The mrma.journal/v1 JSONL journal is append-only. Each event has a monotonic sequence, previous
digest, current SHA-256 digest, run ID, privacy-aware timestamp, type, and bounded data. Normal mode
flushes appends. Durable mode additionally fsyncs the file and synchronizes the parent directory
where the operating system supports it.
The journal covers planning, authorization, budget reservations/updates, rounds, attempts, redirects, observations, cancellation, failure, and completion. Verification rejects duplicate JSON keys, non-finite numbers, malformed digests, sequence gaps, reorder, truncation, tampering, and unknown event types.
mrma.experiment/v9 is strict Draft 2020-12 JSON Schema with nested
additionalProperties: false. It includes run state, authorization summary, plan, budget, journal,
transport/runtime provenance, comparison resource outcomes, charset resolution, observations,
analysis, exploration/confirmation role, assurance dimensions, a complete privacy-safe effective
plan, structured limitations, mutation-pair evidence, and repeated-control pair evidence.
Cross-field constraints and verifier checks prevent decisive verdicts with incomplete sampling, research authorization bypass, insecure unacknowledged TLS, exact replay claims from semantic HTTP, false durable claims, false complete header coverage, and no-influence conclusions after comparison failure. Verification recomputes the v2 effective-plan digest, observation/pair topology, sampling completeness, changed and indeterminate counts, rates, Wilson intervals, similarity and effect summaries, outcome counts, control stability, and the fixed-sample verdict. Published experiment schemas v2 through v8 remain byte-locked; v9 is their immutable successor.
Existing v7 and v8 documents remain schema- and integrity-verifiable. They do not publish the
repeated-control pair classifications required to independently rederive control stability, so the
verification result reports statistical_derivation_verified: false for those versions. The
verifier does not infer unavailable evidence or silently upgrade a legacy conclusion.
Statistical rederivation starts from the privacy-safe public pair classifications. It checks that their topology and every reported aggregate support the recorded conclusion; it does not rerun body comparison from credentials or unbounded raw responses that are deliberately absent from a public bundle.
The evidence plan digest uses run-local fingerprints and is intended for bundle integrity, not cross-run approval. Dry-run output separately exposes a deterministic local approval-plan digest that binds exact request values, privacy mode, authorization digest, comparison policy, and selected transport policy. The approval digest is excluded from journals and shared v9 evidence.
Keyboard interruption, cancellation, authorization rejection, budget exhaustion, transport/policy
failure, and incomplete transport sampling produce v9-valid partial evidence. Status, planned and
completed rounds, stop reason, consumed budget, collected observations, limitations, and
INCONCLUSIVE remain available. Complete sampling requires completed status, exact rounds, and the
exact planned observation count.
Standard and strict evidence use per-run keyed fingerprints and reduced timing/target precision. Forensic mode retains more exact metadata and requires deliberate use. Public evidence contains a summary and digest, not the executable authorization manifest. Raw credentials, proxy credentials, authorization tokens, environment values, mutation secrets, and local paths are forbidden by model and tests. Version 9 declares cross-run correlation as partial because authorization digests, original-source digests, and journal target, origin, address-set, and effective-authority identifiers are deterministic. Mutation-delta identifiers written to the journal are run-local HMAC fingerprints; the deterministic local delta digest remains inside the authorization and transport capability boundary.
An mrma.evidence-bundle/v1 ZIP has deterministic entry order, timestamps, permissions, and
serialization. It contains plan, result, authorization summary, journal, schema, runtime, packaged
release benchmark, replay instructions, and a file-digest manifest. mrma evidence verify checks
the ZIP file set, sizes/digests, schema compatibility, result/journal linkage, observation topology,
benchmark schema, effective-plan digest, statistical derivation, and equality between the result
digest and every journal RUN_PLANNED event.
The hash chain detects modification but does not authenticate who created an entirely new bundle. MRMA does not implement a home-grown signature scheme; organizational trust roots remain deferred.