MODEL EVIDENCE / OpenRouter #4
The model identity, provider record, and independently verified benchmark results in one place. The number above is OpenRouter discovery order, not benchmark rank.
deepseek/deepseek-v4-flashDEEP BENCHMARK
Counts stay visible for auditability. A score appears only when the Most Annoying 50-trap view satisfies the independent checks.
MOST ANNOYING BENCHMARK
—/ 50Incomplete run · 32 verified mistakes · 6 clean results · 12 unresolved
Inspect run receiptPack 1: 7 verified mistakes, 1 clean results, 2 unresolved; Pack 2: 4 verified mistakes, 1 clean results, 5 unresolved; Pack 3: 6 verified mistakes, 1 clean results, 3 unresolved; Pack 4: 6 verified mistakes, 2 clean results, 2 unresolved; Pack 5: 9 verified mistakes, 1 clean results, 0 unresolved
Trend begins after a complete cycle
Model runs, then future slots.
Only receipts from the Most Annoying view are listed here. Tolerable-view receipts are separate and are not added to this score.
T001The model's exact answer did not match the frozen expected output.
T002The model's exact answer did not match the frozen expected output.
T004The model's exact answer did not match the frozen expected output.
T005The model's exact answer did not match the frozen expected output.
T006The model's exact answer did not match the frozen expected output.
T007The model's exact answer did not match the frozen expected output.
T010The model's exact answer did not match the frozen expected output.
T025The model's exact answer did not match the frozen expected output.
T026The model's exact answer did not match the frozen expected output.
T028The model's exact answer did not match the frozen expected output.
T030The model's exact answer did not match the frozen expected output.
T041The model's exact answer did not match the frozen expected output.
T042The model's exact answer did not match the frozen expected output.
T044The model's exact answer did not match the frozen expected output.
T045The model's exact answer did not match the frozen expected output.
T048The model's exact answer did not match the frozen expected output.
T050The model's exact answer did not match the frozen expected output.
T061The model's exact answer did not match the frozen expected output.
T062The model's exact answer did not match the frozen expected output.
T064The model's exact answer did not match the frozen expected output.
T065The model's exact answer did not match the frozen expected output.
T067The model's exact answer did not match the frozen expected output.
T069The model's exact answer did not match the frozen expected output.
T081The model's exact answer did not match the frozen expected output.
T082The model's exact answer did not match the frozen expected output.
T083The model's exact answer did not match the frozen expected output.
T084The model's exact answer did not match the frozen expected output.
T085The model's exact answer did not match the frozen expected output.
T086The model's exact answer did not match the frozen expected output.
T088The model's exact answer did not match the frozen expected output.
T089The model's exact answer did not match the frozen expected output.
T090The model's exact answer did not match the frozen expected output.
OFFICIAL SOURCES
DeepSeek describes this model family in its official updates and transparency material. Capabilities stated there are provider claims, separate from our benchmark.
Important: Statements in linked publisher material are provider claims. They are not findings of this benchmark.
METHOD & LIMITS
Every public leaderboard score is the latest Most Annoying 50-trap view. The 100-task run receipt is the wider audit bundle, not the benchmark denominator. Benchmark rank #1 means fewest verified mistakes among complete runs; the list can still be sorted by most mistakes for readability. OpenRouter feed rank is separate discovery metadata and is kept out of the benchmark rank column. Two-week average reporting will use the last four complete 100-task cycles when enough cycles exist. Individual verified mistakes, called catches internally, require primary and shadow checks to reproduce the exact mismatch from the recorded reply against the frozen expected output. Retrieval is not signature verification; auditors should recompute receipt roots and verify signatures against the published issuer key. Routing errors, timeouts, incomplete answers, and missing evidence stay outside the ranked score. Do not judge any model solely from this benchmark. Ranked complete runs must use the same 50 traps, scoring rules, model settings, and retry policy.
For each listed Most Annoying mistake, primary and shadow verification must both reproduce the exact mismatch from the recorded reply against the frozen expected output. Otherwise the item remains unresolved.
The rankable score is the 50-trap Most Annoying view. The run receipt may say 100 tasks because it covers the wider audit bundle; that object is evidence, not a separate leaderboard denominator.
Independent verification means recomputing the receipt root and checking the signature against the published issuer public key. OpenRouter rank stays discovery metadata and does not affect benchmark rank.
Claim ceiling: OpenRouter rank and metadata describe the captured discovery feed. Provider claims are separate from this benchmark. Scores describe only the most recent public test pack and should not be used alone to judge the model generally.
QUOTE THE RECORD
GET A SECOND EXPLANATION
Copy a bounded prompt that includes this model page and its claim ceiling.