Corrects bf17f42, which claimed a fix that did not work.
That commit added hash_surface to evals/cognition/runner.py::CaseResult. The
test imports run_eval from evals/run_cognition_eval.py, which builds
evals/metrics.py::CaseResult — a DIFFERENT class with the same name. Thirteen
files in this repo define one. I matched on the class name instead of following
the import, so the field never reached the assertion and _diff_rows raised
AttributeError: 'CaseResult' object has no attribute 'hash_surface'. That is why
the count went 99 -> 100: the same 99 divergences, plus one error.
The diagnosis in bf17f42 was right; only its aim was wrong. Confirmed directly
at the pipeline seam:
register=None surface = 'Truth is what is true. pack-grounded...'
hash_surface = 'Truth is what is true. pack-grounded...'
register=socratic_v1 surface = 'Consider the question: Truth is what is...'
hash_surface = 'Truth is what is true. pack-grounded...'
surface varies across the register axis (as it must), hash_surface does not (as
it must not), trace_hash matches. So the field added to CognitiveTurnResult in
bf17f42 is correct and stays; this moves the eval-side plumbing to the class the
test really consumes, and reverts the edit to the one it does not.
[Verification]: PARTIAL. Three registers green — socratic_v1, terse_v1,
manifesto_v1, 24 passed, where all three failed before. The full 99-register run
(~17 min) was still executing when the tree had to be committed; result to
follow, and bf17f42's DO NOT MERGE stands until it is green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FduW6Krm3PPQv3P5iwBYtx
