Pronunciation assessment: numeric tokens in ReferenceText (e.g. "10:17") are marked Mispronunciation with no phonemes

When the ReferenceText contains a numeric/time token such as "10:17", pronunciation assessment marks that token as Mispronunciation and returns a single phoneme entry with an empty phoneme symbol, even though the speaker read it correctly. Other words in the same utterance get normal phoneme-level results. As a result, learners who read the sentence correctly are told they mispronounced the word, and we cannot show them which sounds were wrong because no phonemes are returned. This affects sentences in our service that contain a number or a time.

Request (REST API for short audio, no SDK, region koreacentral)

• POST https://koreacentral.stt.speech.microsoft.com/speech/recognition/conversation/cognitiveservices/v1?language=en-US
• Content-Type: audio/wav; codecs=audio/pcm; samplerate=16000
• Pronunciation-Assessment: {"ReferenceText": "<see below>", "GradingSystem": "HundredMark", "Granularity": "Phoneme", "Dimension": "Comprehensive", "phonemeAlphabet": "IPA"}
• The same WAV file was used for both calls. The speaker says "but it's ten seventeen".

Call 1 - ReferenceText "but it's 10:17"

• Accuracy 79 / Fluency 91 / Completeness 67 / PronScore 74.2
• but: 97 (b 64, ʌ 49, t 46)
• it's: 97 (ɪ 79, t 65, s 53)
• 10:17: 43, ErrorType Mispronunciation, one phoneme entry with Phoneme "" (score 43)

Call 2 - ReferenceText "but it's ten seventeen"

• Accuracy 93 / Fluency 97 / Completeness 100 / PronScore 95.2
• ten: 82 (t 80, ɛ 87, n 74)
• seventeen: 97 (all phonemes returned)

Questions

1. Is text normalization applied to ReferenceText for numbers, times and symbols (e.g. 10:17, 1995, $5, 3rd)? If so, what are the rules?
2. What does the word-level score (43 above) mean when no real phonemes are aligned? Is it safe to show to learners?
3. Is there a recommended way to write ReferenceText containing numbers, or a way to accept multiple readings (e.g. 5:00 as "five" or "five o'clock")?

We could not find any guidance on numeric tokens in the pronunciation assessment documentation. I can share the audio file and the full JSON responses if needed.
pa_10-17.txt