[Pre-review] datalens agent (DeepSeek-V4-Pro) historical 88.7753% submission

[24kGarry]

This is a transparent pre-review request for an unchanged historical run. We are not asserting final leaderboard eligibility.

Agent configuration

• Agent: datalens agent (official-scaffold-v23)
• Team: datalens
• Backbone: deepseek-v4-pro via an OpenAI-compatible gateway
• Hints: Yes
• Tuned prompt: Yes
• Max iterations: 100
• Coverage: 12 datasets, 54 queries, 5 runs/query (270 trials)

Local result

• Dataset-stratified Pass@1: 0.8877533577533577 (88.7753%)
• Raw passing trials: 242/270
• Frozen batch: frozen-20260906T230532Z-4745e3f14a
• Results SHA-256: 95ebb22d8efcf60f84e809aa9e7c9dd903f733e999fb197f9aa6ac579bda53db

All 270 submitted answers match the corresponding final answers in the supplied traces. The score is local and subject to official re-scoring.

Disclosed limitations

• A conservative local prompt screen flagged 55 historical trials; all were classified as requiring a fresh clean run.
• 224 compact historical traces lack the newer full-result artifact manifests; missing payloads were not reconstructed or fabricated.
• The historical aggregate dirty-tree/bytecode fingerprint is not reproducible; recovered prompt-producing source exact-matched 270/270 system and user messages.

Full details are in leaderboard_submissions/datalens_deepseek-v4-pro_submission.md and the trace bundle.

Could you advise whether this historical evidence is useful for pre-review and whether a fresh clean 270-run batch is required for leaderboard eligibility?