Sitelet https://github.com/integrallis/models/commits

Commits

Commits on Oct 1, 2026

Merge pull request #229 from integrallis/feat/two-arm-qualification-worker
bsbodden
authored
c8110bb
Keep new correctness evidence in its own certified run
bsbodden
committed
7cfa0b6

Commits on Sep 30, 2026

Land the library-defaults correctness evidence for every model we publish
bsbodden
committed
169d1a7
Run the plain surface twice so a disagreement names its own cause
bsbodden
committed
91336eb
Record why an unmeasured kernel change hung the pool, and the evidence rule it broke
bsbodden
committed
a002653
Take work-proportional matrix partitioning out of the 0.3.51 release
bsbodden
committed
e5967c2
Stop a published job from shrinking the worker pool's activation gate
bsbodden
committed
fd0de37
Document the four gates between qualifying a model and publishing it
bsbodden
committed
e1d9c3b
Bump user-facing version references to 0.3.51
bsbodden
committed
1a5e713
Release 0.3.51
bsbodden
committed
ea27d64
Rewrap the deepseek2 rotary comment to the formatter's width
bsbodden
committed
9d106f5
Exercise a model the way an application would, across all three surfaces
bsbodden
committed
e6872df
Certify four more, crossing seventy-five usable models
bsbodden
committed
189034e
Certify eight more, including the models three earlier fixes were for
bsbodden
committed
cb5c665
Certify three small models, and stop the lander hiding why nothing landed
bsbodden
committed
9b91c79
Record that text-to-SQL needs its own qualification policy
bsbodden
committed
1541078
Anchor the summarization corpus, and screen every corpus the way the policy will
bsbodden
committed
e5f0b3e
Certify six more qualifications
bsbodden
committed
4c66703
Verify LangChain4j and Spring AI in one pass at the end, not per candidate
bsbodden
committed
508e832
Verify every qualified model through LangChain4j and Spring AI before publishing
bsbodden
committed
e1ea768
Grounding v23: cut the turn, and stop deleting answers that end with a closer
bsbodden
committed
08d9b5e
Certify Granite 4.1 8B
bsbodden
committed
30b6343
Certify Ornith 1.0 9B, and make landing a qualification repeatable
bsbodden
committed
5980a0c
Let a shard choose the workload per job
bsbodden
committed
4e4e9c2
Add a summarization workload
bsbodden
committed
513759b
Size matrix partitions by the work they carry, not by the pool width
bsbodden
committed
4ea4dae
Gate the shard runner on vCPUs rather than instance count
bsbodden
committed
95b3493
Certify six more qualifications from the two-arm fleet
bsbodden
committed
6d0e8d2
Grounding v22: judge the answer after a channel-closed reasoning trace
bsbodden
committed
542a98f
Do not inherit prompt templates from older shards
bsbodden
committed
5c998a7
Choose decode threads per model from the measured sweep
bsbodden
committed
338dfc0
Record the reasoning-preamble finding and its fix
bsbodden
committed
81aa947
Add a gemma4 answer-prefill template for the prose thinking preamble
bsbodden
committed
4d2395d
Run shards four wide now the vCPU quota allows it
bsbodden
committed
3270828
Automate the qualified-verdict import
bsbodden
committed
a7d6865