Upload source.py with huggingface_hub

source.py

@@ -39,6 +39,9 @@ _DEFAULTS = {
     # CODE is the only RANKED benchmark, so it gets the most capable pool model. Measured over the
     # whole 26-task scored pool: kimi-k3 24/26 (92.3%) vs gpt-5.6-luna's reported 84.6%.
     "code_model": "moonshotai/kimi-k3",
+    # Ordered by measured solve rate on the scored pool: kimi 24/26 first, then the models that
+    # cover its gaps. Length also caps the attempts.
+    "code_ladder": ["moonshotai/kimi-k3", "google/gemini-3.6-flash", "deepseek/deepseek-v4-pro"],
     "code_attempts": 3,                         # redraw only when the samples say the code is wrong
     "code_retry_needs_s": 220.0,                # worst observed kimi task was 156s; keep headroom
     "code_max_tokens": 16384,                   # it reasons first; 8192 truncates, 4096 returns nothing
@@ -276,7 +279,15 @@ def build_agent(weights):
             if attempt and (remaining(cfg["budget_s"]) < cfg["code_retry_needs_s"]
                             or not fits(cfg["retry_estimate_s"])):
                 break                                    # no room for another slow call
-            resp = call_model(cfg["code_model"],
+            # MODEL ROTATION, not repetition. Redrawing from the SAME model repeats its blind
+            # spots; different models fail on different tasks. MEASURED on abc399_d — the one task
+            # kimi never solves: gemini-3.6-flash 2/2, deepseek-v4-pro 1/2, gpt-5.6-luna 0/2. This is
+            # also what the #1 miner does (it runs the owner's harness cascade, which escalates
+            # ACROSS models rather than retrying one), and it is the diversity our single-model
+            # retry was missing at 25/26.
+            ladder = cfg["code_ladder"]
+            model = ladder[min(attempt, len(ladder) - 1)]
+            resp = call_model(model,
                               [{"role": "user", "content": prompt + "\n\n" + _draft_instr()}],
                               {"max_tokens": cfg["code_max_tokens"],
                                "reasoning": {"effort": "low"}})