sift-kg reconciliation

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

Reconciliation

Evidence boundary

All four reports use the revision and evidence layers fixed by boundary: implementation and doctrine/design, without supplied observed-run or causal-experiment evidence. The shipped CLI, library stages and agent skill are inside the boundary; actual external-agent scheduling and final responses are outside it. The reports' implemented routes and unobserved operation therefore coexist without a status conflict. This reconciliation uses only the pinned members and makes no additional source reads. No analyst correction answers or declined objections were supplied to this attempt.

Duplicate identities and retained parts

Amendment: EPI-OBJ-extraction-records is superseded by MEM-OBJ-document-extraction. Both identify the document-derived entity/relation candidates held in DocumentExtraction objects and per-document extraction JSON, produced by RT-RTE-document-extraction and consumed by build. Identity evidence is their matching source-native structure, process/filesystem substrates and extraction lineage at SRC-1, src/sift_kg/extract/extractor.py. The epistemic record also describes the constituent ExtractionResult entries; that description does not establish a second retained per-document artifact. This amendment connects the common candidate object, not the upstream source text, transient document summary, graph output or extraction route. The epistemic inventory, parser-check target and retention/cache rows concern this same object. Its prompt-based explicitness requests and unknown semantic fidelity remain the epistemic report's findings; memory's configuration-only reuse, context aggregation and cache-withdrawal findings remain memory's findings. Neither set of values replaces the other.

Amendment: EPI-OBJ-graph-support is superseded by MEM-OBJ-relation-support. Both identify the canonical edge's mentions, support counts/documents, confidence and representative evidence within RT-OBJ-graph-data, on NetworkX dictionaries and graph JSON. Their shared identity evidence is SRC-1, src/sift_kg/graph/knowledge_graph.py, with the same product-complement aggregation passage retained in the memory object and EPI-RTE-support-aggregation. The epistemic record's source_document field describes the same preferred-mention attribution, not another support store. Affected findings are the epistemic support inventory and aggregation/flagging ledger targets, and memory's retained-versus-delivered evidence account. The arithmetic's bounded derivation, missing source-independence check, reload normalization and query projection remain separately attributed findings. The aggregation route is not superseded by its output object.

No other full-record identity is established:

• RT-OBJ-domain-config remains a valid mixed container. EPI-OBJ-domain-vocabulary and EPI-OBJ-domain-guidance explicitly identify its symbolic and interpreted parts. Memory annotates the parent rather than declaring another schema. The parts' different consumers do not invalidate the container, so no split supersession is warranted.
• RT-OBJ-graph-data remains the graph container. EPI-OBJ-graph-assertions distinguishes assertion payload from support metadata. The support records above describe a part, not an alternative graph. Neither the assertion payload nor support part replaces the container.
• EPI-OBJ-narrative-artifacts combines narrative Markdown and description prose for its warrant assessment. MEM-OBJ-narrative and MEM-OBJ-entity-descriptions distinguish those artifacts' persistence and later consumers. The combined record does not assert a common invalidation or automatic read-back route for their full contents; it can remain alongside those parts. MEM-OBJ-communities separately identifies ID assignments and generated labels; the epistemic narrative record mentions that additional sidecar rather than identifying it with narrative.md or entity_descriptions.json. No combined finding requires replacement by a split.
• EPI-RTE-extraction-check is a structural-admission part of extraction. EPI-RTE-support-aggregation, EPI-RTE-graph-cleanup and EPI-RTE-relation-flagging distinguish functions inside RT-RTE-graph-build. Their targets and consequences differ; none duplicates or invalidates the surrounding route.
• MEM-RTE-export-artifacts, MEM-RTE-view-artifacts and RT-RTE-agent-retrieval share retained graph inputs but have different delivery mechanisms: exported files, browser artifacts and CLI retrieval output. Shared inputs do not establish route identity.
• EPI-CLM-merge-approval and EPI-CLM-source-provenance share a README sentence but concern approval and attribution respectively. Neither duplicates RT-CLM-second-brain or RT-CLM-incremental-extraction. The workflow declares no additional records; its annotations extend the supplied runtime records rather than duplicating them.

Cross-report connections

Admission and later mutation. RT-RTE-graph-build, EPI-RTE-relation-flagging and the workflow account converge on graph delivery before relation review. Structural admission and later flags do not certify documentary truth. RT-RTE-review-decisions separates threshold decisions from manual decisions; workflow adds the persisted-YAML handoff and the preservation of earlier decisions on quit. RT-RTE-apply-decisions consumes confirmed merges and rejected relation keys. Memory and epistemic findings agree that confirmed variant suggestions are not insertion commands. These connections support neither human-only admission nor the opposite claim that review has no consequence. The approval-claim mismatch in the epistemic report is a source-claim comparison, not a disagreement with the runtime.

Support and reliance. Memory's relation-support finding and EPI-RTE-support-aggregation converge on repeated mentions affecting confidence without establishing independent evidence. Epistemic's numerical example is an implementation deduction, not an observed run. Memory adds that query omits full mention arrays and node contexts. The inference path therefore runs from retained mention data through aggregation and selection to a narrower delivered projection; storage alone supplies neither source authenticity nor an independent answer oracle. Cleanup's enforcing transformations in EPI-RTE-graph-cleanup coexist with memory's structural-curation account and runtime's lack of complete semantic validation.

Persistence is not synchronized withdrawal. MEM-OBJ-document-extraction identifies the reconstruction base, while RT-RTE-apply-decisions changes saved graph state. Workflow retains the same rebuild limit: reconstruction is not inverse merge recovery. Memory's descriptions, communities and narrative records extend that distinction to independently retained sidecars. None of the reports establishes that a deployed host regenerates every sidecar after every graph change. That uncertainty prevents a deployment-wide freshness or withdrawal conclusion, not the implemented graph-mutation finding.

Schema updates and model components. RT-RTE-schema-discovery and RT-BAP-domain-guidance, with the memory annotation and epistemic symbolic/interpreted parts, connect generated YAML to later prompts and selected code rules. This retained-policy path is distinct from parameter training. RT-CMP-llm-client and RT-CMP-sentence-transformer retain external-version and provider-state limits. The reports do not derive improved capacity, exercised self-improvement or theory criticism from schema persistence alone.

Delivery and external response. RT-RTE-agent-retrieval, memory's read-back annotation and workflow's terminal-boundary annotation agree that stdout reaches a caller while the external host owns further queries and final answers. RT-BAP-host-skill supplies prescribed orientation, paid-call confirmation and answer grounding, not CLI enforcement or observed compliance. These are connected mechanisms, not independent operational corroboration: the reports share SRC-1 and often rely on the same runtime records.

Integration-issue dispositions and conflicts

The five issues in memory's integration section have these dispositions:

1. Metadata filtering: the search exception extends the runtime's exact-ID countercase. The runtime's neighborhood-array filtering finding remains scoped to that projection; memory does not dispute it. Neither report licenses metadata-free results across all agent-facing commands. No inter-report conflict is established.
2. Description/export differences: memory distinguishes CLI non-JSON enrichment, JSON graph export and the library run_export path. Runtime lists descriptions' viewer/search consumers without asserting universally enriched exports. The epistemic narrative record says descriptions reach export wrappers, not that every wrapper joins them. These existential consumer findings coexist with memory's route-specific limits; no universal export-enrichment conclusion follows and no report finding must be replaced.
3. Consistency: unbound review files and sidecars, and unknown host regeneration, remain local limitations connected above. There is no opposed deployment-freshness finding to settle.
4. Identity comparison: memory's nonduplication comparisons concern its supplied runtime records and are retained. Comparing memory with the separately supplied epistemic member establishes the two additional identities stated in the amendments. The valid containers, material parts and distinct export/view routes are not superseded.
5. Unresolved operation and accuracy: host compliance, source/schema fidelity, source independence, further prompt truncation, browser behavior, import fidelity and provider updates remain unestablished where the reports place those limits. These are uncertainties rather than contradictory findings. They neither erase implemented routes nor justify a system-complete negative.

No unresolved disagreement between the pinned reports is identified. No dependent finding requires a supplied record to say something it no longer says. The independent reports' compatible qualifications do not authorize selecting a stronger status, adding observed operation or settling an accuracy claim.