Skip to content

Latest commit

 

History

History
73 lines (63 loc) · 4.44 KB

results.md

File metadata and controls

73 lines (63 loc) · 4.44 KB

Recorded NPU replay results

The whole-provider replacement reduces native logical statements from 264 to 246 and physical lines from 799 to 729, including the entire new typed helper. The separate task 6.3 target requires a positive production statement reduction and comparable source metric definitions. Internal call layers remain unavailable. Task 6.1's stricter all-metric preparation target remains incomplete; task 6.3 neither repairs nor replaces that historical declaration.

Native whole provider/helper metric Baseline Candidate Delta
physical_lines.v1 799 729 -70
logical_statements.v1 264 246 -18
decisions.v1 54 46 -8
functions.v1 36 34 -2
internal_call_layers.v1 unavailable unavailable unavailable

Native definitions are tree-sitter-python@0.25.0 with lesscode-core@0.1.0;ast-grep@0.51.0. Selected scope identities and exact inventories are retained in the native reports. Wider copied source and test accounting reports physical lines 24,319 to 24,249; logical statements 9,388 to 9,370; decisions 1,918 to 1,910; functions 1,108 to 1,106. Its unknown call layers remain unavailable. Category totals distinguish production, tests and dependencies. No dependency/configuration integration changes support the replacement.

The archived NPU01 selected execution/check-family scope used a different boundary and analyzers: physical lines 171 to 101, Radon LLOC 68 to 47, AST decisions 16 to 8, and functions 8 to 6. These are historical selected-family figures, not native whole-module counts. The archived whole modules had the same physical lines 799 to 729, but their Radon LLOC 290 to 269 is not native logical_statements.v1. The native report does not rename or substitute those definitions.

A first adapter run was incomplete despite passing software checks: four candidate relationship reviews named the added helper path rather than its baseline ownership key. first-incomplete.json preserves the exact reasons, source identities and check definition. The adapter now maps that helper to the CPU ownership key whose replacement explicitly includes it. A new trial freezes the corrected runner and repeats observations. No core or upstream product change was needed.

Raw per-case digests and limited exact mismatch witnesses preserve the finite observations. Every check result binds the actual source identity, frozen trial, check definition and raw artifact hashes. A collector's zero exit is not used as differential success. Hardware execution, accelerator validation and performance claims are outside the evidence domain. Hashes establish consistency, not externally authenticated execution or universal behavioral equivalence.

Fresh observation Baseline Candidate Dtype mutant
Frozen differential corpus 3,168 / 0 mismatches 3,168 / 0 mismatches 3,168 / 572 mismatches
Unchanged software tests 339 passed 339 passed not run
Selected mypy modules 2 passed 3 passed not run
Shared-safety audit passed passed not run
Native verdict eligible accepted rejected-behavior

The mutant report retains all three missing candidate checks alongside the known behavior failure. Missing evidence cannot turn that failure into acceptance.

Integrated manipulation Native outcome Exact admission reason
Move helper into support/moved_helper.py invalid, exit 3 declared checked_run.py replacement source missing or excluded
Move production helper into undeclared test path invalid, exit 3 declared checked_run.py replacement source missing or excluded
Clear ownership replacements incomplete candidate owned implementation omitted
Remove required candidate type result incomplete types/candidate: required check result missing
Bind candidate differential to baseline identity invalid, exit 3 differential/candidate: source identity mismatch

The helper/test-path variants physically move the source while retaining the frozen categories and replacement declaration. Native admission rejects those attempts before report generation; exact command, exit code, stdout and stderr are retained instead of fabricated comparison results. Successful relocation would require declaring and counting the replacement under a newly valid scope. The fixtures deliberately test attempted omission/reclassification, not runnable adoption of these altered copies.