Sliced Inference
Runtimelabs
Model
Click Success
Loose Match
Latency
$/step
RuntimeLabs SVLM
self-hosted 4B
0/22
Ready
cirrus-support.internal/tickets
∞
▦
▤
📊
♪
⚙
J. Alvarez
Sep 12
M. Chen
Sep 12
R. Osei
Sep 11
T. Nakamura
Sep 11
P. Dubois
Sep 10
Dispose ticket
✉
💬
🕑
⚑
Loading folders…
Order arrived with wrong lens coating
+
Ticket Dispose
×
📁
Customer Service
×
🔍
×
Searching…
L1
Return
Order Update
Searching…
L2
Refund Issues
Warranty Claim
Delivery Dispute
Searching…
L3
Approved
Rejected
L4
Customer ServiceReturnWarranty ClaimApproved
Select Queue
*
Searching…
Escalations
Delight
Warranty Desk
Assign To
*
Searching…
R. Fontaine
K. Whitfield
M. Adeyemi
S. Kowalski
Disposition Type
Searching…
Pending
Resolved
Escalated
Remarks
*
Reviewed and approved per policy.
Is Callback Required
Callback At
14 Sep 2026, 10:00 AM
September 2026
›
12
13
14
15
16
17
18
OK
SUBMIT
✕
queued
0.00s
flow result
—
RuntimeLabs SVLM · live log
Waiting for the run to start…
gpt-6-astra
OpenAI, hosted API
0/22
Ready
cirrus-support.internal/tickets
∞
▦
▤
📊
♪
⚙
J. Alvarez
Sep 12
M. Chen
Sep 12
R. Osei
Sep 11
T. Nakamura
Sep 11
P. Dubois
Sep 10
Dispose ticket
✉
💬
🕑
⚑
Loading folders…
Order arrived with wrong lens coating
+
Ticket Dispose
×
📁
Customer Service
×
🔍
×
Searching…
L1
Return
Order Update
Searching…
L2
Refund Issues
Warranty Claim
Delivery Dispute
Searching…
L3
Approved
Rejected
L4
Customer ServiceReturnWarranty ClaimApproved
Select Queue
*
Searching…
Escalations
Delight
Warranty Desk
Assign To
*
Searching…
R. Fontaine
K. Whitfield
M. Adeyemi
S. Kowalski
Disposition Type
Searching…
Pending
Resolved
Escalated
Remarks
*
Reviewed and approved per policy.
Is Callback Required
Callback At
14 Sep 2026, 10:00 AM
September 2026
›
12
13
14
15
16
17
18
OK
SUBMIT
✕
queued
0.00s
flow result
—
gpt-6-astra · live log
Waiting for the run to start…
Replay RuntimeLabs SVLM
Replay gpt-6-astra
Where & why gpt-6-astra fails
Knows the neighborhood, not the address
Mean IoU
0.53
(astra) vs
0.88
(ours) · 81.3% lenient collapses to 44.6% at IoU≥0.5. It lands near the element and stops there.
88–100%
Static click
12–17%
Loads / autocompletes
Fails on dynamic UI, not small UI
open_disposition
12%
success, assign-to-random
17%
· plain clicks
88–100%
. It can't know what a screenshot hasn't shown it yet.
✕
✓
Doesn't know when to wait
Racing the folder-tree load is the real product's own
#1 dispose failure
. A one-shot screenshot can't show "still loading."
acc.
tokens
More reasoning tokens ≠ more precision
Kimi:
2,028
tok/call, worst score. Ours:
~30
tok, best. Chain-of-thought doesn't recalibrate a coordinate.
gpt-6-astra
Remarks
✕
RuntimeLabs SVLM
Approved
✓
Sometimes it doesn't understand the field itself
Action precision:
100%
ours vs
82.6%
astra — roughly 1 in 6 answers gets click-vs-fill wrong, a level below "where."
1 GPU · 100 in parallel · 1,000 in RAM
GPUs
1
Serving in parallel
100
Slices staged in RAM
1,000
GPU footprint
~24 GB