V8 Exact Semantics Engine V1 — Independent Adversarial Imple...
Prompt
V8 Exact Semantics Engine V1 — Independent Adversarial Implementation Attack You are a hostile implementation reviewer. Your goal is not to approve V8; your goal is to find a concrete case where the reusable engine produces a result stronger, weaker, or different from the frozen finite-raster semantics. Exact target Use only repository head: f61a618416723fc32977bd7f6cbdc2fc7c528f1a Target artifacts: docs/research/V8_EXACT_SEMANTICS_CONTRACT_V1.md data/v8_exact_semantics_freeze_v1.json scripts/v8_exact_semantics_engine_v1.py tests/test_v8_exact_semantics_contract_v1.py tests/test_v8_exact_semantics_engine_v1.py Independence Do not read or use Claude's review. Do not seek consensus. Assume existing tests may be incomplete and attack what they fail to cover. Attack surface Search aggressively for: malformed/ragged/mixed RGB/RGBA acceptance; Python bool/integer corner cases; very large and negative shifts; prime/composite dimensions and gcd edge cases; W=1, H=1, 1x1, 1xN, Nx1; false TRUE_TRIVIAL or false TRUE_NONTRIVIAL; diagonal/axis-wise confusion; incorrect INTEGER_SUBTILE_REPEAT preconditions; wrong fundamental repeat width/height; exact wrap transition off-by-one or wrong-channel aggregation; mutation/aliasing of caller data after construction; hidden nondeterminism; accidental unqualified “seamless” authority leakage; complexity/pathological inputs that could turn a mathematically exact API into an unsafe operational primitive. Prefer surprising counterexamples not already encoded in the current tests. Scientific boundary The frozen contract is the target. Do not “fix” a failure by changing the definition. No thresholds, SSIM, visual tolerance, renderer assumptions, metadata labels, holdout images, or product semantics. Required output For every candidate finding provide: CLAIM — FACT / MEASUREMENT / INFERENCE / HYPOTHESIS / UNKNOWN EVIDENCE — exact code + exact frozen rule COUNTEREXAMPLE — minimal raster/arguments TEST — deterministic reproducer RESULT — expected vs actual; state NOT EXECUTED if not run VERDICT — one of: VALID_COUNTEREXAMPLE / VALID_IMPLEMENTATION_BUG / VALID_SPEC_AMBIGUITY / DUPLICATE_KNOWN_WITNESS / NON_BUG_PREFERENCE / OUT_OF_SCOPE / UNSUPPORTED_CLAIM / NO_COUNTEREXAMPLE_FOUND Rank findings by severity, but severity must not substitute for proof. If you propose a security or performance concern that does not change semantic truth, label it separately as operational hardening rather than a semantic bug. Finish with: What would falsify your strongest finding?
