heydonna/record-test-bench

The Record Test Bench (rt-v1.1)

Corpus, prepared references, model outputs and results.

A benchmark measuring generative AI models' performance on legal transcription tasks: proofreading, Speaker Correct (speaker correction / SC), and formatting. Five English public-record depositions, proceedings-only certified-transcript-derived references; 20.83 evaluated source-audio hours. The benchmark measures complete editing systems, not readiness to certify a transcript.

Headline snapshot: rt-v1.1-cf310b32baa0 · protocol: rt-v1.1 · corpus: rt-v1 (the same five depositions and references) · evaluation dates: 2026-10-18–2026-10-19 · results schema: rt-public-results/2.

Legacy snapshot (retained unchanged): rt-v1-1c5a96c70469 · single-run rt-v1 · evaluation dates 2026-10-15–2026-10-16 · results schema rt-public-results/1. rt-v1.1 supersedes it; the legacy files stay at their original paths and the aggregate and per_file configurations still point to them.

Canonical leaderboard: https://heydonna.law/leaderboard/

Scope: This package contains the corpus, scores and model outputs; its scores-only companion is https://huggingface.co/datasets/heydonna/record-test-bench-results. The leaderboard's downloadable results JSON is the authoritative record, so match the snapshot ID before comparing.

Tasks, tracks and ranking

• Audio full-chain: audio proofreading → Speaker Correct → formatting; actual upstream outputs feed the next stage.
• Text-only: supplied dataset transcript → Speaker Correct → formatting. Audio proofreading is N/A—not run; do not compare this track directly with audio chains.
• rt-v1.1 (headline): six complete systems, three repetitions each. Repetition 1 is legacy and unpooled; the headline is the mean of repetitions 2 and 3 with the min–max range (n=2, no confidence interval). Ranking, within each track only: descending mean legal_structure_f1, then SC name accuracy, then model ID. This is a legal-structure ranking, not a universal best-transcription claim.
• Legacy rt-v1: six complete 5/5 systems, one repetition each, ranked the same way.
• Legacy rt-v1: one incomplete audio chain is unranked. Its successful stage outputs are included; diagnostics and failure categories are separate from ranked aggregates. Failed formatting outputs/scores are excluded. Attempt counts were not recorded and remain null.

rt-v1.1 headline results

Protocol rt-v1.1 reruns the same five public depositions and reference transcripts. Each model ran three times; repetitions 2 and 3 are the headline. Cells show the mean of the two repetitions' five-file results, then (min–max) across the two repetitions. With two repetitions the range is a spread, not a 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 and have no proofreading stage (n/a, not zero).

Legal-structure ranking, within each track only. Descending mean legal-structure F1; Speaker Correct name accuracy breaks ties, then model ID.

Audio full-chain

Model (track rank)	Proofread WER verbatim	Proofread WER normalized	SC name accuracy	SC boundary F1	SC cpWER (lower better)	Legal-structure F1	Punctuation F1	Word preservation	USD/audio-hour	Spend per repetition	End-to-end min/audio-hour
gemini-3.8-flash (low reasoning; #1)	22.9% (22.8%–23.0%)	13.8% (13.7%–13.8%)	63.9% (61.3%–66.6%)	91.1% (90.7%–91.4%)	51.9% (47.5%–56.2%)	53.7% (48.4%–58.9%)	77.2% (77.1%–77.4%)	99.0% (98.7%–99.2%)	$0.26 ($0.26–$0.26)	$5.32 ($5.32–$5.32)	4.3 (4.2–4.3)
gemini-4-argon (low reasoning; #2)	22.5% (22.4%–22.5%)	13.1% (13.1%–13.1%)	54.8% (53.9%–55.6%)	91.6% (91.5%–91.8%)	50.0% (48.2%–51.8%)	48.1% (41.1%–55.2%)	75.9% (75.9%–75.9%)	99.4% (99.4%–99.4%)	$0.73 ($0.73–$0.73)	$15.16 ($15.12–$15.20)	8.7 (8.5–8.8)
gemini-3.1-flash-lite (minimal reasoning; #3)	20.2% (20.2%–20.2%)	11.9% (11.9%–11.9%)	43.1% (42.0%–44.2%)	88.0% (87.9%–88.2%)	73.6% (71.1%–76.2%)	19.2% (11.1%–27.3%)	76.6% (76.6%–76.6%)	97.4% (97.1%–97.7%)	$0.11 ($0.11–$0.11)	$2.36 ($2.36–$2.37)	2.4 (2.3–2.4)

Text-only Speaker Correct + format

Model (track rank)	SC name accuracy	SC boundary F1	SC cpWER (lower better)	Legal-structure F1	Punctuation F1	Word preservation	USD/audio-hour	Spend per repetition	End-to-end min/audio-hour
claude-sonnet-5-5 (#1)	48.1% (47.4%–48.9%)	87.1% (87.1%–87.2%)	56.8% (56.0%–57.6%)	25.8% (22.9%–28.7%)	77.4% (77.2%–77.6%)	94.0% (92.8%–95.2%)	$0.53 ($0.53–$0.54)	$11.12 ($10.94–$11.31)	3.4 (3.3–3.5)
gpt-6.1-sol (#2)	52.5% (52.0%–53.0%)	89.1% (89.1%–89.2%)	46.0% (44.4%–47.7%)	19.5% (19.3%–19.8%)	77.2% (77.2%–77.2%)	99.8% (99.8%–99.8%)	$0.31 ($0.31–$0.31)	$6.37 ($6.36–$6.37)	4.7 (4.7–4.8)
claude-haiku-5-5 (#3)	41.1% (39.1%–43.1%)	86.6% (86.6%–86.6%)	66.3% (54.0%–78.6%)	8.7% (3.9%–13.6%)	77.1% (77.1%–77.1%)	99.5% (99.1%–99.9%)	$0.03 ($0.03–$0.03)	$0.58 ($0.58–$0.58)	2.4 (2.2–2.5)

USD/audio-hour divides one repetition's recorded spend by the 20.8303 source-audio hours evaluated; spend per repetition is one repetition's recorded spend. Values are generated from results/rt-v1.1-cf310b32baa0.json and match the leaderboard.

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.

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	Repetition	Run status	Proofread	Speaker Correct	Format	Spend (USD)
gemini-3.8-flash	1 (legacy, unpooled)	FAILED	1/5	0/5	0/5	$2.56
gemini-3.8-flash	2	COMPLETE	5/5	5/5	5/5	$5.32
gemini-3.8-flash	3	COMPLETE	5/5	5/5	5/5	$5.32
gemini-4-argon	1 (legacy, unpooled)	FAILED	5/5	3/5	0/5	$11.50
gemini-4-argon	2	COMPLETE	5/5	5/5	5/5	$15.12
gemini-4-argon	3	COMPLETE	5/5	5/5	5/5	$15.20
gemini-3.1-flash-lite	1 (legacy, unpooled)	FAILED	1/5	0/5	0/5	$1.44
gemini-3.1-flash-lite	2	COMPLETE	5/5	5/5	5/5	$2.37
gemini-3.1-flash-lite	3	COMPLETE	5/5	5/5	5/5	$2.36
gpt-6.1-sol	1 (legacy, unpooled)	COMPLETE	n/a	5/5	5/5	$6.35
gpt-6.1-sol	2	COMPLETE	n/a	5/5	5/5	$6.36
gpt-6.1-sol	3	COMPLETE	n/a	5/5	5/5	$6.37
claude-sonnet-5-5	1 (legacy, unpooled)	COMPLETE	n/a	5/5	5/5	$11.28
claude-sonnet-5-5	2	COMPLETE	n/a	5/5	5/5	$11.31
claude-sonnet-5-5	3	COMPLETE	n/a	5/5	5/5	$10.94
claude-haiku-5-5	1 (legacy, unpooled)	COMPLETE	n/a	5/5	5/5	$0.62
claude-haiku-5-5	2	COMPLETE	n/a	5/5	5/5	$0.58
claude-haiku-5-5	3	COMPLETE	n/a	5/5	5/5	$0.58

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. 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.

Metrics

Metric	Definition and aggregation	Direction
Proofread WER, verbatim / normalized	(substitutions + deletions + insertions) / reference words, arithmetic mean over five files. Verbatim retains case and punctuation; normalized lowercases and strips punctuation with the pinned scorer.	Lower
SC name accuracy	Fraction of reference turns assigned the correct named speaker by the mapped-turn scorer; five-file mean.	Higher
SC boundary F1	2TP / (2TP + FP + FN) for turn boundaries matched within two words; five-file mean. Merged and spurious turns reduce recall and precision.	Higher
SC cpWER	Concatenate dialogue by speaker in turn order; minimize summed word-edit distance across speaker permutations using Hungarian assignment and empty streams for missing/extra speakers; divide by reference words. Lowercase, strip punctuation, collapse whitespace; five-file mean. Ignores names and timestamps; not timed tcpWER or DER.	Lower
legal_structure_f1	LCS matching with pinned class-specific normalized line keys for whereupon_open, proceedings_close, exam_header, byline, paren_other. Pool TP, output and reference counts across classes and files; 2TP / (output + reference) = 2TP / (2TP + FP + FN). A line belongs to at most one class. Oaths, appearances, stipulations and certificates are excluded; per-class support/F1 is included.	Higher
Punctuation F1	Micro F1 of punctuation slots between LCS-aligned words in final output versus reference, pooling underlying counts across files.	Higher
Word preservation	Share of formatting-input spoken words retained in formatting output by pinned alignment; five-file mean. Not an accuracy or hallucination measure.	Higher
USD/audio-hour	Recorded total provider stage spend divided by evaluated source-audio hours, not a current price quote. Partial-run spend is retained separately; historical retry/cache/token/pricing detail is incomplete.	Lower
Minutes/audio-hour	Per-file 60 × wall_seconds / audio_seconds, then median over available files. Sum executed stages per file before end-to-end aggregation; text track sums SC + format only. Coverage such as (4/5) describes timing, not output coverage.	Lower

For rt-v1.1, each repetition is first reduced to its own five-file result (speaker-turn and word-error metrics are five-file means; legal structure and punctuation pool their underlying counts across files; minutes/audio-hour are medians over files), and the headline is then the mean of repetitions 2 and 3 with their min–max range.

Quality fractions are stored as fractions in per-file data and shown as percentages on the leaderboard. Reviewed aggregate display strings preserve the site's rounding. WER/cpWER can exceed 100%. Null / N/A means unavailable or not run, never zero. The historical snapshot does not preserve all zero-denominator policies; empty classes do not establish perfect quality.

Files and schema

• results/rt-v1.1-cf310b32baa0.json: rt-v1.1 allowlisted public results (rt-public-results/2): headline aggregates with mean/min/max/n, per-repetition values, draw accounting and the SHA-256 of each of the 18 result files. Byte-identical to the leaderboard download.
• data/rt-v1.1/aggregate.jsonl: six rows (one per model) for the aggregate_rt_v1_1 configuration; numeric headline cells, track rank and configuration.
• data/rt-v1.1/per-file.jsonl: 90 rows (6 models × 3 repetitions × 5 files) for the per_file_rt_v1_1 configuration; metric values are numeric fractions or null (failed or not-run draws are null, never zero).
• data/rt-v1.1/runs.jsonl: 18 run-accounting rows for the runs_rt_v1_1 configuration: status, draws expected/succeeded/failed per stage, stage spend and result-file SHA-256.
• schemas/results-v2.schema.json, aggregate-v1.1, per-file-v1.1, runs-v1.1: strict JSON Schemas for the rt-v1.1 documents and rows.
• Legacy rt-v1 files, unchanged: results/rt-v1-1c5a96c70469.json: six ranked systems; exact canonical-site public snapshot fields, aggregate display values, per-file numeric scores, configuration and class support.
• results/partial.json: one unranked partial chain, diagnostic scores, explicit failure categories and recorded spend; failed-format metrics are null.
• data/aggregate.jsonl: six ranked rows for the aggregate configuration; aggregate display strings preserve canonical rounding.
• data/per-file.jsonl: 30 ranked model/file rows for the per_file configuration; metric values are numeric fractions or null.
• schemas/*.schema.json: strict JSON Schemas for snapshot, aggregate-row, per-file-row and partial-run documents (unknown fields forbidden).
• ATTRIBUTION.json: five individually attributed sources, rights notices, change notices and contact.
• METHODOLOGY.md: protocol, aggregation, limitations, timing/cost provenance and known reproducibility gaps.
• MANIFEST.json: relative payload path, SHA-256 and byte size for every release file except the manifest itself. Manifest checksum is recorded separately in the local review packet.
• data/corpus.jsonl: five rows for the corpus configuration, with relative audio/input/reference paths, durations, source links and audio digests. These path fields are file references, not inline transcript text or automatically decoded audio features.
• audio/: five full source MP3 files used by the frozen corpus. Dean/Degen are prepared multipart joins. Encoded durations differ slightly from integer evaluated durations; both are preserved in the corpus index.
• inputs/: the five provided transcript inputs used by the text-only track.
• references/: 15 prepared proceedings-only reference texts (proofread, SC, formatted), derived from certified originals linked below. Source PDFs are not duplicated.
• outputs/: 86 successful stage-output texts: 75 from complete systems and 11 from the incomplete system; all are model generated, not certified.
• outputs-rt-v1.1/rep<N>/<model>/<case>/{proofread,sc,format}.txt: 190 successful stage-output texts from the 18 rt-v1.1 runs, in the same per-model/per-case layout as outputs/, one directory per repetition (rep1 is legacy and unpooled). Only successful draws are included; failed or not-run draws have no file. Proofread files exist for the three Gemini models only. All are model generated, not certified.
• fixtures/synthetic-example.json: explicitly fictional task-shape example, never scored or mixed into the corpus.

rt-v1.1 snapshot rt-v1.1-cf310b32baa0 is the first 12 hex characters of the SHA-256 of the 18 [result file, repetition, batch, result SHA-256] entries; it pins the 18 rt-v1.1 result files through their individual digests, and was recomputed from those files for this release. Per-file and aggregate values were checked against the committed per-run result JSONs and the leaderboard page. Legacy: the full score-table digest is 1c5a96c70469efcfb18926d16d1636e1d7c63d97c6026acca605a581cc298c74. The snapshot suffix is its first 12 hex characters. Payload digests are separate from the score-table snapshot ID. Metric definitions match the canonical leaderboard; no scores were rerun or recomputed for this mirror.

License

Mixed file-specific rights; no blanket reuse license

Sources and licenses (quoted from the leaderboard)

The following text is quoted verbatim from https://heydonna.law/leaderboard/#sources.

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.

• James Mitchell 5.25 hours KODX Seattle · original source Creative Commons Attribution (CC BY 3.0), as published on YouTube
• Bruce Jessen 4.00 hours KODX Seattle · original source Creative Commons Attribution (CC BY 3.0), as published on YouTube
• John Rizzo 4.00 hours KODX Seattle · original source Creative Commons Attribution (CC BY 3.0), as published on YouTube
• Todd E. Dean 3.20 hours UCSF Industry Documents Library · original source Research / fair-use assertion; see below
• Karen Degen 4.38 hours UCSF Industry Documents Library · original source Research / fair-use assertion; see below

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.

No license is added by this repository: no package-wide or results license is granted beyond these source-specific notices.

Per-file attribution, rights and changes

No single license applies to this whole package. The YAML license: other is a pointer to these source-specific notices, not a grant of rights. YouTube source-audio licensing does not automatically license independently prepared references, model outputs or the scores. Separate results/code licenses remain subject to ownership review; no blanket reuse permission is claimed.

James Mitchell (rt_mitchell)

• Source audio: https://www.youtube.com/watch?v=J-J4FbTtXRY
• Creator / collection: KODX Seattle
• Certified-transcript source: https://www.thetorturedatabase.org/files/foia_subsite/mitchell_james_01.16.17_2.pdf
• Rights: YouTube CC BY 3.0, as labelled on source video. CC BY 3.0.
• Change notice: Audio preparation (extraction; multipart joining where applicable); reference preparation (certified transcript alignment and proceedings-only projection); model-generated edits in stage outputs.
• Scope: Source-audio terms do not automatically license independently prepared references or model-generated outputs. No package-wide license is granted.
• Rights, privacy and takedown: legal@heydonna.law

Bruce Jessen (rt_jessen)

• Source audio: https://www.youtube.com/watch?v=LyzaVyfqbVo
• Creator / collection: KODX Seattle
• Certified-transcript source: https://www.thetorturedatabase.org/files/foia_subsite/jessen_john_01.20.17.pdf
• Rights: YouTube CC BY 3.0, as labelled on source video. CC BY 3.0.
• Change notice: Audio preparation (extraction; multipart joining where applicable); reference preparation (certified transcript alignment and proceedings-only projection); model-generated edits in stage outputs.
• Scope: Source-audio terms do not automatically license independently prepared references or model-generated outputs. No package-wide license is granted.
• Rights, privacy and takedown: legal@heydonna.law

John Rizzo (rt_rizzo)

• Source audio: https://www.youtube.com/watch?v=z7zmQe91OSg
• Creator / collection: KODX Seattle
• Certified-transcript source: https://nsarchive.gwu.edu/sites/default/files/documents/4454543/Document-25-Salim-v-Mitchell-Transcript-of-John.pdf
• Rights: YouTube CC BY 3.0, as labelled on source video. CC BY 3.0.
• Change notice: Audio preparation (extraction; multipart joining where applicable); reference preparation (certified transcript alignment and proceedings-only projection); model-generated edits in stage outputs.
• Scope: Source-audio terms do not automatically license independently prepared references or model-generated outputs. No package-wide license is granted.
• Rights, privacy and takedown: legal@heydonna.law

Todd E. Dean (rt_dean)

• Source audio: https://archive.org/details/opioids_jzhh0257
• Creator / collection: UCSF Industry Documents Library
• Certified-transcript source: https://download.industrydocuments.ucsf.edu/h/n/n/v/hnnv0249/hnnv0249.pdf
• Rights: Redistribution for research and evaluation; fair use is asserted, not guaranteed; no ownership or blanket reuse license claimed.
• Change notice: Audio preparation (extraction; multipart joining where applicable); reference preparation (certified transcript alignment and proceedings-only projection); model-generated edits in stage outputs.
• Scope: Source-audio terms do not automatically license independently prepared references or model-generated outputs. No package-wide license is granted.
• Rights, privacy and takedown: legal@heydonna.law

Karen Degen (rt_degen)

• Source audio: https://archive.org/details/opioids_kzhh0257
• Creator / collection: UCSF Industry Documents Library
• Certified-transcript source: https://download.industrydocuments.ucsf.edu/g/x/n/y/gxny0251/gxny0251.pdf
• Rights: Redistribution for research and evaluation; fair use is asserted, not guaranteed; no ownership or blanket reuse license claimed.
• Change notice: Audio preparation (extraction; multipart joining where applicable); reference preparation (certified transcript alignment and proceedings-only projection); model-generated edits in stage outputs.
• Scope: Source-audio terms do not automatically license independently prepared references or model-generated outputs. No package-wide license is granted.
• Rights, privacy and takedown: legal@heydonna.law

Research/fair-use disclaimer and takedown

Dean and Degen are sourced from the UCSF Industry Documents Library via archive.org. Redistribution is for research and evaluation; fair use is asserted, not guaranteed; no ownership or blanket reuse license claimed. Public availability is not itself proof of redistribution rights. Attribution and source links are not a substitute for checking rights for a proposed use.

Rights, privacy and takedown contact: legal@heydonna.law. Original redactions are preserved where present. Do not use generated outputs as certified records or legal advice.

Limitations and responsible use

rt-v1.1: 5 depositions; two fresh repetitions per model; no confidence intervals. 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. Five files cannot establish performance across all legal transcription, languages, accents, courts or recording conditions. This small development sample is not contamination-free, representative of all legal work, or an internal PR-regression benchmark.

Legacy rt-v1: 5 depositions; single-run exploratory; no variance estimate. There are no confidence intervals, significance claims or winner badges. Five files cannot establish performance across all legal transcription, languages, accents, courts or recording conditions. HeyDonna sponsors this research. No Record Test Bench file, certified transcript or ASR text was used in HeyDonna prompt development, examples or training. Do not describe this sample as contamination-free.

Legacy rt-v1 timing is reconstructed in part; proofread wall includes scoring, format wall excludes scoring. Missing timing is not zero. Cache, retry, token/pricing detail, seeds, temperature, dependencies, region and hardware were not completely recorded. Complete historical prompts and a standalone runner/scorer remain unavailable in this package; exact independent execution reruns are not claimed. See METHODOLOGY.md.

Public legal source material contains real names, sensitive testimony and quoted identifiers. This dataset is not anonymized. No unrelated customer content, internal representative regression data, credentials, raw traces or private operational paths are in scope. Local provenance and automated scan evidence is recorded separately; it does not establish the absence of all personal data in public testimony.

Citation

@misc{heydonna_record_test_bench_rt_v1,
  author = {HeyDonna},
  title = {The Record Test Bench (rt-v1)},
  year = {2026},
  url = {https://heydonna.law/leaderboard/},
  note = {Snapshot rt-v1-1c5a96c70469; single-run exploratory evaluation}
}

@misc{heydonna_record_test_bench_rt_v1_1,
  author = {HeyDonna},
  title = {The Record Test Bench (rt-v1.1)},
  year = {2026},
  url = {https://heydonna.law/leaderboard/},
  note = {Snapshot rt-v1.1-cf310b32baa0; headline is the mean of repetitions 2 and 3; exploratory}
}