Four summarizers, one answer key, eight checks
18 calls, each graded by the same 8 code checks against a reviewed answer key.
Today's model (stand-in)
Stand-in: dvar:glm-5.3-flash
11/18
calls pass every check
1 call with a critical miss or no summary
₹0.067 a call · ₹0.0119 per minutep50 8.4 s · slowest 15.9 s
Closed-source model
google/gemini-4-argon
15/18
calls pass every check
No critical miss
₹0.522 a call · ₹0.0931 per minutep50 7.4 s · slowest 10.0 s
Open model
dvar:gemma-4-31b-it
6/18
calls pass every check
2 calls with a critical miss or no summary
₹0.041 a call · ₹0.0073 per minutep50 26.5 s · slowest 29.3 s
Open model + examples
dvar:gemma-4-31b-it + 8 examples in the prompt
14/18
calls pass every check
No critical miss
₹0.070 a call · ₹0.0125 per minutep50 16.9 s · slowest 28.1 s
Closed-source model: can replace today's model
No critical miss; at least as good as today.
Open model: blocked
- No summary on 2 calls.
- Passes 6 calls, fewer than today's 11.
Open model + examples: can replace today's model
No critical miss; at least as good as today.
Covers the call
Cover, premium and outcome are in the summary.
Today
89%
Closed
100%
Open
78%
+ Examples
100%
Figures from the callcritical
Every amount and percentage was said on the call.
Today
94%
Closed
100%
Open
89%
+ Examples
100%
No mis-selling missedcritical
Every mis-selling or data-leak moment is flagged.
Today
94%
Closed
100%
Open
89%
+ Examples
100%
No false alarms
No flag the reviewer did not find.
Today
94%
Closed
100%
Open
89%
+ Examples
100%
Disclosures right
Each disclosure marked made or not, as on the call.
Today
61%
Closed
83%
Open
50%
+ Examples
78%
Next step captured
The promised follow-up is there.
Today
94%
Closed
100%
Open
83%
+ Examples
100%
No identifiers repeatedcritical
No phone, PAN, Aadhaar, card, OTP or policy number.
Today
94%
Closed
100%
Open
89%
+ Examples
100%
Written in English
English, whatever the call's language.
Today
94%
Closed
100%
Open
89%
+ Examples
100%