# EM-XCT-09 S1 reconciliation — authored by Codex

Both actual browser studies are complete. Claude Opus5 High returned27,269 characters; Grok Heavy returned30,126. Both complete copied final responses were read. Provider self-reports about sources are not independent evidence of full access. Neither executed a validator or verified byte hashes. These are design studies, not implementation acceptance.

## Adopted boundary

Publish a bounded original Enterprise Classification Review companion after implementation and separate frozen audit. Packet and Assessment are local entities; CodeReference is an embedded value; release/profile/assignment/crosswalk inputs are explicitly immutable snapshots of externally owned semantics. Local profile constraints implement an original finite category-set dialect. The original ClassificationScheme, CodeAssignment, SemanticMapping and ProfileConstraint contour needs remain mapped/dispositioned, not silently erased or all declared complete.

Reuse KNW-018 and XCT-020 by semantic reference at the exact verified versions/digests. Do not inherit their whole schemas, canonicalization, legal claims, model-ID reasoning or runtime readiness. The runtime's KNW→XCT parent metadata remains a held navigational signal; no subtype/import is inferred. Both public packages contain only AGENTS, specification, publication manifest and human page, without executable instance validators/fixtures.

## Decisions and disagreements

1. Both support a pure offline assessor, explicit questions, no live classification writes and no model installation. Adopt. Host custody, authenticated actor/time, selected authoritative evidence and mapping-approval competence remain external; input strings cannot authenticate themselves.
2. Claude requests profile-only mode. Adopt, including empty0..0 profiles and legitimate multiple independent categories. The latter is an original all-member category-set contract, not FHIR CodeableConcept translation semantics.
3. Claude proposes required-only binding. Adopt first release; other strengths are explicitly unsupported. This is a local constraint, not a claim that other strengths are invalid in FHIR or other systems.
4. Both require qualified literal codes, complete frozen enumerations, pinned parent/release identity, stricter-only restriction, historical/current distinction and no score selection. Adopt. Completeness remains a trusted publisher/host declaration, not a consequence of SKOS or a digest.
5. Both correctly warn that exactMatch may have multiple equivalents and XKOS supports n:m/empty sides. Yet their field tables collapse mapping sources/targets inconsistently. Preserve full source and target arrays and mapping context/capture in a CrosswalkSnapshot. This is a snapshot value, not a new mapping-table master, so Grok's concern about a third master is addressed without losing provenance.
6. Grok's final sentence permits a single target after profile filtering; this conflicts with its own invariant15/case6. Retain ALL applicable alternatives before profile eligibility. Exclusion never clears ambiguity. Multiple equivalent approval rows for the same sole target may remain evidence, but any live nonexact, contested, unapproved or compound row prevents a candidate in this strict local release. An output candidate is never an assignment or approval.
7. Reject both no-tools-audit suggestions to re-fetch/hash sources: the separate audit will use frozen supplied bytes and no external tools. Codex already verified exact public/local/runtime spec bytes. Auditors will see exact read/provenance limits.
8. Grok compares spec SHA256 against synthesisSha256 and calls them a mismatch. These hash DIFFERENT artifacts. Exact spec-byte hashes match public runtime/local copies; synthesis hashes are distinct metadata in publication.json. This is not a release-integrity failure. Preserve the raw comment and document this correction rather than silently changing it.
9. Claude calls closeMatch intransitive. W3C does not declare it transitive; this does not assert intransitivity. Adopt the precise wording. No chain execution in the reference does not deny SKOS entailments.
10. Claude proposes JCS and digest-only packet identity. Use an explicitly named restricted original JSON encoding instead, and stable packet identity plus immutable revisions/content digest. No RFC8785/JCS conformance claimed. Assessment identity also pins evaluator build and retained approval evidence, because approval context can change the same packet's result.
11. Underlying source semantics remain with original owners. The host owns capture, packet and assessment. Reject the memo's blanket statement that host owns every snapshotted fact. An aggregate assessment is not a federation projection of several subjects.
12. Historical membership may be assessed against a pinned retired code, with its explicit effective interval. New candidate eligibility requires active/selectable status under the local contract. Reject a blanket profile-definition ban on ever listing retired/nonselectable concepts, which would destroy historical evidence. Definitions and whole snapshot/source pins help detect changed bytes; they do not prove true meanings.
13. Initial supported migration recommendation is one source assignment/category and direct1:1 correspondence. Preserve n:m/empty/competing evidence and return human review, with no flattening or arbitrary ranking. General chain reasoning, graph closure, classifiers, native business writes, translations and migrations of existing Dimensions remain deferred.

## Evidence and verification limits

The public sources recorded in primary-notes.md and platform-evidence.json ground adopted clauses. No paid ISO or legal obligations, licensed taxonomy data or actual organizational records are copied. Source observations, local proposals and provider assertions remain distinct. Partial provider source reads are disclosed in manifests. Current prototype code is still unvalidated; neither S1 study approves it.