Gate agentic inline accuracy on the entry the client writes

[anandhu-eng]

The agentic inline accuracy gate never runs on a real submission. It looks up an accuracy entry named agentic_combined, which is the performance dataset's name in the Kimi K3 and DSV4 example configs (Qwen's is agentic_coding). The reference client never writes that name: it records the inline score under dataset_name: "performance", tagged dataset_type: "performance", whatever the performance dataset is called (execute.py, accuracy.py). For Kimi K3, Qwen3.6-35B-A3B and DeepSeek-V4.1-Flash the gate therefore only warned "No point reports a agentic_combined accuracy score", and a submission below the inline threshold passed without --strict.

Fix. The inline entry is found by dataset_type: "performance", which is how the client's own average_accuracy identifies it and how its comments ask consumers to filter. An untyped entry named performance, from clients that predate dataset_type, still counts. An entry named performance but typed accuracy does not.

New validation. An accuracy file with more than one dataset_type: "performance" entry is now an accuracy-valid error. The client scores the performance run once and rejects configs with more than one performance dataset, so no real output trips it.

Scope. The lookup is used only by the agentic inline gate. The §15 single-turn gate, SWE-bench mean-of-N and OSL range are unchanged.

Tests. The fixture now writes the client's shape. New cases cover a native accuracy_scores list, lookup by type rather than name, the untyped legacy entry, and the two entries that must not count as inline. A duplicate-performance case is added to the malformed-file tests. Full suite passes (1535), and ruff and mypy are clean.

Run against a real Qwen3.6-35B-A3B GB300 submission (9 points), the gate now reports inline accuracy at every point (56.43–56.74 ≥ 55.86) where it previously only warned.

Not in this PR.

• SWE-bench is still looked up by the name swe_bench. The client names that entry after the configured dataset, so a submission that names it differently gets a warning instead of a gate.
• The §15 gate averages every entry and does not skip a dataset_type: performance one. No current single-turn example writes one.

🤖 Generated with Claude Code

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅