Skip to content

v0.1.3

Latest

Choose a tag to compare

@github-actions github-actions released this 13 Jul 09:04

What's in v0.1.3

See commit history since the previous tag for details.

Benchmark scorecard for this release

UXLoom benchmark report

Dimension Grade Evidence
Critic correctness A precision 1.000, recall 1.000, F1 1.000 over 20 golden cases (34 planted defects)
Determinism A 25 in-process runs + 2 fresh processes → 1 unique report hash
Performance A 10 screens: 0.1ms · 100 screens: 0.7ms · 500 screens: 5.5ms · 1000 screens: 6.0ms · CLI cold start 185ms
Robustness A 500 fuzzed inputs → 0 crashes (424 rejected cleanly, 76 still-valid accepted)

Per-code detection

Finding code TP FP FN Precision Recall
unreachable 4 0 0 1.00 1.00
dead-end 6 0 0 1.00 1.00
no-final-state 1 0 0 1.00 1.00
target-missing 2 0 0 1.00 1.00
screen-missing 1 0 0 1.00 1.00
target-state-missing 1 0 0 1.00 1.00
state-undesigned 2 0 0 1.00 1.00
contract-drift 3 0 0 1.00 1.00
happy-path-contract 1 0 0 1.00 1.00
contradictory-exemption 4 0 0 1.00 1.00
contrast-below-aa 4 0 0 1.00 1.00
target-too-small 2 0 0 1.00 1.00
label-overflow 3 0 0 1.00 1.00

All golden cases matched ground truth exactly.

Method

Golden projects are generated clean (zero findings by construction), then
defects are injected from a 13-entry catalog with known ground truth
(seeded RNG, fully reproducible). Detection is scored per finding code with
location matching. Determinism hashes the full report JSON. Performance is
median-of-5 on warm runs plus true process cold starts. Fuzzing applies
random structural mutations and requires clean schema rejection.