ponder-lab / Hybridize-Functions-Refactoring / #2276

80%

main: 80%

Build Type

Pull #712

github

Committed by khatchad

Commit Message

Add audit-barren.sh: list barren candidates for a human accuracy pass

The barren precondition (issue 709) hard-blocks or de-hybridizes a candidate
that performs no tensor computation, so a false positive de-optimizes real
tensor code. This script surfaces the set to review: it runs the headless
evaluator with alwaysCheckTensorComputation on (populating functions.csv's
`tensor computation` column for every function, not only P1/P6 candidates),
then lists the tensor-parameter functions computed as barren, grouped by
subject with source locations.

Over the 6-subject corpus at Ariadne 0.52.10 it flags 19 candidates, of which
7 are false positives (six sub-layer-dispatch forward passes, #714; one
keras.backend case, #716) and 12 are genuinely barren (proto builders,
plotting, cv2, shape checks, distribution wrappers). It surfaced SSD.call,
which the failed-precondition CSV missed because it is not a P1/P6 blocker.

Refs #709.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FuhM8SsTRB6BraLt3Ph7ZC

Pull Request Pull Request #712: Add benefit preconditions for barren functions

Coverage Stats

122 of 130 new or added lines in 5 files covered. (93.85%)

11 existing lines in 1 file now uncovered.

1757 of 2200 relevant lines covered (79.86%)

0.8 hits per line