The Record Test Bench
The Record Test Bench is a benchmark measuring generative AI models’ performance on legal transcription tasks. It compares legal transcript-editing systems across five public-record depositions, using proceedings-only certified transcript references. It measures spoken-word preservation, speaker identification and legal transcript structure—not readiness to certify.
Headline: rt-v1.1, mean of repetitions 2 and 3, with the min–max range under each value (n=2, no confidence interval). Compare models within each track only: the Gemini models run the audio chain, while GPT-6.1 Sol and the Claude models start Speaker Correct from the dataset transcript. Caveats and disclosures follow the table.
rt-v1.1 headline results
| Metric | Audio full-chain | Text-only Speaker Correct + format | ||||
|---|---|---|---|---|---|---|
| gemini-3.8-flash (low)Track rank 1 | gemini-4-argon (low)Track rank 2 | gemini-3.1-flash-lite (minimal)Track rank 3 | claude-sonnet-5-5Track rank 1 | gpt-6.1-solTrack rank 2 | claude-haiku-5-5Track rank 3 | |
| Quality | ||||||
| Proofread WER · verbatimlower is better | 22.9%range 22.8%–23.0% | 22.5%range 22.4%–22.5% | 20.2%range 20.2%–20.2% | n/a* | n/a* | n/a* |
| Proofread WER · normalizedlower is better | 13.8%range 13.7%–13.8% | 13.1%range 13.1%–13.1% | 11.9%range 11.9%–11.9% | n/a* | n/a* | n/a* |
| Speaker Correct name accuracyhigher is better | 63.9%range 61.3%–66.6% | 54.8%range 53.9%–55.6% | 43.1%range 42.0%–44.2% | 48.1%range 47.4%–48.9% | 52.5%range 52.0%–53.0% | 41.1%range 39.1%–43.1% |
| Speaker Correct boundary F1higher is better | 91.1%range 90.7%–91.4% | 91.6%range 91.5%–91.8% | 88.0%range 87.9%–88.2% | 87.1%range 87.1%–87.2% | 89.1%range 89.1%–89.2% | 86.6%range 86.6%–86.6% |
| Speaker Correct cpWERlower is better | 51.9%range 47.5%–56.2% | 50.0%range 48.2%–51.8% | 73.6%range 71.1%–76.2% | 56.8%range 56.0%–57.6% | 46.0%range 44.4%–47.7% | 66.3%range 54.0%–78.6% |
| Legal structure F1higher is better | 53.7%range 48.4%–58.9% | 48.1%range 41.1%–55.2% | 19.2%range 11.1%–27.3% | 25.8%range 22.9%–28.7% | 19.5%range 19.3%–19.8% | 8.7%range 3.9%–13.6% |
| Punctuation F1higher is better | 77.2%range 77.1%–77.4% | 75.9%range 75.9%–75.9% | 76.6%range 76.6%–76.6% | 77.4%range 77.2%–77.6% | 77.2%range 77.2%–77.2% | 77.1%range 77.1%–77.1% |
| Word preservationhigher is better | 99.0%range 98.7%–99.2% | 99.4%range 99.4%–99.4% | 97.4%range 97.1%–97.7% | 94.0%range 92.8%–95.2% | 99.8%range 99.8%–99.8% | 99.5%range 99.1%–99.9% |
| Cost | ||||||
| USD/audio-hourlower is better | $0.26range $0.26–$0.26 | $0.73range $0.73–$0.73 | $0.11range $0.11–$0.11 | $0.53range $0.53–$0.54 | $0.31range $0.31–$0.31 | $0.03range $0.03–$0.03 |
| Total stage spendlower is better | $5.32range $5.32–$5.32 | $15.16range $15.12–$15.20 | $2.36range $2.36–$2.37 | $11.12range $10.94–$11.31 | $6.37range $6.36–$6.37 | $0.58range $0.58–$0.58 |
| Stage speed · min/audio-hour | ||||||
| Proofread min/ahlower is better | 2.4range 2.4–2.5 | 3.7range 3.7–3.8 | 1.4range 1.3–1.4 | n/a* | n/a* | n/a* |
| Speaker Correct min/ahlower is better | 1.0range 0.9–1.0 | 2.6range 2.6–2.6 | 0.5range 0.5–0.5 | 1.6range 1.6–1.7 | 2.4range 2.3–2.4 | 1.2range 1.2–1.2 |
| Format min/ahlower is better | 0.8range 0.8–0.8 | 2.3range 2.3–2.4 | 0.4range 0.4–0.5 | 1.6range 1.4–1.8 | 2.2range 2.2–2.2 | 1.2range 1.1–1.2 |
| End-to-end min/ahlower is better | 4.3range 4.2–4.3 | 8.7range 8.5–8.8 | 2.4range 2.3–2.4 | 3.4range 3.3–3.5 | 4.7range 4.7–4.8 | 2.4range 2.2–2.5 |
| Speaker Correct tokens/secondhigher is better | 206.4range 201.0–211.8 | 80.8range 80.2–81.4 | 364.1range 363.1–365.2 | 205.0range 203.2–206.8 | 78.9range 77.7–80.2 | 315.9range 313.6–318.3 |
* n/a means the stage was not run for that model or has no valid timing in the matching result files; missing entries are never zero or borrowed from another run. GPT-6.1 Sol and the Claude models have no proofreading stage. USD/audio-hour divides one repetition’s recorded spend by the 20.8303 source-audio hours evaluated; total stage spend is one repetition’s recorded spend.
Exploratory: two fresh repetitions per model. Headline values are the mean of repetitions 2 and 3, with the min–max range beside each value. Two repetitions give no confidence interval and no significance claims, and failed draws stay in the denominator. Five depositions are not a population estimate. HeyDonna sponsors this research. No Record Test Bench file, certified transcript or ASR text was used in HeyDonna prompt development, examples or training.
Summary of results
Protocol rt-v1.1 is the headline: each model was run three times, and repetitions 2 and 3 are reported. These are recorded results, not statistically established winners; compare models within each track only.
- Audio quality: Gemini 3.8 Flash leads Speaker Correct name accuracy (63.9%); Gemini 4 Argon leads Speaker Correct boundary F1 (91.6%); Gemini 3.8 Flash and Gemini 4 Argon are tied for the lead in legal-structure F1 (Gemini 3.8 Flash 53.7% and Gemini 4 Argon 48.1%); Gemini 3.8 Flash and Gemini 4 Argon are tied for the lead in Speaker Correct cpWER (Gemini 3.8 Flash 51.9% and Gemini 4 Argon 50.0%, lower is better). Models whose means differ by no more than the smaller of their two repetition ranges are treated as tied, and tied values are all highlighted.
- Audio cost and speed: Gemini 3.1 Flash Lite has the lowest spend per repetition ($2.36) and the lowest end-to-end median (2.4 min/audio-hour); Gemini 4 Argon has the highest spend ($15.16).
- Text-only quality: Sonnet 5.5 leads legal-structure F1 (25.8%); GPT-6.1 Sol leads Speaker Correct name accuracy (52.5%), Speaker Correct boundary F1 (89.1%) and Speaker Correct cpWER (46.0%, lower is better). These models start from the dataset transcript, so they are not comparable with the audio chain.
- Text-only cost and speed: Haiku 5.5 has the lowest spend per repetition ($0.58) and the lowest end-to-end median (2.4 min/audio-hour); Sonnet 5.5 has the highest spend ($11.12).
- Reliability: failures stay in the denominator. 35 of 225 expected draws failed across all three repetitions, all in the first (legacy, unpooled) repetition of the three Gemini models; repetitions 2 and 3 completed every draw for every model.
1. Model comparison
Headline: rt-v1.1, repetitions 2 and 3. Each cell is the mean of the two repetitions’ five-file results, with the min–max range underneath; with two repetitions the range is a spread, not a confidence interval. Repetition 1 is legacy and unpooled (see run accounting).
Legal-structure ranking, within each track only. Descending mean legal-structure F1; Speaker Correct name accuracy breaks ties, then model ID. Lower is better for error rates, cost and elapsed time; higher is better for accuracy, F1, word preservation and tokens/second. Quality values are percentages.
Comparability caveat. The three Gemini models run the audio chain: proofreading → Speaker Correct → format, each stage fed by that model’s own previous output. GPT-6.1 Sol and both Claude models have no audio input here, so there is no proofreading stage: Speaker Correct starts from the dataset’s provided transcript and format consumes the model’s own Speaker Correct output. Their results are not comparable with the audio chain’s, and proofreading metrics exist only for the Gemini models.
Bold, shaded values are the best recorded mean within each track, plus any model tied with it. Tie rule: two models are tied when their means differ by no more than the smaller of their two min–max ranges across repetitions. This is a reading aid with n=2, not a significance test. Tables scroll horizontally. Dated per-file scores, per-repetition values and configuration remain in the downloadable results.
Run accounting: failures stay in the denominator
A draw is one model × repetition × stage × file. A failed draw is counted and never replaced, and a failed run forms no aggregate: its values are n/a, not zero. Repetition 1 is legacy and unpooled; the three Gemini repetition-1 runs ended with a failed stage execution (“stage execution failed; requested cases retained”; the result files record no cause), so they are shown here but contribute no headline figure.
| Model | Draws succeeded / expected, all 3 repetitions | Failed draws | Repetitions 2–3 only |
|---|---|---|---|
| gemini-3.8-flash | 31/45 | 14 | 30/30 |
| gemini-4-argon | 38/45 | 7 | 30/30 |
| gemini-3.1-flash-lite | 31/45 | 14 | 30/30 |
| gpt-6.1-sol | 30/30 | 0 | 20/20 |
| claude-sonnet-5-5 | 30/30 | 0 | 20/20 |
| claude-haiku-5-5 | 30/30 | 0 | 20/20 |
| Model | Repetition | Run status | Proofread | Speaker Correct | Format | Complete chains (of 5) | Spend (USD) |
|---|---|---|---|---|---|---|---|
| gemini-3.8-flash | 1 (legacy, unpooled) | FAILED | 1/5 | 0/5 | 0/5 | 0 | $2.56 |
| gemini-3.8-flash | 2 | COMPLETE | 5/5 | 5/5 | 5/5 | 5 | $5.32 |
| gemini-3.8-flash | 3 | COMPLETE | 5/5 | 5/5 | 5/5 | 5 | $5.32 |
| gemini-4-argon | 1 (legacy, unpooled) | FAILED | 5/5 | 3/5 | 0/5 | 0 | $11.50 |
| gemini-4-argon | 2 | COMPLETE | 5/5 | 5/5 | 5/5 | 5 | $15.12 |
| gemini-4-argon | 3 | COMPLETE | 5/5 | 5/5 | 5/5 | 5 | $15.20 |
| gemini-3.1-flash-lite | 1 (legacy, unpooled) | FAILED | 1/5 | 0/5 | 0/5 | 0 | $1.44 |
| gemini-3.1-flash-lite | 2 | COMPLETE | 5/5 | 5/5 | 5/5 | 5 | $2.37 |
| gemini-3.1-flash-lite | 3 | COMPLETE | 5/5 | 5/5 | 5/5 | 5 | $2.36 |
| gpt-6.1-sol | 1 (legacy, unpooled) | COMPLETE | n/a | 5/5 | 5/5 | 5 | $6.35 |
| gpt-6.1-sol | 2 | COMPLETE | n/a | 5/5 | 5/5 | 5 | $6.36 |
| gpt-6.1-sol | 3 | COMPLETE | n/a | 5/5 | 5/5 | 5 | $6.37 |
| claude-sonnet-5-5 | 1 (legacy, unpooled) | COMPLETE | n/a | 5/5 | 5/5 | 5 | $11.28 |
| claude-sonnet-5-5 | 2 | COMPLETE | n/a | 5/5 | 5/5 | 5 | $11.31 |
| claude-sonnet-5-5 | 3 | COMPLETE | n/a | 5/5 | 5/5 | 5 | $10.94 |
| claude-haiku-5-5 | 1 (legacy, unpooled) | COMPLETE | n/a | 5/5 | 5/5 | 5 | $0.62 |
| claude-haiku-5-5 | 2 | COMPLETE | n/a | 5/5 | 5/5 | 5 | $0.58 |
| claude-haiku-5-5 | 3 | COMPLETE | n/a | 5/5 | 5/5 | 5 | $0.58 |
Stage cells are succeeded/expected files; n/a means the model has no such stage. The 18 result files record $115.58 of stage spend in total, including the failed runs. One abandoned partial run that was never reused is not among those files.
Per-run values
| Model | Repetition | cpWER | Name accuracy | Boundary F1 | Turn-speaker accuracy | Legal-structure F1 | Punctuation F1 | Proofread WER |
|---|---|---|---|---|---|---|---|---|
| gemini-3.8-flash | 1 (legacy, unpooled) | n/a | n/a | n/a | n/a | n/a | n/a | n/a |
| gemini-3.8-flash | 2 | 47.5% | 66.6% | 91.4% | 76.2% | 48.4% | 77.4% | 22.8% |
| gemini-3.8-flash | 3 | 56.2% | 61.3% | 90.7% | 71.8% | 58.9% | 77.1% | 23.0% |
| gemini-4-argon | 1 (legacy, unpooled) | n/a | n/a | n/a | n/a | n/a | n/a | n/a |
| gemini-4-argon | 2 | 48.2% | 53.9% | 91.5% | 76.6% | 55.2% | 75.9% | 22.5% |
| gemini-4-argon | 3 | 51.8% | 55.6% | 91.8% | 74.8% | 41.1% | 75.9% | 22.4% |
| gemini-3.1-flash-lite | 1 (legacy, unpooled) | n/a | n/a | n/a | n/a | n/a | n/a | n/a |
| gemini-3.1-flash-lite | 2 | 76.2% | 44.2% | 88.2% | 56.7% | 27.3% | 76.6% | 20.2% |
| gemini-3.1-flash-lite | 3 | 71.1% | 42.0% | 87.9% | 57.0% | 11.1% | 76.6% | 20.2% |
| gpt-6.1-sol | 1 (legacy, unpooled) | 44.3% | 53.0% | 89.2% | 74.9% | 27.6% | 77.3% | n/a |
| gpt-6.1-sol | 2 | 47.7% | 52.0% | 89.1% | 74.1% | 19.8% | 77.2% | n/a |
| gpt-6.1-sol | 3 | 44.4% | 53.0% | 89.2% | 75.0% | 19.3% | 77.2% | n/a |
| claude-sonnet-5-5 | 1 (legacy, unpooled) | 54.4% | 49.4% | 87.2% | 67.7% | 35.9% | 77.3% | n/a |
| claude-sonnet-5-5 | 2 | 57.6% | 48.9% | 87.2% | 67.4% | 22.9% | 77.2% | n/a |
| claude-sonnet-5-5 | 3 | 56.0% | 47.4% | 87.1% | 66.0% | 28.7% | 77.6% | n/a |
| claude-haiku-5-5 | 1 (legacy, unpooled) | 74.5% | 40.3% | 86.7% | 57.7% | 13.8% | 77.2% | n/a |
| claude-haiku-5-5 | 2 | 78.6% | 39.1% | 86.6% | 54.5% | 13.6% | 77.1% | n/a |
| claude-haiku-5-5 | 3 | 54.0% | 43.1% | 86.6% | 66.0% | 3.9% | 77.1% | n/a |
Repetition 1 values for GPT-6.1 Sol and the Claude models are complete but legacy and unpooled; they are not in any headline figure.
2. What the metrics measure
Quality values are percentages. WER means word error rate; F1 balances finding expected items with avoiding incorrect extras. Both omissions and extras can lower F1. All examples below are tiny, synthetic illustrations—not benchmark results. Definitions describe one run’s five-file result; the headline additionally averages repetitions 2 and 3 and shows their range.
- Proofread WER · verbatim word error rate
Lower is better. Counts substituted, missing and added words against the reference after recognized speaker labels are removed. Case and punctuation remain significant. The table is the mean of five file-level error rates.
Example:
- Reference
Q. Did you sign it?
A. Yes. - Output
Q. Did you sign it?
A. No.
replacing “Yes.” with “No.” is one word substitution.
- Proofread WER · normalized word error rate
Lower is better. The same word-error comparison, but both texts are lowercased and punctuation is stripped first. This separates word changes from case and punctuation differences; the table is the five-file mean.
Example:
- Reference
Q. Did you sign it?
A. Yes. - Output
Q. did you sign it
A. no
“yes” → “no” is an error; the case and punctuation changes are not.
- Speaker Correct name accuracy · naming the right person
Higher is better. For each reference speaker turn, finds the output label covering the most aligned words and checks whether it names that person. Some name-spelling differences are tolerated; a missing or unmatched turn is not correct. The table is the five-file mean.
Example:
- Reference
MS. JONES: Yes, I signed it. - Output
MR. BROWN: Yes, I signed it.
the words match, but the speaker name is wrong.
- Speaker Correct boundary F1 · finding speaker changes
Higher is better. Balances finding the reference speaker-change points with avoiding extra ones. Boundaries are matched in aligned word positions with a two-word tolerance, not in seconds; the table is the five-file mean. It does not check speaker names.
Example:
- Reference
MR. SMITH: Did you sign it?
MS. JONES: Yes. - Output
MR. SMITH: Did you sign it? Yes.
merging the answer into the question loses a speaker-change boundary.
- Speaker Correct cpWER · speaker-grouped word error rate
Lower is better. Concatenates each speaker’s dialogue in turn order, then chooses the one-to-one pairing of reference and output speakers with the fewest word edits. Missing or extra speakers are paired with empty text. Edits are divided by reference words, then averaged across five files. Case, punctuation, speaker names and timestamps are ignored; this is not a timed diarization error rate.
Example:
- Reference
MR. SMITH: Did you sign it?
MS. JONES: Yes. - Output
MR. SMITH: Did you sign it? Yes.
all words remain, but combining two speakers’ word streams creates errors. Renaming the two speakers consistently would not.
- Legal structure F1 · retaining legal transcript elements
Higher is better. Matches five classes of legal lines: testimony openings, proceedings closings, examination headings, attorney bylines and other parentheticals. It uses normalized text or event-specific matching keys and pools matching, output and reference counts across classes and files. Missing or extra elements hurt; this is not a check of page layout or placement relative to dialogue. Oaths, appearances, stipulations and certificates are excluded.
Example:
- Reference
CROSS-EXAMINATION
BY MS. JONES:
Q. Did you sign it? - Output
Q. Did you sign it?
the missing examination heading and attorney byline are errors.
- Punctuation F1 · matching punctuation around aligned words
Higher is better. Compares punctuation types in the gaps after words aligned between final output and reference. Missing, extra or changed punctuation types hurt; counts are pooled across files. Unaligned words are not scored as punctuation errors.
Example:
- Reference
Q. Did you sign it? - Output
Q. Did you sign it.
the question mark is missing and a period has been added in its place.
- Word preservation · keeping the formatting input’s words
Higher is better. Compares lowercase word-occurrence counts before and after formatting. It takes the smaller of the input and output shares accounted for by their shared words, so omissions and additions both lower it. The table is the five-file mean. Word order is not measured, and retained words may still be wrong against the certified reference.
Example:
- Reference · Formatting input
Did you sign it? - Output
Did you sign?
dropping “it” lowers preservation even though the remaining words are unchanged.
- USD/audio-hour · provider spend per hour of source audio
Lower is better. Divides recorded provider spend across the executed stages by the total source-audio hours evaluated. Text-only runs still use source-audio duration as the denominator. In the rt-v1.1 headline this is one repetition’s spend. It is historical spend, not a current price quote or an accuracy score.
- Total stage spend · recorded provider dollars for the run
Lower is better. The recorded provider-dollar total across all executed stages and evaluated files, without dividing by audio duration. The text-only track has no proofreading stage. Historical retry and cache accounting is incomplete, so this is not a full operational-cost estimate.
- Proofread min/audio-hour · proofreading elapsed time
Lower is better. For each file, divides proofreading wall time in minutes by source-audio hours, then takes the median. The wall time is the runner’s monotonic stage time. Lower means less recorded elapsed time, not fewer transcript errors.
- Speaker Correct min/audio-hour · speaker-stage elapsed time
Lower is better. For each file, divides the recorded full-file Speaker Correct wall time in minutes by source-audio hours, then takes the median. It measures the complete stage, not just the provider’s token stream.
- Format min/audio-hour · formatting elapsed time
Lower is better. For each file, divides formatting wall time in minutes by source-audio hours, then takes the median. The wall time is the recorded per-file time, before scoring. Failed stages, fully cached files and lost timings are excluded.
- End-to-end min/audio-hour · elapsed time for the executed chain
Lower is better. Sums each file’s executed stage wall times before dividing by its source-audio hours, then takes the median. Audio runs include proofreading, Speaker Correct and formatting; text-only runs include Speaker Correct and formatting only. Cross-file queueing and adapters are excluded; any missing required stage timing makes that file’s total unavailable.
- Speaker Correct tokens/second · recorded output throughput
Higher is better. Divides each file’s reported output-token count by its full Speaker Correct wall seconds, then takes the median. Tokens are model text units, not words; tokenizers differ. This reflects benchmark concurrency, not isolated provider throughput or transcript quality.
WER and cpWER can exceed 100%. Word preservation is measured against the formatting input, not the certified reference. Timing coverage such as (4/5) is separate from completed-output coverage. Missing measurements are n/a, never zero.
3. Methodology
The Record Test Bench corpus (dataset rt-v1) contains 20.8 hours of English deposition audio; protocol rt-v1.1 reruns the same five public depositions and reference transcripts. References are prepared from certified transcripts, aligned to audio and projected to proceedings-only text; dropped pages are excluded. This small development sample is not contamination-free, representative of all legal work, or an internal PR-regression benchmark.
Scoring and aggregation
The metric explanations above describe each scored quantity and its aggregation. Each repetition is first reduced to its own five-file result; the rt-v1.1 headline is the mean of repetitions 2 and 3, shown with their min–max range. Speaker-turn and word-error metrics use the five-file mean; legal structure and punctuation pool their underlying counts across files. The historical export does not preserve every zero-denominator policy or dependency version; do not infer perfect quality from empty classes. Per-class reference support is included in the downloadable results.
Speed, cost and uncertainty
rt-v1.1 timings are the per-file wall times recorded by the runner (monotonic for proofreading; recorded per-file wall for Speaker Correct and format), under benchmark concurrency; the median across files is reported per repetition and then averaged over repetitions 2 and 3. Minutes per audio-hour = 60 × stage wall seconds / source-audio seconds, then the median across files. End-to-end sums each file’s executed stages before taking the median, excluding cross-file queueing/adapters. Text-only sums Speaker Correct + format only. Coverage appears when fewer than five measurements exist.
Fully cached format files and failed stages are excluded from timings. Rates reflect benchmark concurrency, not isolated provider throughput.
USD/audio-hour divides recorded provider stage spend by 20.8303 evaluated source-audio hours. Totals are reported spend, not a current price quote. Cache discounts, retry accounting, token usage and historical pricing were not fully captured.
The rt-v1.1 headline pools two fresh full-chain repetitions per model; the first repetition is reported separately as legacy and unpooled, and failed draws stay in the denominator. With two repetitions the range beside each value is a spread, not a confidence interval; there are no significance claims or winner badges. Confirmatory comparisons require at least three independent full-chain repetitions in the headline, fixed settings, all attempts retained and paired hierarchical bootstrap intervals over depositions and repetitions. Five files cannot establish population validity.
Snapshot and reproducibility
Download rt-v1.1 allowlisted results · Download artifact checksums. Snapshot rt-v1.1-cf310b32baa0 pins the 18 rt-v1.1 result files through their individual SHA-256 digests. rt-v1.1 supersedes the earlier single-run rt-v1, whose results JSON remains available. Downloaded reference and output files have individual SHA-256 digests. Only explicitly selected public RT fields and text artifacts are released; no customer data, internal representative bench data, provider traces or credentials.
Reproduce
Use the Hugging Face dataset for the audio, reference transcripts and the model outputs, and the results dataset for the scores. The rt-v1.1 3-rep outputs and scores for snapshot rt-v1.1-cf310b32baa0 are published there, alongside the legacy rt-v1 outputs and scores (snapshot rt-v1-1c5a96c70469). Immutable links: corpus, pinned commit and results, pinned commit. Consult the dataset card for file layout and evaluation details.
Independent rerun tooling, complete prompts/configuration, and dependency and pricing snapshots remain follow-up work. This launch publishes recorded evidence, not a claim of fully reproducible execution. Future updates must freeze protocol and corpus revisions, retain dated snapshots and publish corrections explicitly. There is no pay-to-rank; request a maintainer-run evaluation or correction via legal@heydonna.law. No submitted code runs automatically.
4. Data & downloads
- Full dataset — audio, reference transcripts, per-model outputs and scores.
- rt-v1.1 results (JSON) — headline aggregates, per-repetition values and draw accounting for snapshot
rt-v1.1-cf310b32baa0. - Scores-only results — the results without the audio and transcript files.
Snapshot IDs: rt-v1.1 headline rt-v1.1-cf310b32baa0; legacy rt-v1 rt-v1-1c5a96c70469. The downloadable results on this page are the authoritative rt-v1.1 record. The Hugging Face datasets carry the rt-v1.1 snapshot; see corpus, pinned commit and results, pinned commit.
Per-source attribution
- James Mitchell — KODX Seattle.
- Bruce Jessen — KODX Seattle.
- John Rizzo — KODX Seattle.
- Todd E. Dean — UCSF Industry Documents Library.
- Karen Degen — UCSF Industry Documents Library.
See the dataset card for attribution, preparation and source-specific rights details. For takedown requests, contact legal@heydonna.law.
5. Sources and licenses
Audio links below go to the original public sources; this site does not host audio. The full dataset is linked above. Reference transcripts are also static research downloads here. No blanket license covers the entire package.
| Deposition | Audio / source attribution | Reference transcripts |
|---|---|---|
| James Mitchell5.25 hours | KODX Seattle · original sourceCreative Commons Attribution (CC BY 3.0), as published on YouTube | proofread · Speaker Correct · formatted |
| Bruce Jessen4.00 hours | KODX Seattle · original sourceCreative Commons Attribution (CC BY 3.0), as published on YouTube | proofread · Speaker Correct · formatted |
| John Rizzo4.00 hours | KODX Seattle · original sourceCreative Commons Attribution (CC BY 3.0), as published on YouTube | proofread · Speaker Correct · formatted |
| Todd E. Dean3.20 hours | UCSF Industry Documents Library · original sourceResearch / fair-use assertion; see below | proofread · Speaker Correct · formatted |
| Karen Degen4.38 hours | UCSF Industry Documents Library · original sourceResearch / fair-use assertion; see below | proofread · Speaker Correct · formatted |
Mitchell / Jessen / Rizzo attribution. KODX Seattle, original videos linked above. Creative Commons Attribution (CC BY 3.0), as published on YouTube. The video pages label the license “Creative Commons Attribution license (reuse allowed)”; YouTube’s CC option is described in YouTube Help. Changes for this benchmark include audio preparation, certified-reference preparation/projection and model-generated edits. Source video licensing does not automatically license independently prepared reference transcripts or generated outputs.
Dean / Degen research and fair-use disclaimer. Sources: UCSF Industry Documents Library, via the archive.org links above. Redistribution is for research and evaluation under an asserted fair-use rationale; fair use is not guaranteed. We claim no ownership or blanket reuse license. Public availability does not itself confer redistribution rights.
For rights, privacy or takedown requests contact legal@heydonna.law. Do not use these materials as legal advice or a certified record.