diff --git a/harness/release-data-implementation.md b/harness/release-data-implementation.md
new file mode 100644
index 00000000..a10b2c3e
--- /dev/null
+++ b/harness/release-data-implementation.md
@@ -0,0 +1,59 @@
+# R6 implementation freeze — schema/source 4
+
+Status: implementation starts in isolated D:/CodexBuilds/Capella2-data-integration.
+State editing acceptance stays on state-integration; merge its navigation fix later.
+The full first-release goal remains incomplete.
+
+## Outcome and contract
+
+Implement the complete data/interface contract in release-data-contract.md before
+accepting schema/source4 artifacts: nature/mechanism, bounded scalar types/defaults,
+field direction/storage, qualified scenario value paths, single specialization,
+identified interface-item allocations/protocols and binary reference associations.
+Use its exact types and validation rules. Existing PortContract and legacy
+resolution/split behavior are authoritative; the old R6 InOut paragraph is superseded
+by release-port-contract.md. R6 never creates a new InOut functional contract.
+
+Retain three crates, Project and validated atomic Command/Document boundaries.
+Use data.rs shared helpers for effective fields, item ancestry, inherited interface
+allocations, carried field paths and value validation. Reference visitors own live
+references; no second model or mutation API. Inherited rows expose original identity.
+Scenario paths disambiguate reused records and stop at Reference storage.
+
+## Sequence and ownership
+
+1. Core owner freezes public shapes/helpers, implements typed data validation,
+   reference/removal/transition maintenance and focused meaningful regressions.
+2. Source/migration owner preserves V1/V2/current3 inputs through an explicit4
+   reader; updates parser/source-project migration without flattening nesting,
+   includes or comments. Frozen strict DTOs only for changed old payloads plus a
+   closed old constructor catalog avoid cloning the whole metamodel.
+3. Desktop owner integrates typed forms, path pickers, readable nested dictionary
+   and interface-allocation rows. Root integrates OpenAPI/spec and actual REST.
+4. Full tests/strict checks, source/JSON/HTTP roundtrips, invalid/stale transactions,
+   Windows/Linux native authoring, compact UI and fresh independent review.
+
+Separate files per owner; communicate model/helper signatures before dependent edits.
+Use a dedicated target for this worktree. No simultaneous Cargo mutations.
+Do not publish or accept intermediate4 files until both payload and dictionary
+slices are complete. R3 semantic reconciliation follows this shape/reference freeze.
+
+## Required evidence
+
+- Physical flow and Operation parameters use bounded reusable types and defaults.
+- Two parameters using one Record retain distinct value paths through calls/replies,
+  reidentify/delete, graphical/source edits and initial transitions.
+- Two interfaces allocate one item with independent protocols/artifact links; removing
+  one allocation retains the item and the other allocation, and Undo restores it.
+- Specialization/inherited fields and associations preserve ownership and identity;
+  invalid cycles/defaults/paths/ranges reject atomically, incomplete drafts warn.
+- V1/V2/current3 migration preserves identities, typed ports, comments, includes and
+  nested ownership; deterministic allocation IDs and unresolved legacy annotations
+  have explicit diagnostics. GUI/source/REST save/reopen agree.
+
+Alternatives: silent changes under schema3 rejected because existing saved3 projects
+are real compatibility inputs. Whole-model frozen DTO/validator duplication rejected
+in favor of strict changed-payload compatibility adapters. Generic JSON diff or
+opaque body replacement is not the R3 solution. Instance literals, unit conversion,
+expression execution and multiple inheritance remain explicitly outside this
+bounded contract; the release requirements for categories/filters are separate.