Sage-wiki report verification

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

Verification

The runtime, memory, epistemic, workflow and reconciliation members support a bounded implementation account. No material blocker was found. One local route-inventory omission is retained below. This is not verification of deployed operation, factual accuracy of generated content or measured benefit.

Evidence and record checks

I read all supplied members, their types, the record contract and the four empty correction-answer inputs. The supplied structural check reports none; there is no structural failure to route. I independently rechecked the source checkout's full HEAD and clean status against boundary, and compared its tracked top-level tree with the coverage table. The evidence remains SRC-1 at the registered revision. Source inspection below is implementation or attributed doctrine/design, not observed operation. No target commands, tests, examples, providers or services were executed.

I checked every declaration and annotation for referent, source support, evidence strength and consumer scope, including the claim records, both behavioral-authority paths and all Part of: relations. No member declares an evidenced absence. The wired and implemented findings describe conditional code paths; they do not assert that a particular workspace enabled or exercised them. The provider records appropriately leave internal parameter changes unresolved and distinguish mutable model labels from immutable model identity. The following groups state the consequential checks and dispositions.

  • Capture and caller-authored writes. RT-RTE-capture-source, RT-RTE-capture-knowledge, RT-RTE-write-article and their retained objects distinguish byte import, model decomposition and caller authorship. SRC-1: pkg/engine/methods.go and internal/mcp/tools_write.go support exclusive source selection, structural vetoes, raw fallback, item admission and separate file/index/manifest effects. These are not mandatory semantic review. The learning-entry result can be part of the broader write route despite their different record kinds; its declaration explains that containment.
  • Compilation and delivery. RT-RTE-compile-sources, the workflow account, and the seven epistemic compile subroutes were compared against the actual handoffs and downstream consumers. SRC-1: internal/compiler/summarize.go, internal/compiler/concepts.go, internal/compiler/fullpipeline.go, internal/compiler/write.go, internal/compiler/evidence_gate.go, internal/compiler/confidence.go and internal/compiler/pipeline.go. Summary hash reuse, 1,000-byte extraction inputs, separately selected article inputs, source-count admission, publication before conditional scoring and the single fresh rewrite are supported. A failed rewrite can leave the original file absent; pass completion can still coexist with individual article errors. Resumability is not rollback or guaranteed repair. All seven compile Part of: relations preserve genuine suboperations rather than replacement candidates.
  • Retained representations and maintenance. The memory objects distinguish raw/derived text, summaries, articles, entries/chunks, dense vectors, manifest lineage, entities, relations, aliases, communities, learnings and pending outputs. Their different consumers justify keeping the mixed containers alongside epistemic content parts. SRC-1: internal/storage/db.go, internal/vectors/store.go, internal/search/pipeline.go, internal/search/graphleg.go, internal/compiler/pipeline.go, internal/compiler/resolve_entities.go, internal/ontology/aliases.go and internal/ontology/community.go. Source pruning is conditional and nontransactional. Applied aliases affect canonical lookup; retained rejection constrains later proposals. Sweep precedes the new-arbitration enablement gate. Community cache preservation depends on membership hash, not complete relation/evidence freshness. No universal cleanup or backend-equivalence guarantee follows.
  • Read-back and provenance. RT-RTE-search-wiki, RT-RTE-query-answer, MEM-RTE-graph-query and MEM-RTE-global-query have different selectors and delivery contracts. SRC-1: internal/mcp/server.go, internal/query/query.go, internal/query/graphqa.go, internal/query/globalqa.go and internal/hub/search.go. Document synthesis rereads prefixes rather than delivering selected chunks verbatim. Chunk IDs precede final assembly; fallback-neighbor paths are not appended to returned sources. The memory qualification is supported, not a contradiction of runtime. Local graph citation filtering does not reject unsupported prose. Global selection follows pickLevel, including its highest-single-summary fallback, rather than the older caller comment. Neither graph handler establishes native answer filing.
  • Trace-fed updates and learning. RT-RTE-scribe-session, its memory annotation, MEM-OBJ-entities and MEM-OBJ-learnings preserve separate chains. SRC-1: internal/scribe/session.go, internal/mcp/server.go, internal/search/graphleg.go, internal/linter/learning.go and internal/compiler/write.go. Filtered transcript text can update durable definitions; ontology traversal supplies a later consumer without making pathless entities document hits. Production caller searches and the unfilled write-template field support the bounded learning-consumer finding. They do not establish absence of external library use, improved capacity or a complete self-improvement loop.
  • Output checking, acceptance and withdrawal. RT-RTE-admit-output, RT-CLM-review-order, and the epistemic agreement, grounding, Verify-promotion and withdrawal records were traced through SRC-1: internal/trust/hooks.go, internal/trust/consensus.go, internal/trust/grounding.go, internal/trust/verify.go, internal/trust/promote.go, internal/trust/predicate.go, internal/trust/source_check.go and cmd/sage-wiki/trust_cmd.go. Agreement, passage-support scoring and disposition are distinct. Zero extracted claims scores one; unreadable source files are skipped; chunk-set distance is not independent source evidence. Conditional proxy acceptance for future reuse is supported, not truth certification. Hooks can confirm before post-transaction promotion, unlike the explicit promotion function's final state write. Direct reads, unrestricted filing and caller-directed promotion prevent universal quarantine or access-revocation conclusions.
  • Other controls and public claims. Pack installation, granted parser execution, restore, generated host guidance and synthesis guidance retain their separate enforcement boundaries. SRC-1: internal/pack/merge.go, internal/extract/external.go, internal/extract/sandbox_linux.go, internal/mirror/hydrate.go, internal/llm/injected.go, pkg/engine/engine.go, internal/serve/auth.go, internal/prompts/untrusted.go and internal/skill/writer.go. Configuration/byte-integrity checks do not confer semantic warrant; Linux parser opt-in is not filesystem isolation; injected-provider cancellation needs cooperation; between-pass cost checks are not a hard spend ceiling. Host instruction delivery does not establish compliance. The three epistemic README claims remain attributed design claims, and their comparisons correctly withhold universal graph grounding, quarantine and contradiction freedom.
  • Lint and evaluation. EPI-RTE-style-check, EPI-RTE-style-fix and EPI-RTE-benchmark-judge have supported targets and consequential consumers. SRC-1: internal/linter/runner.go, internal/linter/passes.go, eval/benchmarks/common/sagewiki.py and eval/benchmarks/locomo/run.py. Style repair adds default metadata without a domain-truth judgment. LOCOMO retrieves through the backend but supplies a separate answerer and reference-answer judge; its labels cannot be attributed to native Q&A or treated as causal component evidence.

Reconciliation and omission searches

I checked all three supersessions: raw source text, local graph answering and global community answering are matching referents with compatible findings. Supersession does not erase epistemic qualifications. The article, output, community and graph prose/state overlaps are not wholesale duplicates of their mixed memory containers. The workflow annotation adds progression to the existing compile route. The remaining distinct routes and claims do not require supersession. No unresolved cross-report conflict is concealed: learning delivery, fallback-neighbor provenance, scribe consumers, community caching and publication/recovery are compatible refinements. No correction answers contest these dispositions.

Omission questions were selected from the frozen tree and producer/consumer dependencies, not only the cited snippets: whether another registered graph admission changes the semantic-gate account; whether a real learning consumer closes the advertised loop; whether graph/global answering bypasses article context and filing; whether cached communities track changed evidence; and whether shipped lint/evaluator callers establish additional acceptance or feedback. Dispatch inspection, production caller searches and the inspected consumer implementations settle the load-bearing comparisons above. The direct ontology tool adds a local omitted admission route, retained as a limit rather than inferred absent.

Specifically, SRC-1: internal/mcp/tools_write.go shows caller-provided relation insertion without the compiler's evidence-span/confidence admission. When temporal behavior is enabled, a functional manual relation can supersede old targets directly; a bare contradiction instead emits a deferred conflict. The report's coverage already identifies the detailed handler as uninspected, so this does not falsify its scoped compiler findings or claim comparison.

if err := s.ont.AddRelation(ontology.Relation{ ID: relID, SourceID: sourceID, TargetID: targetID, Relation: relType, ValidFrom: time.Now().UTC().Format(time.RFC3339), // manual add = asserted now (P3-6) }); err != nil { --- internal/mcp/tools_write.go

Further omission expansion stopped because the selected questions have supported dispositions, and no identified challenge remains that would overturn the retained bounded conclusions. Full batch/provider parity, deep restore/backend consistency, other evaluator results and real host behavior remain outside established comparisons. Their available artifacts do not justify either exhaustive negative conclusions or an open-ended audit. No narrowing of supported positives was needed.

Blockers

none

Limits

  • The graph-admission inventory is partial: MEM-OBJ-entities, MEM-OBJ-relations, EPI-RTE-span-check, EPI-RTE-edge-supersession and the epistemic coverage table do not separately record the direct wiki_add_ontology route. SRC-1: internal/mcp/tools_write.go admits caller-authored entities/relations and can apply functional temporal supersession without the compiler's span/confidence gate. The compiler-specific controls and heterogeneous-admission conclusion remain valid; readers must withhold a complete graph-admission inventory or any comparison that generalizes those compiler gates to every graph write.