Public capability proof · synthetic code

Every finding has a line, a reproduction, and a regression test.

This small Python review demonstrates the complete evidence chain: five intentional defects in a 32-line module, corrected code, and tests that lock in each remediation. It contains no client code and makes no client-result claims.

5reproduced findings
7/7regression tests passing
32reviewed source lines
0external runtime dependencies
High

Export path escapes its base

sample_before.py:9

A parent-relative filename resolves outside the configured export directory. The fix resolves and constrains the candidate path.

Medium

Mutable default shares state

sample_before.py:12–15

Tags leak between otherwise independent calls. The fix allocates a new list for each default call.

Medium

Zero total crashes calculation

sample_before.py:20

An empty dataset raises an unexpected division error. The fix validates the total and reports a descriptive error.

Medium

Negative quantities pass validation

sample_before.py:25–27

Invalid negative values enter the domain unchanged. The fix rejects them immediately.

Low

CSV fields are not escaped

sample_before.py:31

A comma inside a value changes the apparent column count. The fix delegates serialization to the standard CSV writer.

Run the evidence locally

python reproduce_findings.pypython -m unittest -q

Validation is deterministic and reproducible. Service fulfillment uses AI assistance, automated checks, and human review; that disclosure is never hidden.