Engineering artifact

WriteAmp Bench

A small, repeatable test for the three local tiers WriteAmp ships: mini, midi, and max. Each published version is a full artifact. Every input. Every config. Every percentile. And for every case, what happened.

Status

v1 is published. The immutable dataset lives under /benchmarks/v1/. Aggregate metrics, per-case rows, and the exact run manifest. We never change a published version. New versions get new folders (like /benchmarks/v2/), so links you cite stay stable. Quick look:

Tierp50 totalp95 totalPrecision when shown
mini51 ms110 ms13.0%
midi92 ms109 ms17.4%
max123 ms147 ms17.4%

Current-state reference numbers on 31 cases (v1 run, full detail and honest caveats). Confidence-threshold tuning is in progress; a tuned run ships as v2.

What a published version contains

  • aggregate.json: p50/p95 for generation and end-to-end, plus a histogram of why we held back, per-tag outcomes, and the scoring names (correctInsert, correctSuppression, etc.).
  • rows.jsonl, one row per case with the model's top candidate, the suppression reason, and per-phase latency.
  • manifest.json, exact configuration: model/profile/runtime hashes, warm-up, repetitions, decoding flags, prompt and completion budgets, hardware class, macOS version, build mode.
  • corpus.jsonl, the public cases themselves, with per-case provenance and explicit license.
  • methodology, scoring semantics, suppression-reason semantics, and the limitations of mechanical variants.

What this benchmark is not

  • This isn’t “accuracy.” correctInsert plus correctSuppression just track where this model, on this prompt, decided to show or hold back. That’s what you see at the cursor. It’s not a stand-in for a person judging the writing.
  • This isn’t a cross-model ranking. The cases are tuned to WriteAmp’s three rungs. Want to rank other models? You’d need a different set of cases.
  • This isn’t proof of Apple Intelligence speed. That engine lives inside macOS and we bench it separately. Think of it as a canary, report-only.

Questions about the methodology: [email protected].