The 2026-06-04 sealed-breach post-mortem proved the 50-case train_sample has ZERO
predictive validity (its 4 "correct" are overfit; they hid a 5-wrong sealed breach).
This adds the instrument we never had: 500 real GSM8K cases CORE was NOT built on —
the train split minus the 50 train_sample, deterministic sha256(question) sort.
Same scorer as train_sample + the sealed lane, so the three are directly comparable:
train_sample(50): 4/0/46 holdout_dev(500): 0/0/500 sealed test(1319): 0/0/1319
Real GSM8K capability is 0%. The 4 train "correct" generalize to NOT ONE of 500
held-out cases. wrong=0 holds (refuses, never confabulates).
- evals/gsm8k_math/holdout_dev/v1/: cases.jsonl (500), runner, report (0/0/500), README.
- tests/test_holdout_dev_lane.py: floor (wrong==0, forever) + baseline snapshot (0/500).
Discipline: iterate here (open, large enough to resist trivial overfit); the sealed
test stays the final arbiter. wrong=0 is the floor; correct rising is the goal;
"refuse everything" is the FAILING baseline to beat, not a pass. Non-serving (eval only).
