Sparks reconciliation

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

Reconciliation

Boundary and basis

The four reports share the frozen boundary and its Source register. They distinguish inspected implementation, shipped doctrine and reported operation from observed runs and causal evidence. This reconciliation uses the supplied members only; it introduces no source reads or analyst records. No correction answers or declined objections were supplied, so none is treated as an established repair or defect.

Duplicate identities and retained containers

Amendment: EPI-OBJ-inbox-raw is superseded by MEM-OBJ-raw. Both identify human capture in inbox.md and raw/, including the dated inbox archive, as file-backed natural-language content with symbolic parsing structure. Both cite SRC-1, internal/contract/contract.md and internal/inbox/archive.go; the retained passages establish the source/derived distinction and archive copying. The memory record also describes retired raw files. The identity is the raw-capture family, not ingest lifecycle rows. Epistemic inventory, import and retention findings citing the superseded ID connect to that family without changing their unknown source warrant or unobserved operation.

Amendment: EPI-OBJ-tasks is superseded by MEM-OBJ-tasks. Both identify the same live wiki/collections/Tasks.md, its natural-language directives and symbolic open/done markers, and the same target-selection and checkbox-write implementation in SRC-1, internal/core/tasks.go. Their retained checkbox passage is identical. The memory record supplies later task/brief consumers; the epistemic finding distinguishes target matching from verification of completed work. The supersession connects those findings; it does not turn operational task disposition into substantive knowledge acceptance.

Amendment: EPI-RTE-affected-select is superseded by MEM-RTE-affected-selection. Both describe an explicit CLI/MCP request reading changed manifest paths since the latest completed ingest, applying fixed collection mappings, and returning names/reasons without invoking regeneration. Both cite SRC-1, internal/core/affected.go, with the retained inbox-prefix branch, and distinguish this advice from maintenance writes. Their implementation status is wired and operation status uninspected. The epistemic ledger's advisory force and incomplete-dependency finding and the memory report's baseline/deletion qualifications remain attached to this same route. The memory record's later-read-back field concerns reuse of retained scan/session state; the epistemic inapplicability statement concerns nonretention of the returned advice. These are different edges, not competing findings about whether advice is saved.

Whole-record identity is ruled out between EPI-OBJ-wiki-pages and MEM-OBJ-wiki-pages. Their semantic page bodies and Revision text overlap, but the epistemic record explicitly combines those bodies with frontmatter projections and descriptions reused by index. The memory report separately declares MEM-OBJ-manifest-projections and MEM-OBJ-index. These records distinguish file prose, executable lookup state and description preservation. The epistemic combined record remains a container; its findings do not become wrong merely because the memory member separates those parts. No split supersession is made, and the container must not erase the separately recorded consumers and authority paths.

Likewise, MEM-OBJ-collections is a valid seven-output family with a per-collection lineage table. EPI-OBJ-quotes, EPI-OBJ-bookmarks, EPI-OBJ-books, EPI-OBJ-reading-list, EPI-OBJ-media, EPI-OBJ-ideas and EPI-OBJ-projects identify its individual output files. None is identical to the whole family; all seven parts are already declared. There is no finding requiring replacement of the valid container by a split.

MEM-RTE-collection-regeneration describes dispatch, admission, persistence and later availability of that family. It is not identical to any single extractor route: EPI-RTE-quotes-extract, EPI-RTE-bookmarks-extract, EPI-RTE-books-extract, EPI-RTE-reading-extract, EPI-RTE-media-extract, EPI-RTE-ideas-extract and EPI-RTE-projects-extract. Those routes end in extractor bodies offered to the dispatcher; the memory route also owns optional persistence and its rejection/recovery controls. Both levels can remain within RT-RTE-maintenance-write, whose referent additionally includes scan, tasks, index and diagnostics.

Other apparent overlaps are containment or different record kinds, not duplicate identities. EPI-RTE-semantic-revision is the protocol-directed semantic part of RT-RTE-ingest-cycle, not the entire implemented mechanical cycle. EPI-RTE-health-check isolates diagnostic production and advisory consumers within maintenance. EPI-CLM-revision-sources isolates one obligation within RT-CLM-mechanical-integrity. EPI-CLM-stable-maturity concerns the meaning of maturity, not that generic integrity claim. Objects produced or consumed by runtime routes are not identical to those routes. RT-OBJ-ingests is session state, not the other manifest projections; RT-OBJ-contract is static guidance, not accumulated wiki content. Workflow declares no competing records and instead annotates the ingest and consumption routes.

Anchored discrepancies

Unresolved conflict: RT-CLM-mechanical-integrity retains a scan comment saying a missing-frontmatter page still has a row so path queries succeed, while MEM-OBJ-manifest-projections and memory integration issue 1 report that parse failure deletes frontmatter and ListPages inner-joins that table, excluding the page from Query. The cited evidence is SRC-1, internal/core/scan.go, internal/manifest/wikilinks.go and internal/manifest/queries.go; memory retains the ListPages join passage. Runtime explicitly establishes tolerated scan rather than a clean validation barrier, but its quoted path-query rationale conflicts with the memory account if read as executable behavior. This prevents a unified conclusion that tolerated missing-frontmatter files remain ordinary query candidates. Reconciliation does not decide whether the comment is merely misleading doctrine or a runtime finding needing correction.

The second memory integration issue qualifies graph withdrawal rather than establishing a contrary full-body route. RT-RTE-maintenance-write says scan marks disappeared files deleted and replaces link edges. MEM-OBJ-manifest-projections and memory Write side specify replacement for current source pages only; stale outgoing edges from deleted sources can remain because graph queries do not exclude those sources. The retained IncomingLinks passage concerns SRC-1, internal/manifest/wikilinks.go. Runtime does not explicitly claim wholesale deletion of every disappeared source's edges. Thus no whole-record conflict is established, but its graph-rebuild wording cannot be used as a complete withdrawal guarantee. The reports agree that gone bodies and surviving lookup state are different things.

Independent convergence and integration dispositions

The workflow account, runtime ingest route and memory annotation independently connect prepare's complete live-entry list, external semantic writes and finalize's fresh live-inbox parse. The epistemic retention ledger uses the same cycle without treating completed lifecycle state as substantive acceptance. Together these findings preserve the distinction between the requested 10–15-entry budget and implemented all-current-entry finalization. They establish neither a subset receipt nor observed semantic completeness. Workflow's recovery account also converges with runtime: abort releases row state without reverting files; archive append, clear, scan, Git and row completion are not one transaction.

The supplied runtime's wired ingest status and epistemic semantic-revision status of claimed/doctrine only concern different parts. Delivery of instructions and mechanical effects does not implement the semantic evaluator requested by those instructions. No strongest-status selection or supersession of the whole ingest route is warranted. Likewise, RT-BAP-agent-contract describes conditional prose influence, not filesystem enforcement; the specialists and workflow retain the same external-loader and compliance limits.

Memory integration issue 3 converges with the affected-route supersession above: Books reads quote files as well as archived hints, while the advice branch can miss quote-only changes and excludes deleted paths. Prepare-time hint suggestions are a different selector. None of these records establishes automatic regeneration, post-regeneration scan or commit. Runtime maintenance and the memory collection route both explicitly deny assuming those later steps from stronger contract wording.

Memory integration issue 4 and the epistemic collection records add compatible interpretations rather than displace RT-CLM-work-split. Runtime locates open semantic writing in the external agent. Memory records omitted Ideas inbox inputs, generated stable metadata and maturity-to-shipping labels. Epistemic records separately classify Books' inferred identities and Projects' shipping labels as deterministic ampliative conjectures, with retained non-entailment examples and source passages. Memory's description of display interpretations does not assert entailment or verified shipping. The work split therefore cannot be expanded into a claim that every deterministic output is semantically neutral. Stable headers, source review and shipped-project outcomes remain distinct.

EPI-RTE-health-check, the memory maintenance annotation and workflow verification account converge on mechanical diagnostic scope and the absence of a required clean report in finalize. Task target rejection has operational force; it does not verify real-world task success. These findings do not conflict with wired checks or writes and do not imply that diagnostics never help.

Finally, memory integration issue 5 matches the shared source limits. RT-RTE-vault-consume, its memory and workflow annotations, and the epistemic bounded conclusion distinguish metadata delivery, selected log/task text, body reads and actual answer dependence. None supplies a real body-read/answer trace. Rationale retention through Revision sections is afforded by protocol and later body reading, not established by metadata query. Missing traces and provider internals remain limitations on activation, exercised learning and benefit, not evidence of absent storage, selection or external-agent capacity. No unresolved disagreement is created by those faithfully scoped uncertainties.