dot-iu-cutter v0.5 — `evidenced_by` Edge Risk & Alternative Review

Phase: v0.5 v0_5_evidenced_by_edge_APR_draft · Nature: design_only_APR_draft · Date: 2026-05-18 Companion to: dot-iu-cutter-v0.5-evidenced-by-edge-APR-draft-2026-05-18.md
apr_status: DRAFT_ONLY
edge_type_evidenced_by: NOT_APPROVED      # APR-GATED in every mention below
decision_authority: GPT / User ONLY
self_advance: PROHIBITED
Every mention of evidenced_by below is PROPOSED / NOT APPROVED / APR-GATED. This review compares options and gives a recommendation; it does not decide.

1. Options under comparison

A: approve evidenced_by as a new durable universal_edges edge type
   (APR cấp medium Đ32 + amend edge_type vocab + extend DOT Cap-4) — PROPOSED, NOT done
B: keep assembly-local evidence annotation only
   (WS-3 alternative B — relation lives only inside assembly_profile output, not persisted)
C: reuse existing edges as a workaround
   (references / implements / derived_from — no new edge, no APR)

For reference, the WS-3 logical proof currently runs on B (GPT ruling: ACCEPTED), with iu_entity_binding binding_kind=evidences handling the raw-entity case (this is orthogonal to A/B/C and is not a substitute for the IU↔IU edge).

2. Trade-off matrix

Dimension	A — approve `evidenced_by` (NOT APPROVED)	B — assembly-local only	C — reuse references/implements/derived_from
Semantic clarity	High — a first-class, named "X is proven by Y" relation; no overloading	Medium — meaning exists but only at assembly time, ephemeral, not a graph fact	Low — borrows an edge whose definition is mention / spec-realization / creation-provenance; "proves performed" is lost (§4 of APR draft)
Vocabulary bloat	Adds exactly 1 edge (minimized from 6; WS-1). Controlled, justified, §3.3-gated	Zero vocab change — best on this axis	Zero vocab change, but causes semantic drift inside existing edges (worse long-term than a clean +1)
Production queryability	High — durable, indexable, reverse-index served, participates in INV-DUAL / Cap-4 audits	None — relation not persisted; cannot be queried outside an assembly run; no audit trail	Medium-but-misleading — queryable, but queries cannot tell evidence from citation/impl/provenance → false positives in compliance queries
compliance_matrix readiness	Enables a future durable compliance_matrix proof (still DEFERRED; not proven here)	Blocks durable compliance_matrix — would force assuming an unauthorized evidence relation (scope-creep / gate breach per WS-3 §4)	Unsafe — compliance_matrix built on overloaded edges yields semantically wrong attestations
Implementation risk	Medium — needs vocab amend + Cap-4 checker extension + provenance/lifecycle wiring (all post-PASS, none done here)	Lowest — nothing to build; already proven runnable in WS-3	Low to build, High to live with — silent semantic corruption is expensive to unwind later
Governance risk	Low if routed through APR Đ32 (GPT/User sovereign); High only if self-approved (explicitly forbidden)	Lowest — no governance surface	High — bypasses §3.3 by hiding a new semantic inside an old edge; defeats edge-conformance intent (NT4 drift)

3. When NOT to use `evidenced_by` (NOT APPROVED / APR-GATED)

This is the disambiguation contract. Even if the edge is ever approved, it MUST NOT be used in these cases:

do_NOT_use_evidenced_by_when:
  raw_evidence_entity:
    case: the evidence target is a raw_entity — report FILE PATH, verification
          artifact file, Directus item id, code module, SQL entity (NOT an IU,
          fails OQC, not in the relation graph)
    use_instead: iu_entity_binding with binding_kind = evidences (+ entity_reference_registry)
    apr_needed: false        # mechanism already exists (WS-1 binding note §2)

  weak_mention:
    case: only a read-only citation/mention is needed, with NO attestation /
          "proves performed" semantics
    use_instead: existing core edge `references`
    apr_needed: false

  provenance_source_derivation:
    case: target is the SOURCE the unit was created/derived FROM (creation
          provenance), not a proof that a requirement was met
          (e.g. WS-3 R7: lesson N7 derived_from report N6)
    use_instead: existing candidate edge `derived_from`
    apr_needed: false

  implementation_relationship:
    case: source REALIZES a specification target (build-to-spec), not attestation
    use_instead: existing core edge `implements`
    apr_needed: false

  governance_authority:
    case: source is under the normative governance/authority of target (law/agency)
    use_instead: existing candidate edge `governed_by`
    apr_needed: false

Net: evidenced_by (NOT APPROVED) is reserved strictly for durable IU↔IU attestation — "this normative/requirement IU was demonstrably satisfied by that evidence IU" — and nothing else.

4. Recommendation (advisory only — NOT a decision)

recommended_option: A  (approve evidenced_by as a new durable edge) — via APR cấp medium (Đ32)
confidence: medium-high
status: RECOMMENDATION_ONLY — NOT_APPROVED, NOT_SELF_DECIDED
decision_owner: GPT / User (sovereign APR Đ32)

Why A is recommended:

It is the only option that preserves true evidentiary semantics and yields a durable, queryable, auditable relation — the exact thing a future compliance_matrix proof and production evidence chains need (GPT WS-3 review: "preserving the future value of a durable evidenced_by edge for compliance_matrix and production evidence chains").
Vocabulary cost is minimal and justified: exactly +1 edge, already minimized from 6 candidates (WS-1), with a clean §3.3 justification (no existing edge fits — APR draft §4).
Option B is correctly the interim state (WS-3 runs on it, GPT-accepted) but cannot carry durable/production compliance — it structurally blocks the compliance_matrix proof.
Option C is the worst: low build cost but it injects permanent semantic drift into references/implements/derived_from/governed_by, defeating the very purpose of P44-4A edge conformance (NT4) and producing wrong compliance answers.

Bounded recommendation (what is NOT being recommended):

Not recommending immediate creation, vocab amend, or Cap-4 extension — those are strictly post-APR-PASS and GPT/User-owned.
Not recommending any change to the GPT-accepted WS-3 proof: until/unless A is approved, continue on B + iu_entity_binding exactly as proven.
Not recommending any compliance_matrix work — it remains DEFERRED (QG8); named here only as the downstream beneficiary, not proven.
If GPT/User prefer defer: keep B as the standing interim mechanism; revisit evidenced_by before any production compliance/evidence durability is required (BP-FA4).
If GPT/User reject: compliance_matrix durability and production evidence chains must be explicitly re-scoped or descoped — and the WS-1 minimization conclusion ({ evidenced_by }) is closed as "not pursued".

5. Risk register

ID	Risk	Severity	Mitigation (proposed — not executed)
RK-1	Edge approved but used for raw evidence entities → duplicates `iu_entity_binding` semantics	Medium	§5.2 boundary rule in APR draft; DOT Cap-4 endpoint OQC check (post-PASS)
RK-2	Edge approved but overlaps `references`/`implements`/`derived_from`/`governed_by` in practice → drift	Medium	§4 non-overlap justification; anti-drift constraint set; reviewer education
RK-3	Self-approval / self-advance into vocab amend, schema, or compliance_matrix	High	HARD-GATED: GPT/User sovereign only; this phase design-only; QG6/QG8
RK-4	Deferral leaves compliance_matrix permanently blocked without an explicit owner	Medium	Tracked as BP-FA4; recommendation explicitly surfaces the consequence of defer
RK-5	`derived_from`/`governed_by` Candidate edges never ratified (BP-WS2-2) → broader minimization unstable	Medium (orthogonal)	Tracked, NOT resolved here; owner = P44-4A v1.0 ratification (User)

End of risk & alternative review (design-only). evidenced_by remains NOT APPROVED / APR-GATED. GPT/User decision required: approve | reject | defer.

dot-iu-cutter v0.5 — evidenced_by Edge Risk & Alternative Review