Sampled Word Evidence

Owner: vault-ingress. local_media_transcription.transcribe_local_words() returns established source-probe facts and a schema-v2 word receipt. It writes no transcript, timing/quality sidecar, speech profile or catalog record. This lane is separate from the existing transcript acquisition contract.

Acquisition and Native Boundary

The caller supplies a local locator, sample_start_seconds, sample_duration_seconds, optional trusted root and optional current in-memory media/video probe. Missing probe facts are acquired by probe_local_media. Never supply a serialized or manufactured probe. The original source's open descriptor and pathname generation must remain unchanged throughout sampling and transcription. The caller receives no usable result after a failed final generation check or failed workspace cleanup.

The parent creates an owner-private workspace; the authenticated Whisper worker decodes only the requested interval with FFmpeg. Streamed PCM has a strict byte ceiling, a measured duration from frame count and a private WAV wrapper. The sample has its own digest and generation. The original source is not copied or hashed again when a current owner probe is reused. Parent-owned cleanup removes sample files even when the worker times out or is killed.

MEDIA_WHISPER_LIMITS owns wall, memory, process and protocol/diagnostic limits. local_media_sampling.py owns PCM format, clip bounds and decoder refusals. All parsing, normalization, protocol framing, generation checks and cleanup run in CI with real synthetic workers. Only native MLX calls use the documented platform-bound carve-out in the transcription authority.

The supported model ID and immutable Hugging Face revision live in local_media_words.DEFAULT_WORD_MODEL. Renew quarterly with native validation. The optional whisper manifest extra pins mlx-whisper and huggingface-hub; weekly Dependabot tracks their versions. The SDK resolves only the pinned model's configuration and weights into its normal reusable cache; it does not download remote Python code, use stored access tokens, or receive audio. Model acquisition shares the worker's timeout/resource bounds. SDK cache state is separate from per-recording scratch and is not deleted after each sample.

The native call requests real word timestamps, deterministic temperature and no previous-text conditioning. The same model separately measures language probability from the first 30 seconds of the decoded sample. A language label alone, a word probability, or the expected language does not supply that value. This probe does not prove language homogeneity across the complete sample.

Closed Receipt Schema

All listed keys are required; unknown keys and future versions refuse:

{schema_version: 3, pipeline_version: "sampled-words-v3",
 source_sha256, sample_sha256, source_duration_seconds,
 sample_start_seconds, sample_duration_seconds,
 provider: "mlx-whisper", provider_version,
 model: {id, revision}, language, language_probability, language_probe_seconds,
 words: [{text, start_seconds, end_seconds, probability, segment_index}, ...],
 segments: [{start_seconds, end_seconds, compression_ratio,
             average_log_probability, no_speech_probability}, ...],
 token_exclusions: [{token_index, reason: "punctuation_only" |
                     "nonpositive_span"}, ...]}

The root receipt version owns the fixed nested model, word, segment and token exclusion shapes. Words carry one lexical token each; the model's tokenization convention is preserved. Punctuation-only tokens and degenerate tokens whose provider span is zero or negative are omitted, each with an explicit original token index and its reason. Invalid lexical timestamps, overlaps, missing word alignment, invalid Unicode and segment membership failures refuse the sample. No timestamps are interpolated, repaired, stretched or clipped.

A sample is admitted with its degenerate tokens excluded and recorded until their share exceeds the admission bound, above which the sample refuses whole with whisper_word_sample_invalid_word_nonpositive_span. The bound is WORDS_MAX_NONPOSITIVE_SHARE at the top of skills/vault-ingress/scripts/local_media_words.py. Exclusion is never repair: a retained word always carries a positive span, and a receipt that claims otherwise refuses on read.

Word and segment times are relative to the sample, not the full recording. Each word retains its zero-based provider segment_index. Indices are ordered; each word must overlap its declared segment by a positive duration. Native MLX boundary adjustments can retain a segment timestamp inside its boundary word; strict containment is not a provider guarantee. Neither timestamp is changed. All word and segment spans remain positive, ordered, non-overlapping and inside the sample. Source and sample SHA-256 values bind different byte artifacts. sample_start_seconds maps the decoded interval back to the original source, and its actual PCM duration must fit that source. Confidence/probability fields are measurements, not quality verdicts. The separate profile owner records the versioned policy that accepts or excludes each sample.

normalize_word_result() projects bounded provider data into this receipt; validate_word_sample() is the non-mutating reader. Both raise the path-neutral LocalMediaError contract. Neither pure helper authenticates arbitrary supplied digests. Consumers obtain receipts from transcribe_local_words, preserve them unchanged and recheck live source freshness through ingress before reuse. Missing receipts mean no word evidence. Segment-only timing sidecars are not migrated into words; obtain fresh alignment through this owner. Earlier receipts are not accepted or auto-migrated: v1 omitted explicit membership, and v2 refused a sample outright on any degenerate token, so its admitted set is not comparable. Run fresh owner acquisition to obtain v3 evidence.

WordSampleError adds closed numeric failure details: schema_version: 1, word_index, word_count, word_start_seconds, word_end_seconds, segment_index, segment_count, segment_start_seconds and segment_end_seconds. Indices/counts are bounded integers and times are finite sample-relative seconds. A missing segment uses null index/start/end together. Unknown keys, words and source paths are forbidden. Worker transport revalidates these details before exposing them to a caller.

Native Manual Validation

Run on Apple Silicon with the configured interpreter and optional whisper extra. Use an authorized recording whose solo-speaker identity and language are independently known. This procedure covers only the MLX import, native transcription, spectrogram/model access and language-detection call named in the transcription authority; it does not exempt orchestration tests.

1. Resolve the vault root, speaker name and exact interpreter through the strict owner bootstrap in vault-profile. Record the owner's catalog digest and preserve any existing speech profile and transcript artifacts.

2. Require native capability:

   "{python_path}" "{speaker_toolkit_root}/skills/vault-ingress/scripts/check-runtime.py" \
     --lanes core,source-media,speech-calibration --require-lanes core,source-media,speech-calibration

   Include youtube-download in both lists if downloads are authorized.

3. Run a metadata plan through calibrate-speech.py with the explicit speaker, language and --maximum-recordings 1. Inspect the selected source identity; no recording is opened by the plan. Use --allow-download only for authorized YouTube acquisition and retain that same option for the measured run.

4. Repeat with --run, capturing stdout into a fresh private report. Require exit 0, ok: true, one admitted recording and status: "low_confidence" for this intentionally single-recording test. An acquisition or quality exclusion is not a pass. It requires source/provider investigation, not lowered gates.

5. Inspect data.profile.calibration.samples[0].words: current provider version, pinned model ID/revision, exact source/sample digests, actual interval bounds, lexical word start/end times and real language probability are present. Listen to the authorized source at the recorded sample interval and check several word boundaries near the beginning, middle and end. Require plausible word alignment and correct language; evenly spaced invented timestamps fail. A high reported probability does not excuse a wrong detected language.

6. Require all four metric names with their owner thresholds, the recording's actual word count and interval, explicit low confidence and null conservative planning rates. The single-family confidence interval remains null.

7. Repeat the strict catalog read and require the starting digest. Confirm no transcript, timing/quality receipt or prior profile changed and no private sample workspace survives. Model weights may remain in the SDK cache.

A native pass requires all seven checks. A single-recording validation does not establish a representative speaker profile or recorder acceptance. Run the full cohort separately and keep its sparse/missing-mode uncertainty explicit.