The report is an alpha contract. Its JSON Schema is
schemas/report.v0.1.json.
plan identifies the committed test revisions, selected source files and their
SHA-256 hashes. environment describes the supervisor process and is explicitly
labeled with environment_role. target_runtime and each run's runtime record
the actual Python executable, Python version, and pytest version used for tests.
This is provenance, not a complete fingerprint of all installed dependencies or
external state. Replay requires the same recorded supervisor and target metadata
and the original Git revisions; it does not provision or reproduce an environment.
mutants_completed counts mutation pairs attempted, including pairs whose
outcomes were inconclusive. Always read complete, the verdict, and each outcome
together. complete=true means the entire selected plan and controls yielded
valid evidence. It does not mean all possible mutations were selected.
A lost_kill is confirmed with three pairs and final controls. A gain is labeled
gained_kill_provisional because it has only a single paired observation. Partial
confirmed findings may remain visible in incomplete_with_witnesses. Failed or
missing final controls downgrade all findings to unconfirmed_control and the
overall verdict to inconclusive.
Reports and bundles contain source snippets, test failures, runtime paths, and captured output. Review them before sharing. A bundle checksum detects accidental corruption; it does not authenticate its creator. Bundles are trusted local inputs, not portable safe executables or self-contained environment archives.
External configuration records manifest_origin=external, exact UTF-8 text in
manifest_snapshot.text, and its raw-byte SHA-256. Replay checks that snapshot
against both its digest and manifest_hash; the original external file is not
needed. A base-committed manifest retains the default policy.
test_diffs records changed-test unified diffs, full generated-diff SHA-256 and
truncated. Each attempted mutant has diff, diff_sha256, and
diff_truncated. Display text is bounded to 20,000 characters per diff; the hash
identifies the complete generated diff. Unterminated lines have standard newline
markers. These diffs are local evidence and can contain source text.
Actual pytest runs record started_at_utc, finished_at_utc, and monotonic
duration_seconds. A not_completed placeholder represents an execution that
never started, so it has no artificial execution timestamps. limits.suite_calls_max
is the conservative bound 6 + 6 × max_mutants, not a measured runtime estimate.
testdelta export REPORT_DIRECTORY --output NEW_FILE writes a separate
allowlisted summary. It omits raw diagnostics, source/config/diff text, absolute
paths, free-text reasons and arbitrary future fields; relative locations and
hashes remain. The automatic GitHub job summary is more compact and also omits
those raw fields. Full local reports and replay bundles are unchanged. An export
is not a replay bundle, an anonymizer, or an independent attestation of its input.