# PoC-quality record and local path evidence kit

Run only the portable local method/PoC comparison with standard-library Python:

```sh
python3 -B reproduce.py --output /tmp/poc-quality-local-review-1
```

Use a new output directory. No Proof account, native binary, network, third-party
service, pytest or package installation is needed for this local replay. Python
3.10+ is intended; primary actual runtime is recorded in the captured reports.
The wrapper returns 0 when the declared contrasts are reproduced. Weak-helper
and always-deny children deliberately return 1; only the corrected public-method
PoC passes both local path/reply/state cases. That pass means viewer denial plus
editor success under the teaching policy, not an exploit or deployment proof.

Read CONTRACT.md and cases.json first. Each selected attempt uses fresh state.
Six attempts across three groups yield eight actual nested-or-top-level method
entries, not eight separate test cases. The weak viewer-created observation is
true/false/false from those same observations, not three weak passes. Full typed
replies and complete ordered note lists are retained.

## Native checker evidence is a separate retained campaign

The three native projects and raw audit results are supplied for inspection.
They used an exact retained current-source Proof binary after one offline
current input-set/byte comparison: 4,033 local inputs and 578 packages unchanged.
The original reference14 build is distinguished from the fresh reference15 audit commands. The
large binary/full checkout/full compiled-source ledger is not bundled; the
public build identity summary and exact source-region receipt bound the claim
without promising that any installed version reproduces it. To rerun native
audits you need a compatible full Proof checkout/build and prerequisites:

```sh
cd native-projects/S1
proof audit --check poc_quality_checked --stage document --no-cache --format json --fail-level warn
```

That optional command is not invoked by reproduce.py. Native snapshots expect
PASS/PASS/WARN and exits 0/0/2. Each contains one authored low-grade issue; S0
checks zero in scope, while submit intent puts S1/S2 in scope. S1's eight
general true answers are deliberately unsupported. S2 honestly retains three
gaps. Three false Web3-only answers are skips here, not verified criteria.
No native command executes this Python PoC. --no-cache bypasses lookup but may
write derived artifacts. Full toolchain fingerprints are metadata, not solver
proof runs.

## Precision note and limits

The fixed pre-run CONTRACT phrase about a default medium threshold and explicit
video policy must not be mistaken for explicit min_severity YAML: the native
proof.yaml omits min_severity, so medium is the current checker default. Video
policy off and CLI --fail-level warn are explicit. Original pre-run bytes remain
unchanged; this qualification is recorded in execution-review/provenance.

This is new author-created teaching code, not the old illustrative case, an AI
generation record, a real security incident or actual external filing. The
unchanged product guard is already correct; the repaired element is the PoC
path/claim and quality record. A Python method and role string do not establish
a real server/protocol, authentication, current deployment, all-role safety,
economic impact, durable storage, concurrency or authorized reliance.
Only two selected input conditions across three groups are covered. See
current-source-basis.json for the 11-key checker versus 10-item catalog mismatch,
actual scope/N-A/config/video/cache/threshold limits and complete official notes.
