Data trust
Everything this application asserts was produced by a language model reading partial transcripts. This page says exactly how partial, exactly which model, and exactly where it hedged — so you can decide how much of the rest to believe.
Transcripts capture about 32% of the time these calls were said to run
Across 100 calls, coverage ranges from 13% to 83% of the stated duration. Anything said outside the captured window simply does not exist in this database — so every count in this application is a floor, not a total. A theme that appears in 15 calls appeared in at least 15 calls.
All 100 calls are nonetheless marked safe to extract from, and no transcript turn was flagged low-confidence.
Calls
100
3 Feb 2026 – 28 Apr 2026
Cited facts
2,998
backed by 6,025 transcript turns
Mean ASR confidence
92.5%
0 turns flagged low
Partial transcripts
99
of 100 calls
What the model produced
Every one of these resolves to specific transcript turns.
Turn-level sentiment scores are recorded directly against a turn and so are not counted here.
Integrity warnings
The transcript covers only part of the stated call duration, so facts are drawn from an incomplete record.
Transcript timestamps and calendar event times disagree, so absolute times are unreliable. Relative ordering within the call is unaffected.
Where the model hedged
Its own confidence, recorded alongside each fact. Nothing here is inferred by this application.
Product mentions
- High215
- Medium2
Competitor mentions
- Medium5
- High48
Action item due dates?“No date given” means nothing in the transcript pinned a deadline — not that the model was unsure.
- Low63
- Medium48
- None196
- High178
Action item owner attribution?Roster matches are anchored to the meeting's participant list; model-inferred owners are not.
- Model48
- Roster437
Speaker name matching?Initial-only matches are weaker attribution than exact name matches.
- Initial87
- Exact224
Calls where the least was captured
Facts drawn from these calls rest on the thinnest evidence in the corpus.
| Call | Date | Coverage | Captured / stated |
|---|---|---|---|
| All Hands - April Update | 26 Apr 2026 | 13% | 7 min of 49 min |
| Weekly Engineering Standup | 17 Apr 2026 | 16% | 8 min of 45 min |
| Comply v2 - Launch Day Checklist | 4 Apr 2026 | 16% | 7 min of 42 min |
| Aegis / Cobalt Software - Q2 Planning | 11 Apr 2026 | 16% | 7 min of 41 min |
| Detect Outage - Customer Impact Assessment | 12 Mar 2026 | 18% | 9 min of 48 min |
| Aegis / Silverline Brands - Comply v2 Early Access Demo | 3 Apr 2026 | 18% | 10 min of 50 min |
| Aegis / Trailhead Marketplace - Renewal Confirmation | 5 Apr 2026 | 18% | 9 min of 47 min |
| SOC 2 Type II - Final Review | 20 Apr 2026 | 18% | 8 min of 40 min |
How the facts were extracted?A single prompt fingerprint across every call means all 100 were processed identically — no drift between them.
- Model
- deepseek-v4-flash
- Extractor version
- v1.0.0
- Schema version
- v1.1.0
- Extended thinking
- On
- Prompt fingerprints?One fingerprint means every call went through exactly the same prompt.
- 1 across 100 runs
- Total tokens
- 2,231,002
- Average latency
- 114s per call
- Run
- 2 Aug 2026 – 2 Aug 2026
How the taxonomy was built?Themes, action types, metrics, sentiment drivers and competitors were derived from the corpus itself rather than supplied by hand.
- Version
- v1.0.0
- Model
- deepseek-v4-flash
- Built
- 3 Aug 2026
- Derived from
- 100 meetings
- LLM calls
- 51
- Total tokens
- 484,790
- Fingerprint
5f0ea4f9f7d0e6a4
What people have reviewed0
There is no login, so these verdicts are shared by everyone using this application.
Nothing has been reviewed yet
Open the evidence behind any claim and mark it as looking right, or flag it as wrong. Verdicts collect here.