A legacy review wrapper that refuses to invent authority

This executable example wraps synthetic repository-issue/review-letter metadata. It is neither a GitHub connector nor a parser for arbitrary referee prose. It claims no W3C PROV, RO-Crate, COAR Notify, Crossref, journal or repository-schema conformance. It needs no account, network connection, credentials or dependencies.

From the extracted packet root (Python 3.9+ standard library):

python3 -m unittest discover -s adapters/tests -v
python3 -m adapters.demo

Nineteen tests pass after the adversarial adapter revision. The demo writes adapters/results/demo.json, including source/fixture/test/runtime hashes, the loss report, unchanged legacy payload, explicit enrichment, and actual toy-protocol events. No message is sent and no real review is imported.

What is preserved, and what is not inferred

adapt(source) retains a deep copy of the entire JSON-value tree, fingerprints its canonical JSON representation, and maps a small known set of fields. roundtrip checks that the envelope still matches that deterministic projection, then returns the original JSON values. Envelope comparison uses canonical JSON bytes, not Python value equality: Boolean true, integer 1, and floating token 1.0 are distinct. Nonfinite numbers are rejected. The guarantee covers values, not original byte layout, whitespace, signing bytes or ordering fidelity. It is a one-format self-roundtrip, not an independent implementation interoperability test.

Each mapped field has a status: absent, null, explicit_empty, value, or blocked_by_nonobject. A missing dependency field, null, and [] therefore remain different. Empty review limitations do not mean that a check had no limitations. Unknown extension fields are preserved in the raw payload and named in the semantic loss report; they are not silently upgraded to executable protocol assertions.

The exact claim ID, declared git object reference and declared SHA-256 evidence reference are retained. The adapter rejects a mutable URL as its only version reference. It validates reference syntax and the source’s explicit immutability assertion, not the existence of a git object, source authenticity or binding to artifact bytes. The synthetic evidence file’s digest is separately checked by a test. A deliberate negative test shows that a well-shaped fabricated digest still passes the adapter: that missing verification is an explicit trust boundary.

The adapter preserves check scope, declared result, method, limits and recommendation separately. A recommendation to “accept” never supplies a missing result or overrides contradiction. Author names, reviewer names, URLs and team members never establish independence, skill, consent, authority or dependency completeness. Mapped legacy authorization/control fields remain unverified data even when present.

Two executable paths

1. Unaugmented input: execute_fixture(envelope) refuses scoped reliance because no trusted authority, control or dependency assertions exist. A pretty JSON wrapper has not made those facts true.
2. Separately enriched synthetic demonstration: a different fixture file explicitly asserts invented control groups, reviewer capabilities, decision roles, an empty dependency graph, purpose, exact scope, expiry, capacity and limitations. The adapter retains that file separately from legacy facts. It passes the declaration into the actual toy_agents.Protocol offer/check/rely/amend methods. After a synthetic material amendment, the old reliance reports reconsideration_pending.

The enriched run does not rerun the referee’s scientific method. It records the legacy check declaration under explicit fixture assumptions and charges synthetic capacity. Neither the adapter nor the fixture turns a historical letter into consent, authentication, a real editor’s decision or permission to publish private material.

This narrow demo handles one target and an explicitly empty enriched dependency graph. Nonempty legacy dependencies or nonempty authorization/control assertions are refused for reconciliation outside the demo; they cannot be overwritten with more convenient fixture assertions. Unknown/null dependency metadata can be enriched only by a separately visible explicit fixture assumption. Scope widening, contradictory checks, shared author-reviewer control and missing scientific decision roles are rejected. The runtime’s counterexamples and limitations still apply.

The public test fixture contains invented identities and example.invalid URLs. Do not adapt private review letters into a public bundle without appropriate authority and consent. This example makes no claim to implement those permissions.

See ../onboarding/legacy-adapter.md for the corresponding six-question human card.