SEPTEMBER 26–27, 2026 / LIVE FLEET AUDIT
Every answer.
Every failure.
Light questions, harder problems, context, code and grounded answers—tested through the actual ThunderLink service.
Loading verified evidence…
The fleet, model by model
Resources behind the responses
Railway service averages during this run include shared production traffic. They cannot establish the CPU or memory cost of an individual prompt. No GPU telemetry is available; unavailable never means zero.
Accessible resource samples
Protocol and limitations
Local platform regression checks
159 passed; 2 failed because the local parser dependency environment was incomplete. This does not certify the deployed build.
Read all 161 checks
Previous tests, preserved
Historical scores and this new screen use different questions and weights. They are not a controlled improvement comparison.