Citadel reconciliation

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

Reconciliation

Evidence and identity boundary

The four analyst members share the run and revision established in boundary. Its Source register identifies SRC-1. Their implementation findings, model requests and shipped policies are not supplied executions. No report establishes observed host reliance or causal benefit. Whole-system scope does not imply exhaustive inspection: memory and epistemic retain specific available but uninspected paths. This reconciliation used only the pinned members; no source read was needed to resolve an ambiguity.

Supersessions and distinctions

Amendment: EPI-OBJ-node-note is superseded by MEM-OBJ-node-note. Both identify the Markdown Dev session note produced by distill_node_note, not the public acceptance receipt. Their retained passages at SRC-1 kb/session_trace_distill.py identify the same heading and Outcome construction. Memory supplies the extraction limits, mixed transcript lineage and later consumers; epistemic assesses the truth-apt assertions in that output. The epistemic inventory and extraction findings therefore concern the memory-declared object. This identity does not discard epistemic's finding that Outcome and Files changed labels can overstate what the transcript establishes.

Amendment: EPI-RTE-source-enrichment is superseded by MEM-RTE-source-enrichment. Both now identify the optional direct-learning generative transformation, not an enabled stage of HTTP ingest. The connected SRC-1 kb/learning.py and kb/llm_enrichment.py passages trace source input through model text/summary/tags, parsing and learning chunk ingestion to later search. Memory describes admission and revision handling; epistemic separates transformation from shape checking and leaves semantic preservation indeterminate. Both declare wired implementation and uninspected operation. Those epistemic ledger functions remain distinct functions of the same route. EPI-OBJ-enriched-chunk is its output object, not a duplicate route.

EPI-RTE-note-distill remains a valid declared part of RT-RTE-session-capture. Extraction does not replace the parent's transcript-access, byte-cap, transport and admission account. MEM-RTE-share-session instead has volunteered input, a separate endpoint and dual-target admission; MEM-OBJ-session-trace is its distinct compact output. Neither is identical to automatic personal-note capture.

MEM-OBJ-source-revisions, MEM-OBJ-projection-receipts and RT-OBJ-ingest-result are respectively retained bytes/head relations, durable processing state and an immediate public receipt. They are not duplicate objects. MEM-OBJ-search-projections identifies downstream representations, not the worker progression declared as RT-RTE-source-projection. Likewise, MEM-BAP-current-head names selector authority, while EPI-RTE-source-withdrawal names tombstone admission and filtering. No supersession joins these connected objects, routes and authority paths.

MEM-OBJ-optimization-note is the additive retained output of RT-RTE-index-curation, not its static guidance RT-OBJ-optimize-prompt. EPI-OBJ-organization-digest and EPI-RTE-digest-build concern a separate preview/notification product. A common provider and summary-shaped content do not establish identity.

EPI-RTE-feedback-record covers the service retention/fallback and optional improvement trigger within the broader account of MEM-RTE-feedback-processing. The latter also covers scheduled event classification. That combined account remains valid alongside the focused service account; neither replaces the other's full referent. MEM-OBJ-feedback-events is the retained metadata/decision object. Recording a rating, classifying it and proposing an optimization note are not identical operations.

MEM-CLM-retrieval-evaluation concerns answer-span and document-identity tests. EPI-RTE-benchmark-gate traces scoring and enforcing comparison; EPI-OBJ-benchmark-run names saved comparison inputs; EPI-CLM-benchmark-score records an attributed numerical claim. These different referents require no supersession. Nor is the benchmark artifact the permission receipt EPI-OBJ-evaluation-gate.

Integration-issue dispositions

Corrected enrichment attribution. Memory's integration issue 3 now explicitly removes its former containment under RT-RTE-ingest-admission. Both enrichment declarations distinguish direct learning's default permission from the HTTP helper's allow_llm=False. Runtime retains the same exclusion and separately allows later embedding/projection work. The correction removes the cross-report attribution problem; the enrichment supersession above must not be read as restoring HTTP containment. No correction-answer file was supplied, so this disposition rests on the current findings, not an inferred answer.

Rejection versus retention. Memory's issue 1 qualifies the runtime's statement that rejected ingest writes do not enter the store. MEM-RTE-vault-sync and MEM-OBJ-obsidian-revisions retain the JSON save before separate search-admission checks in SRC-1 kb/server.py and kb/obsidian_sync.py. That is a different substrate and ordered gate, not a disagreement with the particular HTTP-ingest route's rejection finding. A universal claim that every rejected searchable write leaves no retained source is prevented.

Withdrawal versus expansion. Memory's issue 2 and EPI-RTE-source-withdrawal converge with runtime's current-binding account. MEM-RTE-document-drilldown adds requested historical revision access, not another task-search declaration. Its retained lookup passage prevents extending current-search exclusion into universal historical-read denial. No report establishes physical erasure or retraction of emitted context.

Capture fidelity. Memory's lossy extraction account and epistemic's failed-tool countercase address different properties of the same note. EPI-RTE-note-distill derives the countercase from selection code, not a supplied execution. It qualifies the apparent outcome/file-effect labels without contradicting deterministic distillation or the retained-write chain. Neither report establishes that all notes are false or that local distillation calls a model.

Promotion and delivery. Workflow declares no new records; its workflow account and annotation extend RT-RTE-promotion-review. They add the scheduled receipt prerequisite, candidate and model-input caps, redacted list handoff, full-candidate reload and admin decision boundary. Runtime and workflow both retain consumed failed approvals and distinguish accepted Central content from indexing and later use. The browser handler's failure to inspect promoted does not contradict backend acceptance accounting. Workflow's queue metadata does not establish the complete rationale trail into later search that memory leaves uninspected.

EPI-CLM-promotion-gate is the public shared-layer claim; RT-CLM-explicit-approval is the ingest-tool instruction. They are not identical claims. The reports independently converge on automatic known-work admission versus queued new-project review. This prevents universal per-item human-approval attribution, without treating credentials or browser confirmation as evidence of an actual human decision.

Opaque maintenance, reasons and feedback. Memory's issues 4 and 5 remain local limits, not conflicts. MEM-OBJ-search-projections retains the opaque Cognee improve/global-context-index successor. That uncertainty does not negate explicit additive optimization-note writes. EPI-RTE-automation-receipt is a declared prerequisite part of the index-curation route, not its opaque transformation. Workflow describes another consumer of the receipt predicate at scheduled promotion. These consumers and the benchmark comparator are distinct enforcement paths; none establishes the receipt issuer's measurements.

Memory and epistemic agree that feedback metadata classification is not an installed ranking successor, and that fallback notes can independently enter search. Neither establishes raw feedback text as optimization-model input. Pointers, summaries and tags are retained, but mandatory reasons and later content-directed criticism are not established by the optimization format. These limits do not erase retention or infer absent learning from opaque provider processing.

Disagreements and remaining limits

No unresolved cross-report disagreement is established by the current pinned findings. The former enrichment attribution is corrected in the supplied memory member and agrees with runtime and epistemic. The rejection and withdrawal issues qualify store and consumer scope rather than contradict their bounded findings. No declined objection was supplied; none is presumed settled or treated as an established defect.

Memory's issue 6 and epistemic's coverage table retain partially inspected connectors, capture paths, repair/backup operations, satellite policies and detailed client behavior. They prevent universal corpus, consent, fidelity, rollback, erasure and policy-agreement conclusions. Missing execution and intervention evidence separately prevents activation and benefit attribution. These are uncertainties, not absent mechanisms. No combined finding requires a split to repair a disagreement on the supplied evidence. Both superseded IDs remain declared; the other compared referents remain distinct or valid containers and parts, without strengthened statuses.