Validation¶

Correctness is established against independent oracles and real-world artifacts, not fixtures we both encoded and graded ourselves. Each claim is labelled by tier — the trustworthiness of the check:

Tier 1 — an independent third party authored the artifact and the answer key, or it is real-world data validated by an independent oracle.
Tier 2 — real engine/tool output confirmed by an independent oracle, but we chose the scenario, so it can miss real-world quirks.
Tier 3 — fixture and expected answer both authored here; legitimate only where no external oracle exists (detection rules defined by spec, robustness properties), never as the sole check of a value-producing path.

The UDIF container and its koly trailer are a reverse-engineered Apple format; dmg-core's block codecs and the trailer layout are cross-checked against hdiutil, Apple's own tooling. Provenance and hashes for every committed fixture are in core/tests/data/README.md.

Reader codec decode — tier 2¶

dmg-core decodes each UDIF block codec against a real hdiutil-made DMG, with hdiutil (a different codebase) as the independent oracle: ADC, zlib (UDZO), bzip2 (UDBZ), LZFSE (ULFO), and LZMA (ULMO). A byte-match between the pure-Rust decoder and hdiutil's output is genuine cross-tool agreement. Tests: core/tests/real_images.rs.

Sparse image readers — tier 2¶

SparseImageReader (.sparseimage, sprs band table) and SparseBundleReader (.sparsebundle, Info.plist + hex-named band files) are validated against a self-minted hdiutil image flattened by hdiutil convert -format UDTO, with a full-image SHA-256 match asserted (core/tests/sparse_images.rs, env-gated on a macOS host). This is tier 2, not tier 1: the image is self-minted (we chose the scenario) and the oracle shares hdiutil's codebase with the minter, so it proves the reader agrees with hdiutil — not that it handles the real-world quirks a captured acquisition (e.g. a Sumuri RECON image) might carry. The full-image SHA is the load- bearing check: it caught a band-table direction error that per-offset spot-checks passed (the sprs table is indexed by physical slot, value = virtual band + 1). Committed synthetic fixtures specify the parse behaviour for CI; a fuzz target (fuzz_sparse) asserts the header parser never panics on arbitrary bytes.

Analyzer koly audit — specificity (tier 2) + sensitivity (tier 3)¶

dmg-forensic parses the koly trailer and grades:

Code	Meaning
`DMG-KOLY-SIGNATURE-INVALID`	trailer signature is not `koly` (not UDIF, or overwritten)
`DMG-KOLY-VERSION-UNEXPECTED`	version field is not the documented 4
`DMG-KOLY-DATAFORK-OUT-OF-BOUNDS`	data-fork offset+length runs past the file end
`DMG-KOLY-XML-OUT-OF-BOUNDS`	XML block-table offset+length runs past the file end
`DMG-KOLY-TRAILER-TOO-SMALL`	file is smaller than a 512-byte koly trailer

Specificity (tier 2): a self-minted hdiutil DMG audits clean — zero anomalies (a_real_apple_made_dmg_audits_clean). The real artifact is the independent oracle for "a well-formed DMG produces no false positives".
Sensitivity (tier 3, by nature): the out-of-bounds and bad-signature rules detect deliberate corruption/tampering that does not occur in benign DMGs, so there is no third-party known-answer artifact. Correctness is defined by the rule plus the koly layout; the crafted trailers in forensic/tests/audit.rs specify behavior. Each finding carries the offending value + field as evidence ("consistent with", never a verdict).

Robustness¶

The koly trailer is read through bounds-checked integer readers that yield 0 out of range rather than panic; every offset/length pointer is range-checked against the file length before use; and a large image is never read whole — audit_reader seeks to the tail and reads only the 512-byte trailer.