The gain is on their own grader
Siemens Fuse used NVIDIA NeMo Gym to evaluate a standard-cell characterization agent by the Liberty files it delivered. The post says the mean composite verification score across rollouts rose from about 55% to 95%. Agent turns fell from 155 to 53, about 66% fewer, and output tokens fell from about 81,000 to about 29,000, about 64% fewer. The fixed model and harness were Claude Opus 4.8 in Claude Code, across about 25 rollouts.
The page time is 16:10 UTC on October 14, which is 00:10 in Beijing on October 15.
[1]
Fail one gate and the score is zero
The reward grades the final file, not whether a single tool call succeeded. Coverage, whether validation tools can use the file, whether the output adds failures beyond the reference, and integrity multiply first, then multiply by the share of timing and power values inside tolerance. If any gate fails, the reward is 0. Integrity is there to catch reuse, truncation, or a file that was only reformatted.
The post says missing cells, missing corners, wrong slew thresholds, and timing or power outside tolerance often show up only when the Liberty file is checked at the end. The team then rewrote skill text and MCP tool descriptions and reran under the same conditions.
[1]Tuning used two cells; the library has 229
The dataset is the open GF180MCU process kit: 229 standard cells and 36 process-voltage-temperature corners, simulated with Solido AFS. Each tuning run used two cells and still covered 36 corners. The post says that subset was enough for tuning. Each run took about 1.5 to 2 hours.
The post does not give a score for all 229 cells together. It also does not compare this agent with another agent or with a human library flow. The move from 55% to 95% is a before-and-after on this grader.
[1]The method can be copied; the result cannot be stretched
The post suggests that other electronic-design and engineering-simulation teams also grade the final artifact and write their own verifier, so an agent cannot return a file that only looks complete. That is a description of the method, not a second experiment.
Standard-cell characterization happens after a foundry updates a process kit: timing, power, and noise models have to be rebuilt from the new transistor models. The post does not say any production line is already shipping libraries from this agent.
[1]