claude-obsidian report verification

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

Verification

The five submitted members support publication of a bounded source account, with the two local limits below. All pinned inputs and governing contracts were read completely. The four correction-answer files are empty; the supplied structural report says none. There is therefore no structural failure to route and no correction argument to accept by default. I rechecked source HEAD against the reviewed commit and enumerated the frozen top-level tree and shipped core/script/skill paths. No target command, test, model, service or example was executed. All additional inspection uses SRC-1, implementation or doctrine/design, under the Source register.

Record and route checks

I checked every declaration and annotation for its referent, evidence layer, consequential wording and consumer, including the load-bearing passages and source anchors. The following groups give the semantic dispositions; deterministic citation and quotation occurrence checks remain the validator's responsibility.

  • Admission, effects and recovery. Independently traced CLI apply into conditional approval comparisons, locked metadata expansion, preparation, journal publication, per-file replacement, completed-result validation and restoration. RT-RTE-save-apply, RT-OBJ-transaction-bundle and RT-BAP-plan-enforcement correctly separate byte-level enforcement from human review and semantic truth. The completed-ID branch validates the stored result; mismatched bundle identity conflicts. Recovery refuses unrelated content drift. The source does not establish simultaneous multi-file visibility. Direct batch capture really calls apply without a plan approval argument, supporting RT-RTE-direct-capture. No universal human-approval guarantee survives in the reports. SRC-1: claude_obsidian/cli.py, claude_obsidian/transaction.py, claude_obsidian/capture.py.
  • Delegation and delivery. Read the Save and Ingest skills and complete ingest-worker contract independently. RT-RTE-ingest-drafts, RT-OBJ-skills, RT-BAP-skill-authority, their memory annotations and the workflow account preserve source/root selection, bounded discovery, current target hashes, partial returns and parent-only mutation. The worker's Bash capability defeats a package-enforced read-only sandbox interpretation; the reports explicitly avoid it. The workflow is a connected shipped procedure, not an invented assembly of unrelated features. An intermediate packet is not an applied vault; final operation ID and paths belong to the parent. SRC-1: skills/save/SKILL.md, skills/wiki-ingest/SKILL.md, agents/wiki-ingest.md.
  • Retained prose and provenance. Checked canonical pages, hot orientation, source metadata/raw payloads, claim ledger and legacy manifest against their distinct producer and consumer contracts. The findings in MEM-OBJ-wiki-pages, MEM-OBJ-hot, MEM-OBJ-source-ledger, MEM-OBJ-claim-ledger, MEM-OBJ-manifest and MEM-BAP-vault-evidence are supported. Setup seeds missing content; adoption can preserve existing bytes. Neither source metadata nor processing success establishes payload truth. Mixed encodings are distinguished locally rather than treating JSON as uniformly symbolic evidence. The epistemic claim/source objects describe the same ledger or its metadata portion without upgrading warrant. SRC-1: claude_obsidian/vault_ops.py, skills/wiki/references/provenance.md, skills/wiki-query/SKILL.md.
  • Evidence checks versus dispositions. Independently read the accepted-support, review-date, independence and contradiction branches, and the prospective-ledger consumer. EPI-RTE-ledger-check correctly licenses violations only within declared validator semantics. EPI-RTE-ledger-disposition correctly restricts the veto to ledger-bearing prepared writes. Nonempty adjudication notes bypass the contradiction error; their semantic adequacy is not checked. The concrete deduction is explicitly conditional, not an executed result. Research, Save, Ingest and Think support requested assessment, separate dossier/merge review and uncertainty, not an actual accepted ampliative claim. EPI-RTE-evidence-assessment and EPI-RTE-canonical-integration retain that distinction. SRC-1: claude_obsidian/ledgers.py, claude_obsidian/transaction.py, skills/autoresearch/SKILL.md, skills/think/SKILL.md.
  • Retrieval and derived state. Traced current-index/chunk/page filtering, no-rerank selection, whole-set rerank fallback, cache publication, prefix tier selection, provider fallback, bounded prefix persistence and chunk/index withdrawal. RT-RTE-vault-query, RT-RTE-optional-rerank, RT-RTE-prefix-generation, MEM-RTE-chunk-maintenance, MEM-OBJ-chunks and MEM-OBJ-embed-cache preserve the difference between ranking input, displayed raw snippet and subsequently read page evidence. Query's prescribed no-rerank branch differs materially from direct default retrieval. EPI-RTE-synthetic-prefix copies title/opening text; its reshaping classification preserves the truncation caveat. Model-produced prefixes remain indeterminate, not established conjecture or accepted evidence. EPI-OBJ-derived-context is a valid broad container with separately described consumers. SRC-1: scripts/retrieve.py, scripts/rerank.py, scripts/contextual-prefix.py.
  • Lifecycle, queue and context. Hook opt-in, vault consent, bounded hot emission and bounded journal warnings support RT-RTE-session-context; host delivery and obedience remain external. Queue token checks, completion metadata, same-host dead/stale resumption, force override and predecessor recovery support MEM-OBJ-queue and MEM-RTE-queue-resumption, not successful capture. The deletion API only proposes review. Separate worker frames and copied return packets are requested working-context controls; neither eviction nor host compaction contents are established. SRC-1: claude_obsidian/hook_adapter.py, claude_obsidian/capture.py, agents/wiki-ingest.md.
  • Maintenance, configuration and views. Fold instructions support the additive, bounded, manually approved rollup in MEM-OBJ-log, MEM-RTE-log-fold and EPI-RTE-extractive-fold, with the status disagreement retained below. Tiling and graph scoring support heuristic diagnostic attention, not semantic merging: MEM-OBJ-tiling-cache, MEM-RTE-tiling-diagnosis and EPI-RTE-diagnostic-selection. CLI mode/extension admission and read-only routing support RT-RTE-mode-extension, MEM-OBJ-mode and MEM-BAP-filing-config. Threshold installation is explicitly uncalibrated. Canvas/Bases support the requested organization surfaces in MEM-OBJ-vault-views, not observed rendering. SRC-1: skills/wiki-fold/SKILL.md, scripts/tiling-check.py, scripts/boundary-score.py, claude_obsidian/cli.py, claude_obsidian/extensions.py, scripts/wiki-mode.py, skills/canvas/SKILL.md, skills/obsidian-bases/SKILL.md.
  • Models and product checks. Checked all four components: RT-CMP-host-model, RT-CMP-haiku-prefix, RT-CMP-ollama-embedding and MEM-CMP-tiling-model. Aliases, dated identifiers and cache model strings do not authenticate immutable weights; client inference does not resolve provider training. Contract subprocess exits, pending manual gates, read-only verifier instructions and audit-before-release replacement support RT-RTE-review-machinery. Optional checkpoint lint can block Git admission without undoing retained notes, as EPI-RTE-checkpoint-gate states. RT-CLM-product-promise is correctly treated as doctrine with shipped pathways, not measured compounding benefit. SRC-1: agents/wiki-ingest.md, scripts/contextual-prefix.py, scripts/rerank.py, scripts/tiling-check.py, claude_obsidian/contracts.py, claude_obsidian/gates.py, claude_obsidian/release.py, claude_obsidian/checkpoint.py, agents/verifier.md, config/product-contract.json.

All route operation statuses remain uninspected. Executable selectors/gates justify wired implementation findings; model-facing semantic behavior remains claimed or afforded rather than demonstrated compliance. No evidenced-absence record is declared. The epistemic overlay distinguishes transformation, checking, disposition, retention, integration and operational use; its entailed derivations are limited to validator/graph/vector domains, not warranted external-world claims.

Checked every Part of: relation: assessment belongs to the ingest draft phase, prospective ledger disposition belongs to shared apply preparation, and synthetic prefix generation belongs to the prefix-tier route. Related Save/Research/Think doctrine in the assessment record does not claim that all those complete workflows are contained in ingest.

Reconciliation correctly supersedes the duplicate claim-ledger object and duplicate fold route with identity evidence, without deleting either declaration or selecting a stronger status. Its decisions not to supersede source metadata/raw containers, diagnostic containers/parts, prefix generation/maintenance, queue/capture, model uses and authority paths preserve different responsibilities. The one explicit conflict names both records, both statuses, shared instruction evidence and the prevented status conclusion. Other integration issues receive dispositions; no missing material part or contradictory guarantee was found that requires a new declaration.

Selected omission questions from the frozen tree and producer/consumer chains were: could another shipped caller defeat reviewed admission; could derived retrieval writes or lifecycle delivery defeat universal read-only/query-targeted memory; could another evidence consumer turn persistence into semantic acceptance; and could research/maintenance instructions establish automatic operational-rule learning? Capture, rerank, hook, prospective-ledger and checkpoint traces answer the first three without relying on analysts' checks. The reports already retain the consequential exceptions. Additional inspection of the research program, Think, mode router and setup wrappers found bounded/user-configurable guidance and explicit mutation protocols, not an evidenced automatic trace-fed rule update. The mode router exposes no set subcommand at this pin despite the skill's legacy warning. This does not invalidate the reported reviewed CLI route. The multi-agent setup script installs host discovery links, not canonical knowledge or learned guidance. SRC-1: skills/autoresearch/references/program.md, skills/think/SKILL.md, scripts/wiki-mode.py, scripts/setup-mode.sh, scripts/setup-vault.sh, scripts/setup-dragonscale.sh, scripts/setup-multi-agent.sh.

Omission expansion stops because these consequential challenges have supported dispositions, and no identified material challenge remains unresolved beyond the two local limits. Full extraction integrations, migration enforcement, transport internals and all syntax/helper branches were not audited. The reports explicitly withhold system-complete semantic, security, withdrawal and learning negatives. More available artifacts alone do not justify extending this review; those uninspected paths cannot support complete route inventories or universal guarantees. Positive shipped persistence/read-back and bounded evidence-gate conclusions remain supported.

Blockers

none

Limits

  • MEM-RTE-log-fold assigns implementation status afforded, while EPI-RTE-extractive-fold assigns claimed, as retained in the anchored disagreement. SRC-1 skills/wiki-fold/SKILL.md supplies a model-facing procedure with a concrete reviewed persistence/discovery path, not an executed extraction. The difference is local and does not change the shared manual/additive, no-evidence-upgrade account or justify a second independent route. Withhold one reconciled implementation-status value and any inference of reliable extractive execution; preserve both assessments and the agreed uninspected operation status.
  • MEM-OBJ-source-ledger, MEM-OBJ-claim-ledger and MEM-BAP-vault-evidence retain a consumer edge that the generic current-support rules do not fully settle: Query has no equally explicit rejected-source/deprecated-claim presentation bullet. SRC-1 skills/wiki-query/SKILL.md, skills/wiki/references/provenance.md and claude_obsidian/ledgers.py support status changes and eligibility checks, not complete automatic prose retraction. The validator and requested Query reliance conclusions remain valid, including EPI-RTE-ledger-check; withhold reliable downstream withdrawal or removal of already-delivered context. The reports already avoid those stronger claims.