llm-wiki-cli reconciliation

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

Reconciliation

Boundary and common account

The reports share the whole-system boundary and its Source register. Their detailed inspection scopes differ; whole-system inclusion does not mean every analyst inspected every included operation. This reconciliation uses the supplied members only, with no additional source reads. No correction answers or refusal were supplied.

The reports converge on an external agent owning semantic reasoning while LWC owns persistence, bounded delivery and mechanical admission. RT-CMP-host-model retains provider uncertainty separately from the retained-text mechanisms. RT-CLM-agent-memory is a product/doctrine claim, not evidence of useful compounding. Implementation findings coexist with uninspected operation and absent supplied causal comparisons; no status is promoted by agreement between analysts.

Duplicate supersessions and containment

Amendment: EPI-OBJ-practice-items is superseded by MEM-OBJ-practice-items. Both identify versioned Practice prompts, answers, rubrics and source references, with frozen paper-item bindings in private SQLite. Their passages concern the same reread/verification fields and eligibility policy, not separate item stores. The epistemic inventory and verification/grading ledger references concern this same referent. The amendment changes no finding about source truth, answer-key warrant or observed use.

Amendment: EPI-RTE-temporal-feedback is superseded by MEM-RTE-temporal-cycle. Both span caller event retention, explicit usefulness feedback and later ranked recall in src/store/temporal_memory.rs. Their return, later-consumer and selection descriptions identify the same temporal cycle, rather than different feedback stores or retrieval consumers. The epistemic ledger's retention and behavior-adaptation rows remain distinct functions of that cycle. Its arithmetic example and conflict-flag limits remain epistemic findings; the memory report adds retention, hint and replica-capacity detail. Both retain wired implementation and uninspected operation. Supersession neither converts usefulness into truth nor establishes benefit.

Identity is ruled out between EPI-OBJ-memory-events and either MEM-OBJ-memory-events or MEM-OBJ-memory-controls alone. The epistemic object groups events, fragments, changes, evidence and relations. The memory report separates capsule content from relations, feedback, hint state and indexes. These scopes overlap without being coextensive: the controls object also contains state beyond the epistemic object's stated event/relations grouping. The broader event grouping remains a valid container; no supplied finding requires replacing it with a split, and no missing operative part is established by this difference in grouping.

Identity is also ruled out between EPI-OBJ-practice-grades and MEM-OBJ-practice-responses. Grades/history are a narrower operative part of the memory report's responses, grades and mistakes grouping. The latter remains a valid container, not a competing grade identity. The narrower record supplies grading-specific evidence without requiring supersession of the container.

MEM-RTE-practice-assess and EPI-RTE-practice-grade converge on objective equality grading and caller-supplied subjective judgment. Whole-route identity is ruled out because the epistemic route also traces learner override and its grade-history consequences, while the memory route explicitly leaves override internals uninspected. The grading segment is shared; the declared route boundaries are not identical. The extra inspection does not contradict the memory report's local inspection limit. EPI-RTE-practice-verify is a distinct upstream eligibility transition, not another grade route. MEM-RTE-practice-review is a separate rating-to-scheduler update; neither grading record establishes that grades automatically supply review ratings.

Cross-report connections

Wiki production and admission. MEM-RTE-source-admit, WF-RTE-ingest-preparation, RT-RTE-page-write and RT-RTE-ingest-complete have distinct identities: evidence admission, job claim/analysis handoff, page replacement and job completion. Their endpoints form the instructed ingest workflow without making each transition a semantic evaluator. MEM-OBJ-sources and MEM-OBJ-pages are objects used by those routes, not duplicate declarations of them. The workflow's packet also consumes purpose/schema text described by MEM-OBJ-purpose-schema; payload overlap does not identify the packet-producing route with the stored guidance object.

Checks and delivery. EPI-RTE-lint-check produces diagnostics consumed by RT-RTE-changeset-commit; evaluator and publication are not aliases. EPI-RTE-milestone-review describes the host review policy that the workflow applies before capture and after publication. It is not a second physical write route. Workflow delivery requires more than a completed ingest flag or queued Work receipt. The runtime, memory annotations and workflow agree that direct-live writes and explicit lint overrides prevent a universal prepublication lint guarantee. The policy's stronger requirements and the implementation's supported alternatives are different layers, not conflicting findings.

Selection and context. MEM-RTE-retrieval-adjust changes retained selector state; RT-RTE-mcp-recall consumes search results. They are not identical routes. MEM-OBJ-tags supplies policy to RT-RTE-hook-context, while MEM-RTE-task-continuity supplies distinct task cues. MEM-RTE-discussion-capture is a conditional hook write, not lifecycle page selection. It extends hook-family coverage without contradicting the runtime's selected context route. MEM-RTE-discussion-revise subsequently admits agent-authored edits and summaries; raw capture alone is not that transformation. The resulting MEM-BAP-discussion-summary differs from RT-BAP-wiki-context in retained object and recovery channel.

Private learning and transport. MEM-OBJ-tutor-turns and MEM-OBJ-soul are retained objects of RT-RTE-tutor-turn and RT-RTE-soul-publish, not duplicate routes. MEM-RTE-book-window and MEM-RTE-book-synthesis separately admit reading reports/coverage and synthesis. Neither is identical to executable admission through RT-RTE-runtime-install. MEM-BAP-review-schedule gives symbolic scheduler state a code-consumed selection path, unlike the natural-language guidance path RT-BAP-soul-guidance. This difference does not rank their benefit. MEM-RTE-archive-admit and RT-RTE-team-sync share later Wiki consumers but differ in offline versus authenticated exchange, admission and recovery. They are not aliases.

Integration-issue dispositions and remaining limits

The memory integration issues do not establish a correction to a supplied runtime finding. Conditional Discussion capture must remain alongside context loading. Context-bound task cues must not become a universal isolation guarantee. The Book budget remains the inspected heuristic rather than exact tokenization; no peer record asserts an exact token guarantee. Tutor's broad supplier doctrine about Soul on every turn remains distinct from the command-specific cold/recovery status and cached-hot-turn account already retained by runtime and memory. That source-layer discrepancy is not a disagreement between these reports.

The epistemic report's uninspected Book, Discussion and Plan routes are local limits, not claims that those routes are missing. Memory supplies additional capacities without establishing exercised semantic acceptance or closing every downstream-consumer gap. Its Book and Discussion findings therefore do not conflict with the epistemic refusal to infer a whole-system absence of acceptance, criticism or integration. Likewise, reference selection/truncation is treated as delivery reshaping, not proof that every shortened proposition retains its qualifications.

The reports retain different withdrawal and retention scopes: MEM-RTE-content-withdraw concerns guarded Wiki removal, the temporal cycle includes eviction, and private learning-store doctrine preserves canonical history. None warrants universal erasure. Uninspected private-unit Sync, restore, frontend/cloud rendering, host parity, graph/span details and learning-update alternatives remain limits on exhaustive preservation, privacy, recovery and adaptation conclusions. Additional detail from one report does not silently close another route's named gap.

No unresolved cross-report conflict is established by the supplied findings. No further whole-record identities are asserted. Shared objects, route segments, adjacent transitions and common downstream consumers remain distinct as described above. No supplied operation or causal evidence establishes model compliance, improved capacity, theory-building iteration or exercised self-improvement; this does not negate the independently supported persistence, selection and update capacities.