Polish: model picker (Apple Intelligence, S1-mini), spoken punctuation, filler fix

[dinooo13]

Summary

A model picker for the experimental polish, the two text bugs the model evaluation turned up, and the parts of the cleanup that code does better than a model.

• Model picker under "Polish dictations" (Processing → Experimental): Apple Intelligence (default, unchanged), S1-mini by Superwhisper at full precision (f16, 1.5 GB) or 8-bit (Q8_0, 805 MB). S1-mini is Qwen3-0.6B fine-tuned for dictation cleanup and runs through llama.cpp on the GPU, leaving the Neural Engine to Parakeet.
  • PolishRouter is the coordinator's one refiner and forwards to the chosen model.
  • ModelFiles downloads a file once from Hugging Face, only while polish is on and that model is picked. Each file is pinned to a commit and moved into place only after its SHA-256 matches. The picker shows progress, and a failed download offers Try Again.
  • S1MiniPolisher uses the prompt the model was trained on and decodes the fixed prefix once at load. It chunks long dictations to fit its 2,048-token context and keeps the 8 s cap and the paste-as-dictated fallback. The model loads at the first key-down and is freed on a switch or when polish goes off.
  • llama.cpp's Metal setup is warmed as soon as S1-mini is chosen. The first run after an install compiles shaders for about 7 s on an M1; macOS caches them after that.
  • llama.cpp is its prebuilt XCFramework, pinned by release and checksum. MLX Swift needs Xcode to build its Metal shaders, and this project has no Xcode project. bundle.sh embeds the framework, drops the Intel slice and signs it with the app's identity.
• Spoken punctuation is a new processor, last in the pipeline. It turns "comma", "question mark", "new paragraph", "Fragezeichen", "signo de interrogación" and similar into the marks, absorbing the punctuation Parakeet adds around them ("verschieben? Fragezeichen." → "verschieben?").
  • Only phrases that are never ordinary words are taken: "period", "Punkt" and "punto" stay words, and Spanish "coma" needs Spanish language evidence. A spoken comma between digits is the decimal comma.
  • Parakeet already writes numbers as digits in English and German, so no number converter was needed.
• Filler fix: a filler that opens a later sentence now hands on its capital. "…geschrieben. Ähm, beim Timeout…" used to paste as "…geschrieben. beim Timeout…".
• CLI: pladder-cli polish --model …, and pladder-cli polish-set docs/polish-set.json --model …, which runs a model over 42 dictations in English, German and Spanish and prints each answer, word error per language, exact matches and timings.
• Docs: CLAUDE.md, BENCHMARKS.md and the README. The README still described polish as a second hotkey.

Polish models (M1 MacBook Air 16 GB, macOS 27.0)

From pladder-cli polish-set, after the app's processors, warm, greedy:

Model	Word error (en / de / es)	Exact of 42	Polish median	p90
Apple Intelligence	0.108 (0.037 / 0.152 / 0.135)	18	1.66 s	2.07 s
S1-mini, full precision	0.071 (0.021 / 0.115 / 0.078)	22	0.49 s	0.86 s
S1-mini, 8-bit	0.091 (0.008 / 0.115 / 0.149)	21	0.34 s	0.61 s

A cold S1-mini start (load plus polish) is 1.1 s at 8-bit and 1.7 s at full precision. The load starts at key-down, while the user speaks.

What the numbers hide:

• Apple's model: translated two dictations into English, turned "halb acht" into "8:00", and never wrote a multi-line list.
• S1-mini: translated nothing and resolved the German and Spanish self-corrections, but:
  • it misreads the German number word "zweihundertfünfzig" as 150 (real dictation arrives as digits from Parakeet);
  • it garbles one German sentence around a code identifier;
  • it ignores "Nächster Punkt" and "Siguiente punto" as list cues.

LFM2.5-1.2B and Gemma 3 1B were tried with Pladder's prompt and dropped; details in BENCHMARKS.md.

Critical path

The engine and the coordinator's release path are unchanged, so the engine benchmark would not see this; I didn't rerun it. What does change is the processor pipeline (a new processor and the filler fix). Measured on the fixtures' text through the whole pipeline, median of 31, M1. "Typical" has a filler in every sentence; "worst" adds a spoken question mark to every sentence:

Fixture	main, typical	this branch, typical	main, worst	this branch, worst
10s	1.53 ms	1.97 ms	1.51 ms	1.62 ms
60s	1.71 ms	1.81 ms	1.73 ms	2.04 ms
2m	1.96 ms	2.14 ms	2.00 ms	2.61 ms
10m	3.77 ms	4.83 ms	4.13 ms	6.85 ms

At most 3 ms at 10 minutes, against a ~230 ms release-to-paste median. The 10 s difference is noise; the ~1.5 ms floor is the language recogniser the filler remover already calls. With polish on, the polish stage itself drops from about 1.7 s to 0.3–0.5 s with S1-mini.

Testing

• swift test: all targets pass (338 + 21 + 16 + 10 + 6). New tests cover spoken punctuation, the filler fix, S1-mini's prompt format and chunking, model-file mapping and pinning, the download's checksum and failure paths (from a file URL), the polisher without its file, and the setting's persistence and fallback.
• pladder-cli polish-set on all three models, which exercises the download, verification, load and generation end to end.
• bundle.sh with a certificate and ad-hoc. Checked that an ad-hoc hardened-runtime binary refuses the framework ("different Team IDs"), hence no hardened runtime for ad-hoc builds.
• A test copy with its own settings, launched from dist/, showed the new processor row and the picker. I did not dictate through the app with S1-mini selected; that end-to-end check is yours.

Notes

• License: S1-mini is Apache 2.0 with a naming clause, so it appears as "S1-mini" by "Superwhisper". The 8-bit file is mradermacher's quantisation of the same release, because Superwhisper publishes only f16 and 4-bit.
• Privacy: the Hugging Face exception now covers a polish model the user explicitly picks. Nothing downloads unless polish is on and S1-mini is chosen.
• Possible follow-up: someone has published a German fine-tune, Joni000000000/s1-mini-de-v3. It could be judged with polish-set before adding it as a picker entry.

🤖 Generated with Claude Code