WikiSkill reconciliation

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

Reconciliation

The reports share the frozen boundary and Source register. Their positive mechanism accounts describe doctrine/design, not inspected implementation or observed full-loop execution. Runtime's claimed design and uninspected implementation/operation statuses do not disagree with epistemic's doctrine-only architectural status. This reconciliation uses the current pinned reports; no source reads, correction answers or declined objections were supplied or needed. The previous reconciliation is not evidence that its conflicts still hold.

Duplicate object identities

Amendment: EPI-OBJ-raw-traces is superseded by MEM-OBJ-raw. Both identify raw filesystem trajectories containing reasoning, calls, results and answers, supported by SRC-1 §3.1's trace-description passage. The epistemic distinction between recorded occurrence and truth of embedded reasoning concerns the same payload. This connects the inventory and acquisition lineage without upgrading declared immutability to verified enforcement.

Amendment: EPI-OBJ-wiki-patterns is superseded by MEM-OBJ-patterns. Both identify Markdown pages under wiki/patterns/ with descriptions, causes and remedies; both retain SRC-1 E.2's root-cause requirement. Candidate-content assessments and memory consumer findings apply to those same pages, not separate explanatory stores.

Amendment: EPI-OBJ-wiki-index is superseded by MEM-OBJ-index. Both identify wiki/index.md and its problem/cause/fix summaries for relevance selection, citing SRC-1 E.2. This connects condensed claims to the proposer's access aid without asserting semantic preservation during complete replacement.

Amendment: EPI-OBJ-evolution-log is superseded by MEM-OBJ-log. Both identify the maintainer findings/actions narrative and cite SRC-1 E.2's append-log instruction. Both retain the logs.md versus wiki/log.md naming uncertainty. Their shared referent is the described log function, not a resolved filename or the separate harness-written audit.

Amendment: EPI-OBJ-skill-impact is superseded by MEM-OBJ-skill-impact. Both identify wiki/skill-impact.md, the harness-written proposal/diff, score and disposition history from SRC-1 §3.2.4. Audit scope and retained-intervention findings concern the same object. This does not settle whole-file reconstructability, complete maintainer delivery or causal attribution.

Amendment: EPI-OBJ-skill-content is superseded by MEM-OBJ-skill. Both identify procedural SKILL.md under skills/, using SRC-1 E.3's skill-content schema. Efficacy/applicability claims and provisional-validation/later-active consumers belong to that same file; neither every instruction's truth-aptness nor acceptance before every exposure follows.

Amendment: EPI-OBJ-skill-purpose is superseded by MEM-OBJ-purpose. Both identify companion PURPOSE.md containing origin, patterns addressed and evolution history, quoting SRC-1 E.3's rationale schema. This connects provenance assertions to retained reasons without establishing separate downstream delivery, synchronized patches or companion-file rollback.

These supersessions state identity only. All declarations and each analyst's fields remain; no finding or status is replaced by the other record's value.

Valid containers and distinct parts

EPI-OBJ-task-evidence now remains a valid container alongside EPI-OBJ-imported-evidence and EPI-OBJ-produced-predictions. SRC-1 §2 and Appendix B support incoming external evidence versus model-emitted answers/artifacts; the current inventory and ledger assign acquisition and generation to the respective parts. The prior missing-part conflict therefore no longer holds. No supersession is required because the container disclaims uniform lineage, producer and warrant. These parts are not duplicate identities of one another, of the raw training archive, or of RT-RTE-task-execution: current evidence and outputs can be retained within raw trajectories without becoming the same referent as that archive or the whole task route.

EPI-RTE-task-import is a distinct acquisition phase of RT-RTE-task-execution, not its complete reasoning/action/return route. EPI-RTE-wiki-diagnosis, EPI-RTE-wiki-formulation and EPI-RTE-wiki-retention distinguish requested checking, authoring and persistence within RT-RTE-wiki-maintenance, supported by SRC-1 §3.2.2 and E.2. Containment does not establish duplicate identity; valid runtime containers need no supersession.

Likewise, EPI-RTE-validation-scoring, EPI-RTE-skill-disposition, EPI-RTE-impact-retention and EPI-RTE-skill-integration separate checking, deciding, retaining feedback and integrating accepted instructions within RT-RTE-validation-gate, citing SRC-1 §3.2.4 and Appendix A. They neither duplicate nor supersede that valid container. Candidate validation exposure precedes acceptance; retaining wiki candidates is not post-acceptance integration.

Independent convergence and distinct claims

MEM-BAP-wiki-guidance and RT-BAP-skill-injection are not duplicate authority paths. Default wiki advice reaches maintenance/proposal through wiki context and reads; active skills reach task system prompts. SRC-1 §5.1's wiki-enabled inference alternative qualifies that separation without establishing default direct wiki access or its alternative selection budget.

MEM-CLM-rejection-history and WF-CLM-loop-evolution share SRC-1 §5.3's narrative but assert different scopes: history reuse versus the numbered ALFWorld rejection, later acceptance and refinement sequence. Neither supersedes the other or the admission mechanism in RT-RTE-validation-gate. Workflow leaves the Iteration 4 admission outcome unestablished; memory's refinement account does not contradict that qualification.

EPI-CLM-knowledge-growth assesses broader warrant language using that case, while EPI-CLM-objective-audit assesses the audit's claimed license. These are not the same historical proposition or the numerical bundled maintenance/access contrast owned by RT-CLM-wiki-benefit. Shared evidence does not make them duplicate claims. Their attribution limits converge: reported scores and simplified cases do not independently isolate a file effect or establish accepted causal explanations.

Memory's write-side account and epistemic's bounded conclusion preserve the independent theory and improvement qualifications in RT-RTE-wiki-maintenance, RT-RTE-skill-proposal and RT-RTE-validation-gate. Claimed content-guided pathways coexist with uninspected content-directed criticism and exercised causal dependence. Workflow's documented skills/wiki return adds a delivery boundary, not evidence that the improvement pathway operated or caused benefit.

Integration dispositions and remaining disagreement

The log-name question remains an exact-path limitation in MEM-OBJ-log and EPI-OBJ-evolution-log, not evidence of two logs. Diff versus full rejected-content wording remains unresolved in MEM-OBJ-skill-impact and the current EPI-RTE-impact-retention. SRC-1 §3.2.4's unified diff and E.3's full-content wording do not jointly establish reconstructable whole files. Epistemic now preserves this limit rather than requiring stronger serialization.

The prior conflict attributing definite maintainer delivery to EPI-RTE-impact-retention no longer applies to that record: its current authority, read-back and evidence-limit fields explicitly retain the same uncertainty as MEM-OBJ-skill-impact. The remaining runtime-to-specialist discrepancy is separate and remains below.

Unresolved conflict: RT-RTE-validation-gate describes later proposer/maintainer history contexts without qualifying maintainer tracker delivery, whereas MEM-OBJ-skill-impact and EPI-RTE-impact-retention retain that delivery as unresolved. Runtime's general current-wiki/history account draws on SRC-1 §3.1–3.2; the specialists distinguish broad history/full-wiki prose from E.2's explicit index/log/pattern input list, which omits the tracker, while E.3 establishes proposer access. These positions prevent a joint conclusion that the maintainer necessarily receives the complete intervention tracker. The disagreement concerns delivery completeness, not whether history exists or proposer access is declared. Reconciliation does not select either finding or correct the runtime record.

Rationale delivery and synchronized patch/restore coverage remain unknown in MEM-OBJ-purpose; EPI-OBJ-skill-purpose establishes no contrary downstream consumer. Training raw retention in MEM-OBJ-raw does not establish persistent validation trajectories or replay checkpoints merely because RT-RTE-validation-gate produces validation outcomes and restores skills. The memory annotations and runtime limits also leave no_action dispatch, malformed-edit recovery, context eviction, ALFWorld window selection and provider state unresolved. These are bounded unknowns, not disagreements from unequal detail. They prevent complete replay, crash-safe recovery, maintained rationale exposure and long-run context/isolation guarantees without erasing the described retention and read-back routes.