llm-wiki report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The supplied reports support publication as a bounded source-level account, with the limits below. I read the complete supplied reports, types, contracts, reconciliation, prior verification and correction answers. The artifact check reports no structural failures. I judged the records afresh against frozen source text rather than accepting the previous verification or the analyst's correction answer as evidence. Checkout HEAD matches the boundary and its working tree is clean. No target, test, example, provider or service was executed. Evidence remains SRC-1, implementation and doctrine/design only, under the Source register.

Correction and workflow ownership

The earlier epistemic blocker is repaired. I independently read single-path credibility review, Plan Mode dispatch and cross-path compilation in SRC-1 claude-plugin/commands/research.md. Each path-agent receives the full search → credibility → ingest pipeline, launches its own search agents and writes selected raw sources before returning. The foreground compilation barrier subsequently reads those files. The corrected owner prose, delegated visibility, selection, effects, ledger timing and bounded conclusion in EPI-RTE-source-review and EPI-RTE-source-selection all preserve that distinction. No dependent passage still assigns plan-path pre-ingest review to the foreground model. The load-bearing dispatch and compilation passages are retained in those records.

The selected workflow and RT-RTE-research-paths agree with this progression. User plan approval is not per-source acceptance. Search-worker findings need not include an entire fetched document; later compilation separately requests source-file reads. Path prefixes and deferred indexes coordinate intended writes but supply no enforced ownership. SRC-1 claude-plugin/skills/wiki-manager/references/research-infrastructure.md confirms per-path source targets, retry/skip choices, retained partial raw files, ephemeral plan state and durable provenance. It also confirms the filename-prefix variation and one-session-per-wiki limitation. The exact failed-path skip transition remains unspecified, as the reports already state.

Runtime and memory findings

I reviewed every declared record, its source anchors, evidence strength and consequential fields, including annotations and behavioral-authority paths. Independent source tracing checked the following groups:

  • Source compilation and access metadata. SRC-1 claude-plugin/commands/compile.md, claude-plugin/skills/wiki-manager/references/compilation.md, claude-plugin/skills/wiki-manager/references/ingestion.md and claude-plugin/skills/wiki-manager/references/indexing.md support model-operated admission and synthesis, separate raw bodies/article bodies/frontmatter/indexes, incremental date selection, conversation provenance and strict-guide confirmation. Self-validation follows article writes and is not rollback. Count-only index detection cannot certify current summaries. These support RT-RTE-compile-wiki, MEM-RTE-admit-sources and the raw, wiki, frontmatter and index objects in memory. Collection version preservation and refresh overwrite are genuinely inconsistent source policies; neither establishes filesystem immutability.
  • Query and host authority. I read both Pi launchers, SRC-1 profiles/ds4/pi-query-tools.ts, profiles/query-lite/SKILL.md and claude-plugin/commands/query.md. They support the requested compact tool surface, caller-argument forwarding, provider configuration and payload controls, not host isolation, fixed weights or reliable model compliance. Quick, deep, resume and ambient retrieval must not inherit one uniform write policy. RT-CMP-host-model, RT-OBJ-query-profile, RT-RTE-query-wiki and RT-BAP-query-instructions retain those distinctions.
  • Local maintenance and withdrawal. SRC-1 scripts/llm-wiki dispatch, lint, archive and retract paths support structural reporting/fixes, access-changing archive moves, and literal mutation followed by a rescan. Partial effects can precede failure. Literal replacement is not source-dependent semantic cleanup or factual refutation. These support RT-RTE-lint-wiki, MEM-RTE-withdraw-memory and EPI-RTE-privacy-retract, without certifying every fix branch.
  • Operational memory. I traced SRC-1 scripts/llm-wiki-session configuration loading, feedback classification, state updates, digest rendering, hook output, rehydration and both promotion functions, with plugins/llm-wiki/hooks/hooks.json. The chain supports RT-RTE-session-bridge, RT-BAP-session-context, the memory state/digest/feedback objects and MEM-RTE-feedback-candidates. Rehydration supplies state-derived summaries and digest pointers, not digest bodies. Manual summaries bypass payload redaction; the template is not semantic transcript compaction. Missing configuration defaults enabled/balanced despite opt-in doctrine. Digest promotion and candidate promotion consume different inputs, neither with an interactive factual acceptance decision.
  • Executable capability and export admission. SRC-1 scripts/llm-wiki request/response checks, subprocess dispatch, specialist enablement and checkpoint seal/verify support the bounded authority account. Adapter checks precede and follow effects but trust remote attestations; filtered environment is not a sandbox. Specialist admission validates structure, not quality. Checkpoint hashes, scanner findings and the word floor do not prove confidentiality or semantic completeness. Final copy/import are separate. See RT-OBJ-adapters-registry, RT-RTE-adapter-execute, RT-RTE-specialist-enable and RT-RTE-checkpoint-export. Their passages and SRC-1 claude-plugin/skills/wiki-manager/references/specialists.md and claude-plugin/skills/wiki-manager/references/checkpoints.md support the stated boundaries.

Topic guidance and specialist text have interpreted instructional authority, distinct from factual article evidence and symbolic allowlists. Compilation preflight and librarian adoption rules support MEM-OBJ-schema, MEM-BAP-topic-guide and RT-BAP-specialist-methods. Persistent editable methods afford later behavior changes but establish neither exercised improvement nor method quality. Research provenance remains distinct from harness-session state in MEM-OBJ-session-provenance.

Epistemic checks and reconciliation

The epistemic object inventory identifies candidate content and diagnostic judgments, not accepted truths. SRC-1 claude-plugin/commands/research.md supports thesis decomposition, opposing evidence, verdict criteria and weakest-side next-round search. SRC-1 claude-plugin/skills/wiki-manager/references/audit.md and claude-plugin/skills/wiki-manager/references/librarian.md support substantive evidence-directed review and separate reports, not a universal query gate. SRC-1 claude-plugin/commands/refresh.md supports sampled comparison and user disposition; claude-plugin/commands/ll.md supports lesson extraction, article appends and separately approved instruction proposals. Thus EPI-RTE-thesis-verdict, EPI-RTE-trust-review, the refresh and lesson records, and all three epistemic claim comparisons preserve the distinction between doctrine, retention and evidenced acceptance. The memory counterparts agree. Today's verified date is not independent evidence of human confirmation. Missing source-to-output histories leave synthesis semantics indeterminate; they do not establish absent inference or absent checking.

I independently checked the classifier, byte-replacement and term-grader deductions without running them. The wrong and thanks classifier cases follow the code; a required-term test can accept a negated term. SRC-1 scripts/benchmark-token-efficiency guards fixture paths before delivery, truncates large files and reports file bytes separately from delivered content. The three cited JS assertion bodies have their reported local limitations. These support EPI-RTE-feedback-classification, EPI-RTE-feedback-promotion, RT-RTE-query-evaluation and EPI-RTE-fixture-evaluation, not general factual accuracy or measured improvement. SRC-1 tests/promptfooconfig.yaml supplies configured model-rubric, skill-use and cost checks, not outcomes.

All explicit Part of: relations are substantively supported: the three epistemic research functions belong within research; fixture delivery/checking belongs within benchmark evaluation; session state and digest are retained data within the session route. The separately inspected JS comparisons are not asserted to be benchmark components. Both reconciliation supersessions identify the same lesson and refresh workflows while preserving lens-specific fields. The broader raw/article/feedback inventories and combined withdrawal account remain valid overlapping containers, not interchangeable whole records. No additional supersession is required, and no evidenced-absence record is declared.

The earlier research-owner conflict is now resolved by corrected records, not by reconciliation selecting a stronger status. The remaining refresh delegated-visibility conflict cites both positions and their evidence. Agent permission alone establishes no concrete worker packet; the foreground protocol remains supported while alternatives are uninspected. Its comparison uncertainty is local, retained below. Raw preservation, query repair, enabled defaults and verification dates are source-level discrepancies properly qualified across the members, not hidden contradictions. Theory-builder conditions, learning, reflection, autonomy and self-improvement remain independent and do not acquire observed or causal status from retention or requested iteration.

Omission questions and stopping decision

I used the tracked top-level tree, runtime-specific entry points and state producer/consumer dependencies to select consequential alternatives. Could another query route write despite compact policy? Explicit query versus ambient indexing already bounds that claim. Could session delivery conceal digest bodies or automatic canonicalization? The complete selected transformation chain supports summary/pointer delivery and separate promotion. Could maintenance defeat raw preservation or human-verification readings? Refresh and compilation do, and the reports retain both counterexamples.

Could unassessed admission planes overturn the epistemic conclusion? Additional inspection of SRC-1 claude-plugin/skills/wiki-manager/references/ideas.md confirms explicit shaping approval and a retained decision snapshot; claude-plugin/skills/wiki-manager/references/datasets.md specifies bounded observations, samples and native-access recipes. These are differentiated surfaces, not evidence of a product-wide absence of acceptance. MEM-OBJ-inventory, MEM-OBJ-datasets, MEM-OBJ-projects and the epistemic inventory already retain partial coverage rather than that negative.

I also searched portable/package entry points and synchronization paths and rechecked identical hashes for canonical and packaged CLI/session helpers. Source-level copying does not establish cross-host behavioral equivalence. These added reads supply implementation or doctrine/design only, never observed operation. Further omission expansion stopped because the selected challenges have supported dispositions and no identified material challenge remains undisposed. Uninspected collection, native-payload, output, package and evaluator detail prevents exhaustive comparisons; more artifacts alone do not undermine the positively traced host/model/helper account. Unsearched behavior is not treated as absent.

Blockers

none

Limits

  • MEM-OBJ-datasets, MEM-OBJ-inventory, MEM-OBJ-projects and the epistemic coverage inventory do not establish a complete enumeration of native dataset transformations, collection fetch branches, output admission, package transformations or evaluation consumers. Dataset profile/sample/recipe doctrine and Idea approval demonstrate additional differentiated surfaces, not an absence of checking or acceptance. Withhold exhaustive route counts, cross-host equivalence and system-complete epistemic negatives; the positively traced compile/query/session and selected maintenance mechanisms remain supported.
  • MEM-RTE-refresh-sources leaves delegated visibility uninspected, while EPI-RTE-source-refresh calls it inapplicable to the foreground protocol; reconciliation preserves both. SRC-1 claude-plugin/commands/refresh.md declares Agent permission but specifies no delegated packet or return. Withhold a complete comparison of refresh context isolation; the sampled comparison, user-choice and overwrite findings hold.
  • The workflow recovery account and RT-RTE-research-paths support retry/skip and retention of partial raw artifacts, but SRC-1 claude-plugin/skills/wiki-manager/references/research-infrastructure.md does not specify the exact skip-state transition at the all-path compilation barrier. Withhold a complete failed-path delivery/recovery guarantee. The successful-path serialization and offered user recovery choices remain supported.