sift-kg reconciliation

Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md

Reconciliation

Boundary and evidence

The four reports describe the same shipped document-to-graph system and external-host interface under the boundary. Their implementation and doctrine findings use SRC-1 at the frozen revision; none supplies observed execution or causal comparison. Convergence below is agreement between analytical lenses, not independent experimental corroboration. No source files were read for this reconciliation. No correction answers or declined objections were supplied.

Record identity and containment

Amendment: EPI-OBJ-domain-config is superseded by MEM-OBJ-discovered-domain. Both identify the corpus-discovered DomainConfig, its symbolic constraints and interpreted descriptions, and its reusable discovered_domain.yaml representation. Both derive it from RT-RTE-schema-discovery and identify extraction/build consumers. The epistemic inventory and discovery-ledger citation concern this same guidance object. This supersession connects identities only: it does not change either report's schema-adequacy, admission or theory findings, nor identify the object with static shipped domains.

EPI-OBJ-extracted-content identifies the assertion-bearing content within the persisted results described by MEM-OBJ-document-extraction. The memory object additionally includes cache identity, paths, errors and processing metadata consumed for reuse. These are a content part and a valid retained-result container, not identical scopes requiring supersession. EPI-OBJ-document-text and EPI-OBJ-document-summary are separate transient inputs to the extraction route, not duplicates of its persisted output. In particular, the summary's delivery to later chunk calls does not identify it with a retained summary field in DocumentExtraction.

EPI-OBJ-graph-data is a valid graph-content container spanning the entity/provenance state in MEM-OBJ-graph-data and the canonical edge evidence/support part in MEM-OBJ-relation-support. Neither memory part alone replaces the epistemic container. Their distinct aggregation, retained fields and delivered evidence remain applicable; no split supersession is asserted.

Likewise, EPI-OBJ-review-state validly groups the decision material that memory separates into MEM-OBJ-merge-proposals and MEM-OBJ-relation-review. The combined finding distinguishes reasons from operative statuses and does not require that entity merges and relation decisions have identical effects. Keep the container and its separately described file/control parts; do not supersede it by either file alone.

EPI-OBJ-derived-prose groups the narrative, descriptions and generated labels. MEM-OBJ-narrative, MEM-OBJ-entity-descriptions and the label portion of MEM-OBJ-communities describe its distinct outputs and consumers. Community assignments also have structural lineage and consumers beyond prose. These are overlapping parts and a valid container, not interchangeable referents; no supersession is warranted by their grouping alone.

The epistemic subroutes EPI-RTE-extraction-shape, EPI-RTE-relation-normalization and EPI-RTE-style-rewrite identify material operations within, respectively, RT-RTE-document-extraction, RT-RTE-graph-build and RT-RTE-narrative-generation. They do not duplicate their entire parent routes. The parent accounts retain their wider selection, persistence and recovery findings.

The three claim records are distinct: RT-CLM-merge-approval concerns approval and source linkage; EPI-CLM-explicit-extraction concerns the explicit-source extraction policy; EPI-CLM-graph-grounding concerns answer grounding. The last is also distinct from RT-BAP-agent-skill: a warrant claim and a host-directed behavioral-authority path are different record kinds and referents. Workflow declares no new records; its annotations add handoff and delivery roles to runtime records without creating identities to supersede. No additional component or absence counterpart is declared in the other reports.

Cross-report disagreement

Unresolved conflict: EPI-OBJ-structural-results and its entailed-derivation ledger row attribute counts from an “Info/topology/query request” to RT-RTE-graph-retrieval, while MEM-RTE-project-info explicitly identifies info as a distinct operation outside that runtime route's query/search/topology progression. The runtime route itself names query/search/topology, not project-status inspection. The epistemic evidence anchor is SRC-1 src/sift_kg/cli.py; memory's info record retains the extraction-file count witness and its separate configuration, status-file and existence-check inputs. Both positions permit structural computation, but the supplied runtime route does not cover all of the invocation family the epistemic row assigns to it. This prevents treating that row's whole info/topology/query scope as supported by the cited runtime route alone. It does not establish that the computed counts are false, or that the structural-results container must be superseded by the info route. The verifier must dispose the route-attribution discrepancy; this reconciliation changes neither declaration.

Convergence and integration dispositions

Search filtering is a scoped addition, not an established contradiction. Memory's retrieval annotation distinguishes search matching over all nodes from query's metadata-stripped neighborhood selection. RT-RTE-graph-retrieval locates metadata removal in substantive neighborhoods and describes search separately; it does not explicitly guarantee DOCUMENT-free search matches. The memory integration issue therefore limits any broader reading of the runtime account without replacing its supported query finding. The epistemic selection analysis likewise cannot turn neighborhood filtering into a universal source or metadata filter.

Persistence, freshness and withdrawal remain separate. The memory integration issues extend RT-RTE-document-extraction and RT-RTE-graph-build: cache identity omits source content; retained extraction inputs can survive original-document deletion; rebuild starts a fresh graph and does not preserve arbitrary surgery. Workflow gives the same rebuild consequence for the host's update procedure. Memory's independent sidecar freshness findings connect MEM-OBJ-communities, MEM-OBJ-entity-descriptions, MEM-OBJ-narrative and RT-RTE-view-export. These additions do not contradict wired persistence: reloadability does not entail revision-linked coherence or coordinated withdrawal.

Approval and factual acceptance have different scopes. Runtime's confirmed-status surgery gate, memory's status-versus-reason distinction, epistemic's separate check/disposition rows and workflow's optional maintenance branch converge on RT-RTE-review-application. Automatic thresholds and manual edits can supply statuses; flags do not quarantine the already saved graph; confirmed variant proposals are not inserted by application. Epistemic's lack of established evidence-consuming factual acceptance does not negate the implemented operational admission gate. The reports also retain the possibility of careful human acceptance without asserting that every status records it.

Evidence delivery is narrower than retained provenance. MEM-OBJ-relation-support distinguishes saved mentions from representative evidence returned by query. Memory's narration annotation further distinguishes name-matched, capped contexts from full source evidence. These connect runtime's retrieval/narration selection and epistemic's bounded source-warrant conclusions without asserting that retained evidence is absent. The specific conditional origin-to-destination mapping in EPI-RTE-relation-normalization adds a semantic finding to the broader build account; it does not classify every normalization as conjecture or establish an observed bad extraction.

Host delivery is not demonstrated host uptake. Workflow traces artifact production and machine-readable delivery separately from the host's final response. Memory distinguishes caller-directed pull from the skill's session-start protocol; runtime distinguishes the same protocol from executable scheduling. These agree with EPI-CLM-graph-grounding's bounded comparison: the graph can shape later answers without this evidence demonstrating their truth or benefit. Missing host traces, dependency pins, accuracy evaluations and controlled comparisons remain shared limitations, not cross-report disagreements or reasons to promote an implementation finding to observed operation.