PRD — J/token instrumentation via powermetrics

Goal

Measure energy-per-generated-token across the model zoo and across training runs. Surface a j_per_token column on the SLM leaderboard and a per-epoch number on training runs. Answers “what does it cost in laptop battery to use / train this specialist?” — a question real users actually ask.

Why now

The hardware bench framework is the right home for this — it’s one more column next to TTFT / ITL / tok/s / RSS. The bench harness doesn’t measure it.
Long-running training on battery is increasingly a thing; even a rough J/step number lets users budget overnight training honestly.
powermetrics needs sudo, so this PRD documents the auth shim alongside the measurement.

Scope — in

scripts/bench_energy.py — wraps either bench_decode.py (decode) or a tinygpt train --steps N (training) in a powermetrics sidecar that samples package power at 1 Hz, integrates over the window, and divides by tokens generated (or steps run).
Output JSON gains joules_total, joules_per_token, joules_per_step.
Optional integration with the leaderboard aggregator — new optional energy_json: field on the manifest row.
One-time sudo helper: scripts/setup_powermetrics_sudoers.sh walks the user through adding a passwordless sudo rule for /usr/bin/powermetrics only.

Scope — out

Cross-device energy accounting (CPU vs GPU vs ANE breakdown). Package power is the V1 number; per-engine is V2.
Training-loss-per-joule optimization passes — V1 measures, doesn’t optimize.

Files to touch

File	Change
`scripts/bench_energy.py`	new — sidecar wrapper
`scripts/setup_powermetrics_sudoers.sh`	new — one-time auth helper
`scripts/build_slm_leaderboard.py`	optional `energy_json` field in manifest reader
`docs/research/mac_decode_baseline_m5pro.md`	add energy column to Run 5/6
`docs/PLAN.md`	B9 ⬜ → ✅ on ship

Acceptance criteria

scripts/bench_energy.py --decode --model google/gemma-3-12b --n 20 produces a joules_per_token number ± 10% on rerun.
scripts/bench_energy.py --train --steps 100 --base shakespeare.tinygpt produces a joules_per_step number.
Without sudo set up, the script prints a clear “run setup_powermetrics_sudoers.sh once” error.
Energy row appears on the leaderboard when present.

Reference patterns

scripts/duty-cycle-throttle.sh — existing powermetrics invocation; same auth pattern.
C5 decode-jitter-thermal — companion (thermal + energy are often co-measured).

Open questions

Whether to integrate into tinygpt bench itself (Swift-side) vs keep it a script wrapper. Recommendation: Python wrapper — powermetrics has too much shell glue to comfortably go in Swift.