Actuary
← All models

Model monitor

Google Gemini 3.5 Flash Lite

google/gemini-3.5-flash-lite · via openrouter

As of 2026-08-05: no alarm in 5 of 5 task families · 11 complete panels behind the record.

Served route, trailing 30 days: median latency 0.5 s · p95 0.9 s across 85 probe calls · served by Google AI Studio (85). Observed at panel time — a few calls a day, not continuous uptime.

Certified claim, per family: H0: E[panel_mean_t | prior completed panels] ≥ frozen_null_floor. Fixed panel composition does not assume or establish temporal independence. Recent panels below link to their hash-committed evidence.

One email when this model's record changes.

Task family

Structured output

No alarm
Frozen null floor
0.700
Anytime level α
0.050
Completed panels
2
CS lower
0.000
CS upper
1.000
Panel-mean trajectory, confidence-sequence band, and frozen null floorPlot generated from complete fixed panels in the epoch. The band is anytime-valid under the signed conditional-mean claim; alarms come from the e-process. Panel mean Completed panel 0.0 0.2 0.4 0.6 0.8 1.0 0 1
Complete-panel means with anytime-valid confidence-sequence bounds. The red line marks the frozen null floor — printed before the data.
Panel record — 2 completed panels
Cycle Completed Member scores Mean Evidence
44 2026-08-05 14:01 UTC 1.001.001.000.651.00 0.930 3405be86ac7b…
43 2026-08-05 12:59 UTC 1.001.001.000.651.00 0.930 6bddc02dce80…

Newest first. A failed member scores 0 and stays on the record. Each evidence link opens the panel’s full hash-committed record.

Epoch & design record
Current epoch
caa69f8d-e087-41f8-bb0a-52cd59541bb3
Opened
2026-08-05 12:27 UTC
Previous epoch
6ad2b80b-d607-4a71-aa71-8daf8e39c985
Transition
checker_refresh
Design hash
8a4f5431d708d308ccd859fc31d097a5db0297535ace2795cab39b5f087eb318
Panel hash
a37ad21aa4b8c3e16dbae958c68a70064a9d95487e131d832f2c601f7fdacf14
Pool hash
8e9496011afbb949e38b3facc664e6dd560f1c5d2155fd424fcb720d084c69ee
Panel size
5
Terminal failure
zero — a failed member scores 0 and stays on the record
Pool version
structured-output-v3
Checker version
structured_output-v2
Reference mean μ₀
0.700
Minimum effect δ
0.000
Evidence log e
-0.17 · max 0.00 · alarm at 3.00 (= ln 1/α)
Raw scores kept
10
Provider spend, this epoch
$0.0035996

The persisted monitor state is the authority a certificate replays byte-for-byte; the chart band above is a faithful replay of the same recorded panel means. Spend sums completed probe runs and includes reserved estimates where a run’s true cost was unrecoverable.

Task family

Extraction

No alarm
Frozen null floor
0.700
Anytime level α
0.050
Completed panels
2
CS lower
0.000
CS upper
1.000
Panel-mean trajectory, confidence-sequence band, and frozen null floorPlot generated from complete fixed panels in the epoch. The band is anytime-valid under the signed conditional-mean claim; alarms come from the e-process. Panel mean Completed panel 0.0 0.2 0.4 0.6 0.8 1.0 0 1
Complete-panel means with anytime-valid confidence-sequence bounds. The red line marks the frozen null floor — printed before the data.
Panel record — 2 completed panels
Cycle Completed Member scores Mean Evidence
44 2026-08-05 13:59 UTC 1.001.001.000.001.00 0.800 e1fd0942bb3b…
43 2026-08-05 12:56 UTC 1.001.001.000.001.00 0.800 b94c6e812d53…

Newest first. A failed member scores 0 and stays on the record. Each evidence link opens the panel’s full hash-committed record.

Epoch & design record
Current epoch
99138791-e8c1-4c4f-b911-abc42f9a84aa
Opened
2026-08-05 12:27 UTC
Previous epoch
b2b316ba-d856-41de-8d55-086cd7b97642
Transition
checker_refresh
Design hash
a2749265eac9731c94d1b8cada6fedc0e453bbc919053c455268b3a7eb248a61
Panel hash
1703e9221c9ce6ce1a9f69974cf0fd77056ac1ee4009dec47e3e6f23be2d4351
Pool hash
51802fc8964f7d9927bcabfae23b4bd2f0905e8977c9f8fb4ce50884ab0fbba6
Panel size
5
Terminal failure
zero — a failed member scores 0 and stays on the record
Pool version
extraction-v3
Checker version
extraction-v2
Reference mean μ₀
0.700
Minimum effect δ
0.000
Evidence log e
-0.11 · max 0.00 · alarm at 3.00 (= ln 1/α)
Raw scores kept
10
Provider spend, this epoch
$0.0008281

The persisted monitor state is the authority a certificate replays byte-for-byte; the chart band above is a faithful replay of the same recorded panel means. Spend sums completed probe runs and includes reserved estimates where a run’s true cost was unrecoverable.

Task family

Instruction following

No alarm
Frozen null floor
0.700
Anytime level α
0.050
Completed panels
3
CS lower
0.000
CS upper
1.000
Panel-mean trajectory, confidence-sequence band, and frozen null floorPlot generated from complete fixed panels in the epoch. The band is anytime-valid under the signed conditional-mean claim; alarms come from the e-process. Panel mean Completed panel 0.0 0.2 0.4 0.6 0.8 1.0 0 1 2
Complete-panel means with anytime-valid confidence-sequence bounds. The red line marks the frozen null floor — printed before the data.
Panel record — 3 completed panels
Cycle Completed Member scores Mean Evidence
44 2026-08-05 13:39 UTC 1.001.000.831.001.00 0.967 e883d313f041…
43 2026-08-05 12:36 UTC 1.001.000.831.001.00 0.967 dd3337ab25d0…
42 2026-08-05 11:36 UTC 1.001.000.831.001.00 0.967 e58eaa654870…

Newest first. A failed member scores 0 and stays on the record. Each evidence link opens the panel’s full hash-committed record.

Epoch & design record
Current epoch
8b2f4401-b21d-4102-8ff4-06da76a6f79b
Opened
2026-08-05 10:46 UTC
Design hash
4f5cc6b9ff4ce026ccab93f5a3c8241f78eb005bc8b8068e4c701e9236504412
Panel hash
e7dfea1f343f7bf37e066a1117ca21f3438e804ad61437bf72d810640cef7e71
Pool hash
9c5d065322394a6fb18c6d7f90c5bb63f00781f3e4c4aaa663536f76bb8ba7f9
Panel size
5
Terminal failure
zero — a failed member scores 0 and stays on the record
Pool version
instruction_following-v3
Checker version
instruction_following-v1
Reference mean μ₀
0.700
Minimum effect δ
0.000
Evidence log e
-0.20 · max 0.00 · alarm at 3.00 (= ln 1/α)
Raw scores kept
15
Provider spend, this epoch
$0.0005634

The persisted monitor state is the authority a certificate replays byte-for-byte; the chart band above is a faithful replay of the same recorded panel means. Spend sums completed probe runs and includes reserved estimates where a run’s true cost was unrecoverable.

Task family

Code reasoning

No alarm
Frozen null floor
0.700
Anytime level α
0.050
Completed panels
1
CS lower
0.000
CS upper
1.000
Panel-mean trajectory, confidence-sequence band, and frozen null floorPlot generated from complete fixed panels in the epoch. The band is anytime-valid under the signed conditional-mean claim; alarms come from the e-process. Panel mean Completed panel 0.0 0.2 0.4 0.6 0.8 1.0 0 1
Complete-panel means with anytime-valid confidence-sequence bounds. The red line marks the frozen null floor — printed before the data.
Panel record — 1 completed panels
Cycle Completed Member scores Mean Evidence
43 2026-08-05 13:30 UTC 0.000.000.000.000.00 0.000 d3aa5a01f463…

Newest first. A failed member scores 0 and stays on the record. Each evidence link opens the panel’s full hash-committed record.

Epoch & design record
Current epoch
d572d53f-0a30-4823-a821-0af4de95a995
Opened
2026-08-05 12:27 UTC
Previous epoch
8ef31a64-b55f-4b46-9d09-fe8fd21da992
Transition
checker_refresh
Design hash
bc3668bd52d9d48803b744c2c90f4ed816d7e394525b32b4219f6bab5c9c6df2
Panel hash
827de544c7641d140ac37e321bdb2578a41e291915f519183abbc5bd68153e81
Pool hash
cbfba66601ff51f43c83017d6de9e49988afcee72235f5dca89d705f4f8bb4cc
Panel size
5
Terminal failure
zero — a failed member scores 0 and stays on the record
Pool version
code_reasoning-v4
Checker version
code_reasoning-v2
Reference mean μ₀
0.700
Minimum effect δ
0.000
Evidence log e
0.39 · max 0.39 · alarm at 3.00 (= ln 1/α)
Raw scores kept
5
Provider spend, this epoch
$0.0005887

The persisted monitor state is the authority a certificate replays byte-for-byte; the chart band above is a faithful replay of the same recorded panel means. Spend sums completed probe runs and includes reserved estimates where a run’s true cost was unrecoverable.

Task family

Math

No alarm
Frozen null floor
0.700
Anytime level α
0.050
Completed panels
3
CS lower
0.000
CS upper
1.000
Panel-mean trajectory, confidence-sequence band, and frozen null floorPlot generated from complete fixed panels in the epoch. The band is anytime-valid under the signed conditional-mean claim; alarms come from the e-process. Panel mean Completed panel 0.0 0.2 0.4 0.6 0.8 1.0 0 1 2
Complete-panel means with anytime-valid confidence-sequence bounds. The red line marks the frozen null floor — printed before the data.
Panel record — 3 completed panels
Cycle Completed Member scores Mean Evidence
44 2026-08-05 13:34 UTC 0.001.001.001.001.00 0.800 19654523e581…
43 2026-08-05 12:32 UTC 0.001.001.001.001.00 0.800 584118de1900…
42 2026-08-05 11:33 UTC 0.001.001.001.001.00 0.800 bbd74aff3392…

Newest first. A failed member scores 0 and stays on the record. Each evidence link opens the panel’s full hash-committed record.

Epoch & design record
Current epoch
fd84db18-dbe5-4a02-8554-9e6e9f8179f9
Opened
2026-08-05 10:46 UTC
Design hash
fdf94065c55c2ee7d86701efc50be0f5abdc5f02a236f6df52c2ac6f7db801eb
Panel hash
882f3932d07aeb85c4f6406a550d8ad14512e70d3df2432fe7aab03d0508bc75
Pool hash
7ed49fa9a63ba0e0907ff8326de00a843a926618f6d0dec40aa8e491d041e99b
Panel size
5
Terminal failure
zero — a failed member scores 0 and stays on the record
Pool version
math-v4
Checker version
math-v1
Reference mean μ₀
0.700
Minimum effect δ
0.000
Evidence log e
-0.11 · max 0.00 · alarm at 3.00 (= ln 1/α)
Raw scores kept
15
Provider spend, this epoch
$0.0040367

The persisted monitor state is the authority a certificate replays byte-for-byte; the chart band above is a faithful replay of the same recorded panel means. Spend sums completed probe runs and includes reserved estimates where a run’s true cost was unrecoverable.