[pgmpablo157321]

Fixes #174

Problem

When LiveCodeBench cannot be scored, the client's LegacyMLPerfDeepSeekR1Scorer writes a non-null score from the text subsets only and sets complete: false. The CLI never read complete, so _gate_accuracy gated the partial score as the full aggregate, and it could pass.

Change

• _gate_accuracy (models/aggregate/context.py) now checks every entry for complete: false. It reports an accuracy-gate / incomplete-datasets error naming those entries and skips the threshold comparison. It sits alongside the null-score check from fix: reject incomplete accuracy aggregates with null scores #170; an entry with both a null score and complete: false gets both errors.
• New message text and fix hint in data/messages.yaml.

Notes:

• Only an explicit complete: false counts as incomplete. A missing complete field is gated as before, so older result files keep working.
• The inline performance-run entry is checked too, because its score also feeds the weighted aggregate.
• Out of scope: the agentic accuracy path (_dataset_score) also ignores complete.

Tests

• New: an incomplete DeepSeek-R1 entry (score 81.3355, which passed before) is rejected in both the keyed and native accuracy_scores layouts, and an entry with complete: true is still gated normally.
• The new tests fail without the fix and pass with it. All accuracy-gate tests pass.
• Full suite: the only failures and errors (MCP server, check-submission and runs command tests) are the same set on a clean main in my local environment, so they look like missing dependencies there rather than this change.
• Ruff was not run locally.

🤖 Generated with Claude Code

[github-actions]

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