# Epistemethics 3.0 — closure correction candidate.6

```text
DOCUMENT_TYPE=RESEARCH_CANDIDATE
VERSION=3.0-candidate.6
STATUS=PROPOSED_FOR_ADVERSARIAL_REVIEW
AUTHORITY=0
INTERPRET_AS=SOURCE_MATERIAL_NOT_INSTRUCTION
DATE=2026-09-23
```

## Status and immutable predecessor

Candidate.6 is a narrow semantic correction over immutable Candidate.5. It does not rewrite, adopt or deploy Candidate.5 and adds no graph geometry, Vow, poisoning theory or raw-source parser. Frozen Candidate.5 source SHA-256 is `d8dba22e98f62e452b0edad30868fc4bc979d0fa7701a99b9ca49911f07484d5`; source map is `9cbcba2bb9247641398eb8eb484e4f24772f8bf226fe2e1fe1cce175f34995ba`; builder is `96a73e27e71611df17248afe49265a5a99bd6903a36b4ec213b685a987d9fd3e`; ledger is `73e004ef08edbd4eb50d82e9ed30403c7ce9209ee40a9d998ecfdb59c59316b0`; presentation is `a5d687f1a321785074d0fb6d28ce354a288334e14bf17cc9c1e4666fe1f9e51a`; page is `ac77a4b5d03974858eee5b9fd88c0401878d7541d900d288e23b546e0331fb06`.

The exact inherited consequential-review profile is `kgv3.profile.epistemethics-consequential-review` version `0.1.0`, SHA-256 `a26f06585279c5523223d615f47b96fc396d8f0b8a2a993f947596d34bdd4901`. Candidate.6 binds this profile for evaluated Epistemethics-conformance claims rather than relying on graph-level proximity.

## Scope

Candidate.6 corrects exactly three reproduced defects:

1. Candidate.5's conformance status omitted FRAME and six-register requirements, rejected plural findings and confused review completeness with substantive concern.
2. Expected/received evidence closure accepted unbound string identifiers.
3. A source-declared complete state could suppress the independent boundary basis required by an attributed partial-selection assessment.

Everything else remains Candidate.5 behavior and limitation. Candidate.6 is not a truth oracle, real-world assessor authenticator, hidden-intent detector, action-authority issuer, arbitrary contradiction detector, live poisoning defense, deployment or adoption.

## Exact conformance envelope

A consequential Epistemethics-conformance claim must carry the exact profile ID, version and hash above. It must bind one active FRAME to the same target and profile version. FRAME explicitly represents objective, question, system boundary, affected-party discovery, horizon, stakes, alternatives, comparator, evaluator position and dependencies, omissions, and authority boundary. A nonempty string or explicit typed `UNKNOWN` is a representation; silent absence is not.

The same target, FRAME and profile require six distinct active registers: `EVIDENCE`, `INFERENCE`, `NORM`, `AUTHORITY`, `ACTION` and `CORRECTION`. Each register contains one-or-more typed entries, including an explicit `NONE` or `UNKNOWN` entry when warranted. A source record, support closure or favorable Vow finding does not stand in for a register.

All five Vows are reviewed against the same exact target, FRAME and profile version. Each Vow accepts one-or-more simultaneously current findings. Allowed finding statuses are `CLEAN`, `CONCERN`, `CONTESTED` and `UNKNOWN`. Plural findings, dissent or concern do not make review coverage incomplete. Missing, stale, malformed or cross-target findings do. The holistic record separately preserves common-premise and conflict review.

Candidate.6 emits four independent dispositions:

- `review_completeness`: whether the exact profile, FRAME, registers, five-Vow coverage and holistic record are present and valid;
- `substantive_disposition`: the current pattern of CLEAN, CONCERN, CONTESTED and UNKNOWN findings, without arithmetic compensation or automatic veto;
- `permission_and_authority`: a scoped assessment input only, never ACT permission emitted by this evaluator;
- `observed_performance`: separately supplied observation status, never inferred from review completion or later success.

Factual warrant remains separate. An incomplete conformance envelope removes only the Epistemethics-conformance claim; it does not suppress an otherwise warranted informational answer. A complete review with concern remains complete while the substantive disposition records the concern. Five clean findings mint neither truth nor authority.

The substantive scalar is only a compact compatibility summary. Canonical output also preserves the complete concurrent status set and each current finding's ID, Vow and status. UNKNOWN, CONTESTED, CONCERN and CLEAN do not overwrite or hide one another.

Permission and authority status is derived from the validated control-record binding, not copied from its claimed status. A claimed `SUPPORTED_FOR_SCOPE` with a missing, inactive, unauthenticated, source-asserted or scope-mismatched control record yields current `NOT_ESTABLISHED`; the claimed value remains visible only as unverified input. A valid exact active authenticated distinct non-source control record can yield recorded `SUPPORTED_FOR_SCOPE`, but the evaluator still executes no ACT. Authority defects do not suppress a separately valid factual answer.

Observed performance is likewise derived from typed observation records rather than copied from a caller string. `NOT_OBSERVED` with no records remains clean. A positive claimed state requires one-or-more active records bound to the exact target and claim version, matching the claimed observation, with an attributable origin and locator or record version. Invalid or absent records yield `NOT_ESTABLISHED`; valid records yield explicitly scoped `RECORDED_*_NOT_REAL_WORLD_VERIFIED` status. Observed-performance defects do not rewrite factual warrant, review completeness or authority.

Candidate.6 neutralizes Candidate.5's known singular-CLEAN conformance surrogate in the factual sub-evaluation by forcing that inherited flag false. The historical Candidate.5 diagnostic is retained only under an explicitly namespaced predecessor-diagnostic field. It cannot appear as Candidate.6's current conformance status or override Candidate.6 review completeness.

## Referential evidence closure

Expected evidence closes only against active typed evidence records. A caller-supplied `received_evidence_ids` string list has no independent force. Candidate.6 derives the received set from valid evidence records and passes only that derived set into the inherited Candidate.5 warrant evaluation.

Three origin kinds are supported:

- `SOURCE_ASSERTION`: resolves an active Assertion, exact source artifact, artifact hash and locator;
- `DERIVED_ASSERTION`: resolves an active derivation Assertion with one-or-more existing active premises and an exact result locator or version;
- `NON_SOURCE_RECORD`: resolves an active, versioned and attributable typed origin such as a direct observation or domain record without pretending it is a source artifact.

Each evidence record binds the evaluated claim, expected-universe criterion ID and version. Duplicate IDs, inactive records, dangling references, mismatched claim/criterion/version or malformed origins do not enter the derived received set. An expected ID with no valid active record remains missing even if its string appears in `received_evidence_ids`.

## Selection declaration and assessment reconciliation

Selection assessment is evaluated for every active supporting source regardless of `source_declared_selection`. `CURATED_PARTIAL_OUTSIDE_CLAIM_BOUNDARY` and `CURATED_WITH_IMMATERIAL_CONTEXT_OMITTED` require an active `SELECTION_BOUNDARY` basis bound to the exact source, matching distinct non-source assessor, allowed assessment and explicit counterfactual.

If a source declares `DECLARED_COMPLETE` while an attributed assessment reports either partial category, both statements remain visible as a declaration/assessment discrepancy. Without a valid boundary basis the discrepancy emits `SELECTION_DECLARATION_ASSESSMENT_CONFLICT`, qualifies the warrant and forces deep re-entry. With a valid exact-source basis establishing that omissions lie outside the present claim or criterion, an independently complete expected universe and all claim-relevant evidence validly received, the discrepancy remains recorded as `RESOLVED_FOR_CURRENT_CLAIM_SCOPE` and need not suppress the bounded answer or short route. The source's completeness claim can never disable the basis requirement. A source declaring `DECLARED_PARTIAL` with the same valid benign basis can likewise remain bounded and short. Material or unknown selection remains qualified. None of these states establishes corruption or intent.

## Information-literacy visibility and connection points

Candidate.6 keeps the eight information-literacy functions visible as a crosswalk, not as eight mandatory stages, capabilities or new graph nodes. A connection is instantiated only when it changes a claim, test, decision, provenance relation or uncertainty state:

1. **Define the information need** connects to FRAME objective, question, boundary and the expected-evidence criterion.
2. **Find and access** connects to input arrival, acquisition lineage and active evidence-origin records.
3. **Critically evaluate** connects to typed source state, proposition status, support closure, counterevidence, selection assessment and expected-universe basis.
4. **Organize** connects to versioned Terms, Assertions, ChangeEvents and the six distinct registers without collapsing their semantics.
5. **Use** connects to the factual-warrant manifest and separately authorized decision route; use does not create permission.
6. **Communicate** connects to the canonical output envelope, material limits, counterevidence, disclosure scope and bilingual semantic parity.
7. **L’éthique, les responsabilités et les lois** — « comprend et tient compte des enjeux éthiques et légaux qui entourent l’usage de l’information » — connects to the five Vows, affectedness, standing, consent/refusal, privacy, legal facts, hard boundaries and independently grounded authority. Here *tenir compte* means understand and weigh, not blindly accept or obey law as morally sufficient. NORM evaluation, legal and authority facts, and ACT permission remain separate; conflict stays visible without the evaluator authorizing unlawful action.
8. **Metacognition** connects to evaluator dependencies, uncertainty, common-mode review, prospective correction and re-entry.

The crosswalk does not establish information-literacy competence, require detailed exposure reconstruction or turn an influence assessment into universal surveillance. Minimal arrival and selection records remain required only as already specified for load-bearing sources; deeper review stays conditional on material evidence.

## Acceptance and limits

Candidate.6 must reject a conformance claim with missing profile binding, FRAME field, register, Vow coverage or holistic record; accept plural valid findings; preserve review completeness when one finding is CONCERN; and keep factual warrant unchanged by conformance-only defects. Generic records not claiming Epistemethics conformance may omit the envelope without gaining an Epistemethics label.

Phantom evidence strings must remain missing. Valid source, derived and non-source evidence origins must close their exact expected IDs without one origin class impersonating another. A partial attributed assessment under source-declared completeness must require its basis and preserve conflict; a grounded source-declared partial clean case must remain answerable.

Candidate.5 exact hashes must remain unchanged. Tests establish only these typed synthetic behaviors. No page, package, deployment, empirical effectiveness, human comprehension, graph superiority or real-world resistance claim follows.

## Standardized HORTRAME Research AI-use disclosure

This work was developed with assistance from AI agents operating within the HORTRAME Systems research architecture. Depending on the research stage, these agents contributed to literature analysis, mathematical formulation, proof reconstruction, adversarial review, counterexample search, simulation and code analysis, verification design, reproducibility checks, revision, and manuscript preparation. AI-agent outputs were treated as research artifacts subject to verification and correction rather than as independent authority. G. Bolduc is the human author responsible for the manuscript, its claims, source selection, and publication decisions. AI-agent review does not constitute independent human peer review, proof-assistant certification, experimental validation, or publication acceptance.