[Tristan-Stoltz-ERC]

Summary

Starts a grounded, executable policy-engine bake-off for Mycelix DeSci using the existing DefaultScientificAuthorizationPolicy as the semantic oracle. The comparison is split into authorization decisions, evidence qualification, and non-delegable authority/commit gates.

Added

• Frozen authorization corpus: 30 cases (18 allow / 12 deny) extracted from source revision 061e34f24d406f19f6de0e61867305b6989f455f.
• Rust adapter in scientific_governance.rs that builds typed actors, signed synthetic events, and claim projections then invokes the real DefaultScientificAuthorizationPolicy::authorize for every case.
• Metamorphic invariance test covering role-order changes, duplicate/irrelevant roles, unrelated actor-organization additions, and stable decision reasons.
• Cedar 4.14.0 schema and policy plus 30 native Cedar CLI test requests generated from the same corpus.
• OPA/Rego v1 policy plus 30 generated corpus tests and eight additional positive/negative controls.
• Separate evidence-projection corpus: 10 cases covering review qualification, organization/person source de-duplication, replication outcomes, and withdrawal. The Rust adapter rebuilds signed event streams and checks exact evidence-profile counters.
• Eleven separate outer authority/commit gates: event signature, actor binding, key validity/revocation, acting-organization membership, event-time bounds, idempotency/replay, signed receipts, receipt-chain continuity, and durable commit behavior.
• Dedicated CI workflow and benchmark plan covering semantic hard gates, adversarial mutations, fair in-process vs CLI/service comparisons, p50/p95/p99, allocations, RSS, cold/warm load, concurrency and policy-diff analysis.

Actual workflow evidence — partial, not an overall pass

The latest dedicated workflow is run 42 (ID 41047474710) on PR source head f89b4bcdcfb901961e2ca11a577775acbf53a74d. Results:

• Cedar: schema validation passed and all 30 native Cedar cases passed.
• OPA/Rego: strict checks passed and all 38 tests passed (30 corpus vectors + eight focused controls).
• Corpus integrity: denominator and adapter checks passed.
• Rust oracle / projection: blocked before our tests could run. Compilation reports E0583 because the repository's lib.rs declares and re-exports scientific_credentials, but that module source is absent. Downstream unresolved registry calls also fail. This is the pre-existing build-integrity issue tracked in #5189, with a protocol-resolution contract in PR #4968.

The missing registry must be restored through its authentic protocol and compatibility boundary; this PR does not add a stub or bypass the compiler. Consequently, Cedar and Rego are individually green against the extracted expected outcomes, but three-way Rust/Cedar/Rego parity and the 10/10 projection suite are not yet proven.

Research position

A policy-language rewrite must not weaken Mycelix's signed identity and authority kernel. The first objective is semantic parity, not language aesthetics or a synthetic latency leaderboard. The projection corpus intentionally records an open normative issue: should a non-creator individual without a declared organization count as an independent replication source? The fixture preserves current behavior; it does not endorse it as the final scientific standard.

No performance result or "beats SOTA" claim is made. The benchmark plan requires equivalent workloads and exact decisions before speed comparisons.