| Today there's one autosave path: any change to reviews, utterances, or the trim marks hasUnsavedChanges dirty, a $effect waits 1s of inactivity, then saveDraft() fires one PUT review_draft with the entire current reviews array + entire utterances array + trim. The response is reconciled back against snapshots (savedReviewSnapshot/savedUtterancesSnapshot) taken at send-time to avoid clobbering clicks made mid-request (MeetingReview.svelte:1003-1052). | |
| Two things make this fragile: | |
| - Snapshot drift: reviewSnapshot() is a string built from decision:start_at:end_at. Anything that changes server-side but isn't reflected identically in that string (rounding, reordering, a field the reconciliation didn't anticipate) makes reviewDecisionKey never re-equal savedReviewSnapshot, so hasUnsavedChanges stays permanently true โ the 1s effect refires forever โ indefinite "Savingโฆ". | |
| - Backend semantics assume a full snapshot. apply_phi_reviews/apply_bystander_reviews treat absence from the array as an explicit signal ("this m |
| [ | |
| { | |
| "meeting_id": 862, | |
| "original_transcript": [ | |
| { | |
| "end": 9.56, | |
| "text": "ใใใซใกใฏใๅปๅธซใฎๅณถๅใงใใใใๆฌใใๆฌไบบๆง็ขบ่ชใฎใใใใซใใผใ ใจ็ๅนดๆๆฅใใ้กใใใพใใ", | |
| "start": 0.0, | |
| "words": [ | |
| { |
Dataset #6 "Sample Import 2026-09-09 13:41" โ 7 meetings.
Each meeting's original transcript is corrected two ways and compared:
gemini-3.1-flash-litetranscript_issues (claude-opus-4-6)For every meeting: first the plain-language list of what each model changed,
Dataset #6 "Sample Import 2026-09-09 13:41" โ 7 meetings.
For each meeting the original transcript is corrected two ways: applying the issues found by gemini-3.1-flash-lite and, separately, the stored Opus (claude-opus-4-6) transcript_issues. Each is rendered as human-readable text ([Speaker] ..., consecutive same-speaker turns merged, like Meeting#formatted_transcript), then diffed.
The diff is gemini-corrected โ opus-corrected, character-level: [-only in gemini's version-], {+only in opus's version+}. Lines with no markers are unchanged context.
| Removed. 57 tests, 4 skips (pending), 0 failures. | |
| โโโโโโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโโโฌโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ | |
| โ Original Phrase โ Translation โ Actual Split โ Correct Split โ Incorrect Split Translation โ | |
| โโโโโโโโโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค | |
| โ โ "I'm Shimada, your โ ๆ ๅฝ | ๅป | ใฎ โ ๆ ๅฝๅป | ใฎ | โ "in-charge" + "medicine" + "'s" + "Shimada" + "am" โ the title "attending โ | |
| โ ๆ ๅฝๅปใฎๅถ็ฐใงใ โ attending physician" โ | ๅถ็ฐ | ใงใ โ ๅถ็ฐ | ใงใ โ physician" collapses into "in-charge" plus a bare, meaningless "medicine" โ | |
| โ โ โ โ โ character โ | |
| โโโโโโโโโโโโโโโโโโโโผโโโโโโโโโโโโโโโโโโโโโโโโโผโโโโ |
What it costs to run one meeting through the transcription pipeline, measured against meeting 850 with live LLM calls on 2026-09-09.
Meeting 850 โ online GP consultation (ja-JP, 5.0 min): patient with recurrent unilateral pulsatile headache since their early 20s, now increasing in frequency, with visual aura, nausea, photophobia/phonophobia, and sleep-deprivation / red-wine triggers. Doctor rules out red flags, does a brief video neuro exam, diagnoses migraine with aura, prescribes a triptan + antiemetic + prophylactic, and
| require "google/cloud/ai_platform/v1" | |
| module ClaudeClient | |
| DEFAULT_MODEL = "claude-opus-4-6" | |
| # us-east5 is a region-pinned endpoint with a data residency guarantee, | |
| # unlike the "global" endpoint which may route requests to any region. | |
| LOCATION = "us-east5" | |
| ANTHROPIC_VERSION = "vertex-2023-10-16" | |
| MAX_TOKENS = 8192 |
Generated 2026-08-07 13:40 UTC from training/eval_history.json. Re-run python training/generate_report.py after evaluating new checkpoints to refresh this report.
This report evaluates whether the synthetic mozu-research-corpus transcript dataset carries
enough real signal to train a model to perform a clinically meaningful task, by measuring how
much fine-tuning on it improves an open-source model's output on a held-out test set. The
underlying question is about dataset quality, not model selection: if fine-tuning on this
| #!/usr/bin/env ruby | |
| require_relative "../config/environment" | |
| def rss_kb | |
| `ps -o rss= -p #{Process.pid}`.to_i | |
| end | |
| $measure_depth = 0 |
Source: https://contributors.rubyonrails.org/contributors/bogdan-gusiev/commits (#76 all-time contributor, 143 commits, 2011โ2026)
pluck, and improved #last to leverage SQL-level ordering.Overall focus: code quality, performance, and bug fixes rather than new feature work.