at `HwShunt`): the new test fails on the reading itself, "a live counter's reading is not graded as a hardware reading" (ModelBased, SocTotal) vs (HwShunt, SocTotal), 75 / 1. The read path is what the test pins. Co-Authored-By: Claude Opus 5.5 <[email protected]>
70 lines
3.3 KiB
Markdown
70 lines
3.3 KiB
Markdown
# L3 binding — JouleClaw model tier → Joule Code cascade
|
|
|
|
A consumer **binding example**: how to wire JouleClaw's **L3 (Model)** tier into
|
|
the Joule Code cascade. It is a standalone workspace, so the standard's reference
|
|
library keeps no JouleClaw dependency.
|
|
|
|
A coding request that misses the cheaper tiers (L0 cache, L1 lawful, L2 embed) is
|
|
generated by a local model at L3, producing a content-addressed `CodingOp` and a
|
|
joules-class `Receipt`.
|
|
|
|
## Three backends, same slot
|
|
|
|
All implement JouleClaw's `jouleclaw_llm_cheap::LlmBackend`, so they drop into
|
|
the same cascade stage:
|
|
|
|
- **`EchoBackend`** (from JouleClaw) — deterministic, model-free; used by the
|
|
unit tests so the wiring is verifiable with no model present.
|
|
- **`JouleclawRuntimeBackend`** (here) — runs the model **in-process on
|
|
JouleClaw's own from-scratch GGUF tensor runtime** (`jouleclaw-runtime`): no
|
|
external binary, no FFI. Tokenize → build the decode-step graph → compile
|
|
against the registered kernels (`jouleclaw-backend-reference` /
|
|
`-apple`) → autoregressive decode → detokenize, all in Rust the family wrote.
|
|
This is the **default** backend for the `run` binary.
|
|
- **`LlamaCompletionBackend`** (here) — drives a real on-device GGUF via
|
|
`llama-completion` (llama.cpp). Set `JOULE_CODE_L3_ENGINE=llama` to use it.
|
|
|
|
Either local path: no datacenter, no network. In production the cheapest capable
|
|
backend is chosen by `jouleclaw_dispatch::ModelBank`'s cost table.
|
|
|
|
## Run it against a local model
|
|
|
|
```bash
|
|
# Default: in-process on jouleclaw-runtime (the from-scratch tensor engine).
|
|
cargo run --release --bin run -- /path/to/model.gguf "Write a Rust function that reverses a string."
|
|
# or: JOULE_CODE_L3_MODEL=/path/to/model.gguf cargo run --release --bin run
|
|
|
|
# Use the llama.cpp shell-out backend instead:
|
|
JOULE_CODE_L3_ENGINE=llama cargo run --release --bin run -- /path/to/model.gguf "..."
|
|
```
|
|
|
|
The `run` binary generates with the local model, content-addresses the result as
|
|
a `CodingOp` (tier L3), and puts the **host** energy spent during the inference on
|
|
the receipt, tagged with the provenance of whatever metered it:
|
|
|
|
- **Apple Silicon whose IOReport CPU energy rail is live** — JouleClaw's IOReport
|
|
counter, a live vendor energy counter: a real local measurement, tagged
|
|
`HwShunt` (with `measured_uj` set).
|
|
- **Apple Silicon whose CPU rail is not live** — macOS 27.0.1 for an unprivileged
|
|
process, observed on an M5 Max and an M3 Ultra — JouleClaw's load-average
|
|
estimator, tagged `Estimator`. Not a measurement.
|
|
- **Anywhere else**, Linux and Windows included — the token-count estimate,
|
|
tagged `Estimator`. This binding builds `jouleclaw-energy` with the `ioreport`
|
|
feature only, so it reads no RAPL or NVML counter even where one exists.
|
|
|
|
The printout names the meter that ran and the IOReport status, and says whether
|
|
the figure was measured or estimated.
|
|
|
|
Verified live: the llama.cpp path against a local 350M GGUF, and the in-process
|
|
from-scratch runtime against `ternary-bonsai-1.7b-q2_0` (qwen3, Q2_0) — real text
|
|
generated, a real `op_id`, the cascade closing at L3.
|
|
|
|
## Tests
|
|
|
|
```bash
|
|
cargo test # EchoBackend path — no model required
|
|
|
|
# In-process from-scratch runtime against a real GGUF (opt-in):
|
|
JOULE_BONSAI=/path/to/ternary-bonsai-1.7b-q2_0.gguf \
|
|
cargo test --release --test runtime_l3 -- --ignored --nocapture
|
|
```
|