The ASR flagged its own doubt. The subtitle editor rendered every word the same.

Start with the arithmetic, because it settles most of the argument. The formula in the write-up divides substitutions plus deletions plus insertions by reference words ^[3], which means every entry in that numerator weighs exactly one, whichever word it lands on ^[4]. Take the author's own example of a single wrong word in a hundred ^[15]: that is a 1 percent error rate, a 99 percent word accuracy figure ^[5]. The same 1 percent buys you a harmless "and" turned into "an" or a dropped "not" that reverses the sentence ^[6]. The metric counts errors. It does not price them ^[7].

The more interesting part is where the information went missing. The recogniser had already emitted per-word timing and confidence, and the renderer flattened all of it into uniform text ^[2], so the token reading "gp" sat in the same weight and colour as everything it was surrounded by ^[1]. Nothing was absent from the model output. The editor declined to show it.

Two defaults in the author's pipeline are worth copying. A word arriving with no usable raw score is normalised to 0, the lowest review score, so it lands in the queue instead of letting silence read as certainty ^[11]. Confidence stays optional on the editor word, with a source field marking asr against manual, so a human-typed word never inherits a score the model never produced ^[12]. Malformed responses are rejected at the boundary rather than patched with invented fields further in ^[13].

The caveat the author raises is also the ceiling on how far flagging can be pushed. A confidence number is model output, not automatically a calibrated probability; research on ASR word confidence proposes explicit calibration so values can be compared across models, and a high-confidence token can still be wrong ^[10]. A threshold tuned against one engine is a local setting, not a portable quality bar.

That is why two lists of ids do more work than a colour scale. Pending is candidates minus reviewed ^[17], which gives review a denominator that belongs to the workflow instead of the model. Confidence ordering answers only the question of where a reviewer should look first ^[9], and even a perfect ordering says nothing about whether words land at the right time or whether a line is readable before it disappears ^[8]. The one thing the author wants surfaced is the thing no aggregate score contains: how much of the transcript remains unchecked ^[8].

Sourcing is worth stating plainly. This is one engineer's build, published on dev.to, with the gp/HP case coming out of his own editor, checked against his code, and a disclosure that AI helped edit and structure the piece ^[18]. It is not a benchmark and it does not claim to be. The conclusion he draws is narrow enough to survive that: an accuracy score describes a model result, and does not define a finished review workflow ^[19].