Honest labelling

wattop marks every figure that is an estimate, and never shows a zero where it has no data.

• Claude context-fill is an estimate (dashed bar edge); Codex’s is exact from rate_limits (solid edge).
• Costs are estimates from the pricing table and ignore subscription plans (Claude Pro/Max, ChatGPT/Codex) entirely.
• An unpriced model renders $—, never $0.00. Those mean different things.
• Per-session GPU is measured as ms/sec; the derived percent is a rescale and never the default sort key.
• Bandwidth renders — only where no source exists; a channel that resolves and reads zero renders 0.0. On this chip no DRAM byte counter resolves at all, so DRAM shows one power-derived total, marked as an estimate (Total ~14.3 GB/s), with no read/write split.
• The temperature row shows the three semantic sensors (CPU, GPU, SOC) and omits any that did not resolve, rather than every raw SMC key the chip exposes.
• $/hr counts only spend wattop watched happen. Cost already on disk when it starts is baselined, so a fresh launch reads $0.00/hr rather than extrapolating hours of history into a rate.
• Token rates are 60-second transcript averages, not instantaneous model generation speed. Unknown usage is —; observed inactivity is 0.0.

“This chip” is the M5 Max wattop has been tested on. wattop doctor shows which channels resolve on yours.

A nonzero live $/hr has not been observed end to end

From limitations: “A nonzero live burn has still not been observed end to end: across two 30-second and one 3-minute --json capture, every watched transcript was static after the launch backfill, so total_cost never moved. The positive path rests on internal/pricing’s unit tests ($3.60/hr and $60.00/hr exactly, against controlled timestamps), not on live data.”

See limitations for the full list with evidence, and the manual QA checklist this release was verified against.

Self-CPU overhead

Measured over a 75-second --json run on an M5 Max watching 12 Claude/Codex sessions with 1,293 subagents: Snapshot.self_cpu_pct settles to mean 5.19% (max 6.94%) in steady state. The first ~5 samples after startup read 95-304% while wattop catches up on existing transcripts (and, on M5-class chips, calibrates DRAM bandwidth); see limitations and docs/manual-qa.md item 11. A monitor that distorts what it measures must be accountable for its own overhead; this is that accounting.