Bump sampling temperature 0.3->0.5 -- single-variable test on v11 (score 0.1303, best-of-day banked)

Switch to mild sampling (temp=0.3) instead of greedy -- single-variable test on v10 (score 0.0786, best-of-day banked)

Trim SAFETY_MARGIN_SECONDS 180s->90s -- single-variable test on v9 (score 0.0786, last sub of the day)

Bump MAX_NEW_TOKENS_FULL 1536->2048 -- single-variable test on v8 (score 0.0786)

Revert capitalization-preservation guard (regressed) -- back to v8 baseline (score 0.0786)

Add capitalization-preservation guard -- single-variable test on v8 (score 0.0786)

Revert morpheme-hyphen guard (regressed) -- back to v8 baseline (score 0.0786)

Guard against morpheme-hyphen leakage into FINAL ANSWERS -- single-variable test on v8 (score 0.0786)

Add context-length truncation guard -- single-variable test on v7 (score 0.0786)

Revert repetition_penalty to 1.15 (1.08 collapsed EM to 0) -- back to v7 baseline

Tune repetition_penalty 1.15->1.08 -- single-variable test on v7 (score 0.0786)

Bump MAX_NEW_TOKENS_FAST 512->768 -- single-variable test on v6 (score 0.0786)

Add bare-form output constraint (no quotes/punctuation/alternates) -- single-variable test on v5

Revert self-consistency (score collapsed to 0.0) -- back to v5 baseline

Add self-consistency (N=3 sampled + majority vote) with correct time-budget accounting

Relax FINAL ANSWERS marker regex -- was silently failing to match if model put an answer on the same line as the marker

Revert exact-match grading note -- regressed EM to 0.0, restore v4 baseline (score 0.0445)

Add exact-string-match grading note (isolated test vs v4 baseline)

Add mild repetition_penalty=1.15 (no ngram ban) -- retry repetition-loop fix without banning literal repeats

Add explanation field for Human Evaluation Challenge track

Bump MAX_NEW_TOKENS_FULL 1024->1536 (isolated test vs v2 baseline)

Revert step-1 trim -- regressed EM to 0.0, restore v2 baseline (score 0.045)

Trim step-1 to terse notation, dedup patterns (isolated test vs v2 baseline)

Add morpheme-segmentation instruction to step 1 (isolated test)

Revert count_expected_items dispatch (regressed chrF in isolated test)

Add task_type-aware count_expected_items (isolated test)

Revert repetition guard (regressed exact_match in isolated test)

Add repetition guard to v1 (single-variable test)

Revert to v1 script (score 0.0271) after two regressions

Add repetition guard, trim step-1 to skip already-logged patterns

Fix per-row token budget bug (comparing against wrong variable caused mid-generation cutoffs), trim SYSTEM prompt, task_type-aware count_expected_items

Fix count_expected_items to be task_type-aware (was returning 0 for match_letters/fill_blanks, disabling the pad/truncate safety net)

Upload script.py with huggingface_hub

Upload script.py with huggingface_hub