Frozen coding fixture and independent grader

Seven pure-Python tasks were collected three times per model. The frozen suite.json SHA-256 is 3f7f061135313c81e9002f9971f682ee80d03ae485f566cfdb2fd17ea9b7c57a. The original suite and grader are byte-identical to the reviewed private inputs. Text markers establish collection only; executable held-out cases establish the reported functional scores.

From this evidence directory, reproduce a frozen list-pair grade with:

python fixture/grade.py native/quasar-text32k-coding-raw.json --reviewed --expected-repetitions 3
python fixture/grade.py native/swift15-text32k-coding-raw.json --reviewed --expected-repetitions 3
python fixture/grade.py --self-check

A behavioral failure returns exit status 1 and still emits its complete grade. Only the original metadata fields named in redaction-lineage.json were redacted; prompts, visible code, reasoning and case values were preserved. Grade artifact_sha256 values identify the pre-redaction native bytes; the lineage ledger maps these to the public copies.

The separately named *-coding-grade-tuple-edges.json files are post hoc input representation sensitivity analyses. They reconstruct only the stable_toposort edge pairs as tuples immediately before candidate invocation, using the same five cases and expected outputs. They do not replace the frozen list-pair grades or change the collected model responses. The prompt describes (before, after) pairs without explicitly requiring lists.

All retained sources were manually reviewed before offline execution; see source-review-notes.md. --reviewed, the AST allowlist and subprocess timeout are safeguards for this bounded fixture, not a hostile-code security sandbox. This fixture does not measure repository-level SWE performance.

These synthetic prompts and cases were authored for this campaign and follow the repository license. No external benchmark corpus was incorporated.