The last gate before a fine-tune ships
Promote DeskShould this fine-tuned checkpoint ship - and in what format?
Paste the checkpoint's evals. Free in your browser, the four-stage gate runs: data quality, capability drift against a 1 / 5 point budget with honest error bars, a paired arena with a real confidence interval, and a canary for production - ending in PROMOTE, REJECT or RERUN with exactly one fix. Once promoted, it picks the export format for your hardware and workload. A run writes the promotion report, or the export plan with its smoke test.
1 The checkpoint's evidence free
2 The gate free
Capability drift, with 95% error bars
Paste the drift suite above.
Paired arena
What the gate flags
3 The export free
4 The report or the plan metered
Pricing as soon as the session is ready...
Sign inThe reply could not be read. Here it is as it came back:
Stage by stage
Remediation
Caveats
Steps
Smoke test
What failure looks like
The page's check of this reply
5 Saved checkpoints
How it works
Drift is the drop from baseline to checkpoint in percentage points - absolute accuracy, never relative percent. Each benchmark is banded by the source skill's budget: up to 1 pt is noise, 1 to 5 pts needs a seed-variation rerun (which must land back at 1 pt or less), over 5 pts is a hard fail with no exception for task gains. Each row carries its 95% half-width, 1.96 x sqrt(p(1-p)/n); when a breach of the 5 pt line is smaller than that half-width the verdict is REJECT (uncertain). A 200-item suite resolves about +/-6 pts at 70%; a 2.5 pt half-width takes about 1,300 items.
The arena counts ties as half a win and needs 50% + your margin with the 95% Wilson interval clear of 50%. Stages gate in order - data, drift, arena, canary (production only; local deployments stop at stage 3) - and a REJECT names one fix: for a drift breach, the next untried rung of the forgetting ladder (swap the replay mix, lower the learning rate, fewer epochs, a smaller rank).
The export pick follows the format map: FP8 on FP8-capable GPUs, AWQ INT4 on older ones, GGUF Q4_K_M from an imatrix for llama.cpp, NVFP4 only for Blackwell at scale and never on GB10, and never INT4 for long-context, code or math work (W8A8 or GGUF Q8_0 instead). Sizes are approximate bits per weight times parameters.