What the polish pass changes, and what it never touches: the 31-case golden corpus

After the speech model returns text, and before Lven types it into the focused app, a single-pass token-stream engine (LvenStructuralPolisher, shipped as C++ on Windows and Linux and as Kotlin on Android, byte-identical contracts held by a shared golden corpus) applies seven rules in a fixed order: hyphen-fragment pruning; prefix-stutter pruning; n-gram repetition collapse for n = 1, 2, 3; filler purge (uh, um, eh, er, ah, uh-huh); contraction, proper-noun and phonetic-spelling restoration from a fixed table; sentence casing; terminal punctuation, where the last sentence decides ? versus period.

It is not a language model. It cannot rephrase, restructure, change tense, match tone to an app, or decide what you meant. It has no network access. Given the same transcript it produces the same text every time, on every platform.

The corpus is the polisher's own golden regression test — the file CI runs on every commit and regenerates for the Swift ports, so that all implementations are held to one contract. We compiled tests/polisher_golden_test.cpp against the shipped Lven-Windows/src/instant/lven_polish.cpp with a stock LLVM MinGW g++ (-std=c++17) on 2026-08-15 and ran the binary; it asserts every expected output and returns non-zero on any mismatch. Word counts and change classes were then computed over the 31 pairs with a token diff (case- and apostrophe-insensitive alignment, then classification of each difference).

Every word that survives is a word the speaker said. The only substitution in the corpus is a spelling restoration from a fixed table; nothing is paraphrased and nothing is added.

Note what is preserved: 'you know like you know' in the first case stays, because it is not in the filler list and the engine does not judge whether a phrase is worth keeping. 'had had' and 'that that' stay, because the collapse pass guards legitimate doubles.

A generative cleanup layer optimises for text that reads well. That is a real product goal and some people want it. Its cost is that every sentence has to be re-read for meaning, because the layer is allowed to decide what you meant; in July 2026 one cloud dictation vendor traced accuracy complaints to its cleanup default having changed words users had not asked it to touch. A deterministic pass optimises for the text being what you said. Its cost is that it will not turn a rambling sentence into a clean one. We chose the second, and this corpus is what that choice looks like in practice.

From a checkout of the Lven Instant source, at the repository root: g++ -std=c++17 -I Lven-Windows/src/instant tests/polisher_golden_test.cpp Lven-Windows/src/instant/lven_polish.cpp -o polish_golden && ./polish_golden. The binary prints the pass count and exits non-zero on any mismatch. The Kotlin and Swift ports are held to the same corpus through tests/polisher_golden_cases.json.