OpenWiki reconciliation

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

Reconciliation

The pinned reports share the boundary and Source register. Their evidence establishes inspected implementation and shipped guidance, not observed execution or causal benefit. This reconciliation uses those members without additional source reads. No correction-answer files or declined objections were supplied. Memory's integration issues and the revised workflow account expose the corrections relevant to their connections. The previous reconciliation's context conflict was reassessed against the current records; it is not carried forward as a current disagreement.

Duplicate supersession

Amendment: EPI-OBJ-wiki-markdown is superseded by MEM-OBJ-wiki-page. Both identify generated repository and personal/local-wiki Markdown pages, interpreted prose and filesystem storage. Both connect that output to RT-RTE-wiki-retrieval, whose SRC-1 src/retrieval/wiki.ts evidence supplies current Markdown reload and section delivery. Memory additionally distinguishes deterministic byte transformations and later consumers. The affected epistemic inventory and ledger concern the same retained page output, not a second artifact. Both IDs remain declared. This identity does not replace either analyst's assessment of transformation, warrant or consumer authority.

Containers, parts and distinct identities

Claims have different operative parts. EPI-OBJ-page-claims describes a mixed Claim/PageClaims container. MEM-OBJ-claim isolates interpreted statements; MEM-OBJ-evidence isolates symbolic resource/version identities; MEM-OBJ-page-claims describes the symbolic persistence envelope while separating its semantic parts. Their SRC-1 schema and compact-display passages support different forms and consumers. No individual memory part replaces the epistemic container, and the container's mixed-form findings do not become wrong when these parts are separated. No split supersession is warranted. EPI-OBJ-verification-events isolates producer/time events in sidecars and their Markdown projection, with separate eligibility and withdrawal rules. It is not identical to the entire envelope. EPI-OBJ-page-sources is a whole-file provenance projection, not precise Claim evidence: statements, ranges and versions are omitted from that projection.

Queue, assignment, coverage and conversations are distinct. RT-OBJ-repository-run-state is transient lifecycle state; RT-OBJ-page-guidance separates its interpreted assignment material. Memory and workflow annotate them rather than declare replacements. MEM-OBJ-page-manifest survives finish as a coverage ledger. MEM-OBJ-chat-checkpoint has graph/thread consumers and an opaque serialized payload. MEM-OBJ-conversation-history instead identifies a home-directory filesystem mount with direct middleware writes and backend reads. It is neither SQLite payload, queue state nor wiki output. Runtime describes this channel inside generation but declares no object counterpart. None of these connections warrants an identity amendment.

Generation containers and maintenance functions are not duplicates. RT-RTE-native-generation and RT-RTE-host-generation are distinct invocation paths sharing the lifecycle core. MEM-RTE-claims-reconciliation spans their Claim maintenance subflow rather than replacing either complete invocation. EPI-RTE-evidence-preflight, EPI-RTE-claim-decisions, EPI-RTE-claims-persistence, EPI-RTE-verification-projection and EPI-RTE-source-projection separate checking, disposition, retention and projections. Their declared native-generation containment and described host-core use do not establish identity with the generation container. The broader memory route remains valid alongside those subfunctions; no supersession is warranted.

Authority paths carry different material. RT-BAP-page-instructions carries planner objectives. MEM-BAP-claims-maintenance carries retained propositions and evidence issues. RT-BAP-managed-guidance is host instruction policy, whereas RT-BAP-retrieved-sections supplies repository context. Shared consumers do not make these channels or forces identical.

Trace documentation, checkpoint continuation and history offload remain separate. MEM-RTE-trace-documentation starts with code-mode external operational samples and joins repository generation. It is not RT-RTE-personal-ingestion, which uses configured personal sources and another update graph. MEM-OBJ-trace specializes MEM-OBJ-connector-data with reduced execution structure and a separately traced transformation; specialization is not duplicate identity. MEM-RTE-chat-continuation supplies thread-based graph continuation and SQLite pruning. MEM-RTE-history-offload is a cross-cutting storage/access subflow spanning native generation and personal/chat agents, not either complete invocation. Its infrastructure is wired, consumer access afforded and operation uninspected. Those separate findings do not establish automatic compaction/reload or cross-worker reuse. No supersession is warranted.

Evaluator routes retain separate functions. MEM-RTE-ledger-evaluation spans checkpoint generation, capture, evaluation and carry. EPI-RTE-ledger-judgment isolates model judgment against source excerpts; EPI-RTE-ledger-reporting isolates verdict consumption and subsequent evaluation state. The valid broader route stays alongside these functions, without supersession. EPI-OBJ-ledger-verdicts is assessment output, not product Claim membership. MEM-RTE-evaluation-transfer copies generated/cached knowledge to a downstream task host; MEM-RTE-visualizer-delivery projects pages for humans. Neither duplicates requested section retrieval or scheduled Git publication.

Correction and integration dispositions

The earlier context disagreement is no longer present. RT-RTE-native-generation, RT-CLM-native-confinement, MEM-OBJ-conversation-history and MEM-RTE-history-offload distinguish fresh initial graphs from the shared durable history channel. The current workflow account and page-guidance annotation now make the same distinction. Planner research and failed-attempt conversations are not automatically transferred in initial prompts or restored through a generation checkpointer; this does not rule out separate history retention or later access. The retained SRC-1 src/agent/agent-backend.ts mount passage and test/agent/conversation-history-offload.test.ts direct-write/read-back passages support that connection. Model-tool write denial does not govern direct middleware backend writes. Workflow also explicitly limits page/Claim snapshot recovery: it does not establish rollback or deletion of history files. No current dependent finding requires the old exclusion of all conversation-derived material. Payload, selector, automatic replay and exercised cross-worker use remain unresolved across the reports; this faithful uncertainty is not a conflict.

Pending-page exclusions do not imply projection noninterference. The workflow page-admission account now distinguishes exclusions from Claim-session persistence and eligibility entries from the verification projector's traversal of all discovered pages. Its SRC-1 src/okf/claims-verification.ts passages show that absent active events can remove existing OpenWiki-family stamps on another pending page. EPI-RTE-verification-projection supplies the same producer-family filtering and absent-event rule. This qualifies a possible inference of complete pending-page noninterference; it does not contradict RT-CLM-page-durability, which claims persistence ordering and durability checks, not universal immunity of other pages' metadata. The passage retained by MEM-RTE-claims-reconciliation that unrelated Claim state remains intact concerns session persistence, not preservation of every projected Markdown stamp. No inspected execution establishes a race, semantic corruption or delivery failure. The reports therefore connect without a new conflict or an upgraded failure claim.

Independent convergence and remaining limits

Durability is not semantic warrant. RT-CLM-page-durability, memory's reconciliation route and epistemic persistence agree that Claims persistence precedes queue completion. EPI-CLM-grounded-claims is a broader public lineage/warrant contract rather than a duplicate durability guarantee; its README passages explicitly exclude connector-derived facts. Evidence identity, protocol admission and proposition truth remain separate. No supplied report establishes an observed evidence-consuming acceptance of a produced ampliative proposition.

Workflow delivery and recovery remain bounded. The native-generation annotation adds quickstart ordering, same-attempt correction, one fresh retry and fatal-error behavior to the supplied route. The durability annotation agrees with runtime that skipped restored pages and interrupted metadata can coexist with complete finish. Native snapshots do not cover external host writes, Git publication or history rollback. RT-RTE-maintenance-admission separately permits a PR to preserve completed partial progress after CLI failure; actual merging remains unobserved.

Withdrawal is consumer-specific. The memory retrieval annotation establishes search/direct-read deprecation asymmetry. Runtime claims current-byte reload rather than universal withdrawal. Epistemic's retrieval assessment leaves explicit-read behavior unassessed and rejects a universal withdrawal inference. Different inspection coverage is not disagreement. Claim retraction, metadata withdrawal and removal of prose or historical copies remain separate effects. The memory integration issues therefore constrain stronger synthesis claims without requiring runtime correction.

Evaluator detail does not override another report's uncertainty. Epistemic judgment/reporting records add bounded criteria and consequential verdict consumers to memory's broader LEDGER route. Memory leaves detailed evaluator internals uninspected; epistemic retains incomplete extraction and evidence-selection coverage. Both lack observed accuracy and an inspected score-feedback gate into product generation. A shipped benchmark semantic evaluator thus coexists with no independent semantic answer oracle in the inspected page-admission sequence.

Other limits remain local. RT-CMP-chat-model bounds native parameter mutability and immutable version identity. SQLite payload, history-production/reload policy and external host/provider behavior retain different unresolved facts. The epistemic bounded conclusion leaves learning, content-directed criticism and self-improvement independently unestablished within assessed paths; it does not deny memory's supported trace-fed authoring capacity or history infrastructure. Memory's grounding, trace and residual coverage issues remain limitations rather than cross-report disagreements. These connections identify no further duplicate supersession, required missing operative part or current unresolved conflict. Named inspection limits still prevent complete-system negatives, deployed compliance, exercised reliance and empirical benefit conclusions.