Record that incremental transcription was not built, and why

[bojieli]

Why

docs/INCREMENTAL.md currently recommends engaging the live segmenter at 45 s and cutting every 30 s, taking it from 8% to 20% of dictations. That was written before the measurements that mattered, and it is now wrong in the direction that would cost users quality.

What the measurements said

Segmenting does not recover content. Ten recordings, 136–312 s, whole vs Silero-policy segments, no screen context:

	whole 3.5	segmented 3.5	whole 3.6
content, chars/s	7.79	7.64	7.77
divergence from reference	0.120	0.129	—
closer to reference	7/10	3/10	—

segmenting recovered >15% more text on 0/10 recordings. Six Mandarin recordings were consistently worse.

The maintainer reviewed the disagreements by ear and preferred whole-file. 258 aligned differences judged against the audio — the only ground truth in the investigation.

So blanket segmentation trades a frequent small quality loss for protection against a rare large one, and the rare one now has a direct fix in the model (#8). Bad trade.

What it actually found

gemini-3.5-flash silently truncates long recordings — 6 runs in 10 on one 90-second clip, returning fluent text missing two thirds of what was said. Nothing noticed, because HallucinationGuard's rate ceiling is a maximum and 1.1 chars/second passes it from below.

The near-miss suite is structurally blind to this. Its cases are short clips; truncation only happens on long audio. It reported 3.6 ahead by 4 points — inside its own noise — when the real margin was 0 truncations against 6 in 10. That blind spot is the most reusable lesson here and it now leads the document: ask what a suite cannot see before citing it.

What changed in the doc

• Opens with what was proposed, what survived, and the bug — instead of a design pitch for something that should not be built.
• The body/tail split, rolling context, the whole policy and the aggressive live policy are removed as recommendations and kept only as recorded rejections.
• The scorer fix and the Silero boundary source are re-stated against the policy that actually ships, not the proposed one, with the Silero trade (better seams, longer final chunk) and the reason it is not yet shipped.
• Four more entries in Mistakes already made against this code, each of which cost real time here: trusting a suite outside what it measures, a Latin-only tokenizer on Mandarin, a synthetic fixture with less pause than real speech, and C# record structs ignoring primary-constructor defaults.

Documentation only.

🤖 Generated with Claude Code

https://claude.ai/code/session_015XKX6zEAGiZFE5wAxJgYbV