Semiont report verification

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

Verification

The corrected reports support a bounded implementation account. The earlier generated-citation blockers are repaired. I read all supplied members, correction answers, report-check, governing types, record contract and previous verification. I judged the current wording against frozen source rather than treating the answers or earlier verdict as evidence. The structural report-check is none; there is no structural failure to route. Checkout HEAD equals the boundary's revision. No target code, tests, examples, services or model calls were executed. Source findings below are SRC-1 implementation or expressly attributed doctrine/design, not observed operation or causal benefit.

Citation correction and its consequences

I independently followed Dispatcher assignment, Worker upload/link/citation ordering, Python SDK commit and Archivist creation/commit checks. Dispatcher assigns the job on its original focus resource. Upload checks the assignment there through generatedFrom.resourceId, then creates a fresh resource stream. Worker citation commit instead targets the new resource, and the SDK attaches the original job ID. Archivist checks that target's own log before recording annotations. No assignment-copy step appears in the inspected creation progression. A nonempty batch therefore encounters the missing-assignment refusal. The SDK rethrows non-timeout refusal; the subsequent Worker completion call is not reached. An empty batch sends no commit request. SRC-1, apps/dispatcher/handlers/src/handlers.rs, apps/worker/src/semiont_worker/held_job.py, packages/sdk-python/src/semiont/claims.py, apps/archivist/src/commands.rs.

The minimum passages are now retained in RT-RTE-yield-generation and EPI-RTE-citation-retention. Their wired and implemented assessments explicitly concern construction, attempted commit and enforced refusal, not successful citation retention. Their operation assessments remain uninspected. This is a supported architectural account of a failed transition, not an unsupported positive retention value moved into a caveat.

I checked dependent wording as well as those records: Runtime account and RT-CLM-source-grounding; memory Core ideas, generation annotation, Write side, Read-back and Integration issues; epistemic citation-processing persistence wording, retention ledger, both affected claim comparisons and Bounded conclusion. They consistently preserve generated bytes/events and separately successful source binding/linking, exclude refused citation annotations from later read-back, and distinguish empty-batch completion from nonempty-batch failure. Resource creation is not rolled back by the later refusal. The three corrected answers are supported; none of their previous blockers remains.

Runtime and selected workflow

I checked every runtime declaration and its support, keeping identity, effect ordering, guarantee strength and evidence status separate. RT-CMP-worker-inference, RT-CMP-gatherer and RT-CMP-embedding-provider identify distinct configured clients, not immutable remote weights. Service/factory wiring and the inspected OpenAI request support configured access and generation without establishing provider-internal training, fixed backing weights or every provider's capability boundary. RT-OBJ-job-params correctly separates queued input from narrower announcements and lasting knowledge. SRC-1, apps/worker/src/semiont_worker/service.py, apps/librarian/src/semiont_librarian/service.py, packages/inference-python/src/semiont_inference/factory.py, packages/inference-python/src/semiont_inference/openai.py, packages/vectors-python/src/semiont_vectors/factory.py, apps/dispatcher/handlers/src/admission.rs.

RT-RTE-annotation-gather and RT-BAP-generation-prompt preserve the consequential producer/consumer distinction: returned linked-target text/summary is not separately rendered by the annotation prompt; the first textual comment/assessment body can be rendered; semantic matches are independently sorted, limited to three and clipped. Resource-focused content has separate per-resource clipping. Returned context is not a universal prompt-content guarantee. RT-BAP-worker-config describes boot configuration authority, not generated self-modification. SRC-1, apps/librarian/src/semiont_librarian/annotation_gather.py, apps/worker/src/semiont_worker/generation/resource_generation.py, apps/worker/src/semiont_worker/service.py.

RT-RTE-resource-admission, RT-RTE-mark-detection and RT-RTE-pdf-repair retain supported forcing cases. Direct bytes need not pass generation; principal/holder checks are provenance controls, not truth checks. Gateway identity replacement and service-authenticated forwarding are separate trust paths. File storage precedes command admission. Mark commit precedes checkpoint advancement; cancellation is not all-or-nothing withdrawal. Typst executes through fixed argv with no subprocess timeout argument, and compiler diagnostics can trigger at most two whole-document repairs. Its acceptance domain is compilation, not source truth. Queue retry/retention and async inference bounds do not extend these into exactly-once effects or immediate provider cancellation. SRC-1, apps/gateway/src/routes/bus.rs, apps/gateway/src/routes/content.rs, apps/archivist/src/surface.rs, apps/worker/src/semiont_worker/held_job.py, apps/worker/src/semiont_worker/processors.py, apps/worker/src/semiont_worker/generation/typst.py, apps/worker/src/semiont_worker/inference_call.py, apps/dispatcher/jetstream/src/lib.rs.

The workflow account selects a real shipped MCP gather-to-delegate caller. I reread that handler and the downstream processor: it requests the 2,000-code-point gather window, supplies neither cite nor output format, defaults to Markdown downstream, and returns success text/progress rather than generated bytes or the completion payload's resource ID. SRC-1, packages/mcp-server/src/handlers.ts, apps/worker/src/semiont_worker/processors.py.

I explicitly assessed the reconciliation's workflow conflict. The workflow's sentence “commits any generated citation annotations and completes the job” must be read within its stated selected-wrapper boundary, which explicitly excludes requesting citations. The processor produces an empty citation list without cite, and the SDK skips that empty batch. Thus the selected ordinary progression remains supported. I dismiss this as a material cross-report contradiction: the workflow does not establish a citation-enabled success path, and its failure section already allows post-upload commit failure. The sentence cannot support a broader nonempty-citation comparison, but no such comparison is asserted for the selected wrapper. The corrected parent record supplies the alternative branch. No workflow correction is required to preserve this bounded account.

Memory records

I reviewed all two component, ten object, ten route, one claim and one authority-path declarations, plus the four annotations, against their evidence and consumers. MEM-CMP-smelter-embedding produces retained vectors; MEM-CMP-matcher ranks candidates. Neither duplicates the runtime's query/gather/generation clients. SRC-1, apps/smelter/src/semiont_smelter/service.py, apps/librarian/src/semiont_librarian/service.py.

The object distinctions hold: MEM-OBJ-content is mutable file content; MEM-OBJ-event-log retains events and derived views, not old byte copies. MEM-OBJ-annotation-text differs from symbolic MEM-OBJ-annotation-links. MEM-OBJ-graph, MEM-OBJ-vector-points, MEM-OBJ-vector-payload and MEM-OBJ-anchored-text separate projection, opaque numerical search data, readable passages and extraction outcomes. MEM-OBJ-declined and MEM-OBJ-weaver-checkpoint are later-consumed maintenance state, not learned prose. SRC-1, apps/archivist/src/content.rs, apps/archivist/record/src/record.rs, packages/vectors-python/src/semiont_vectors/store.py, packages/vectors-python/src/semiont_vectors/qdrant.py, apps/smelter/src/semiont_smelter/anchored_text_store.py, apps/archivist/src/anchored.rs, apps/weaver/src/semiont_weaver/checkpoint.py.

MEM-RTE-record-replay, MEM-RTE-content-revision and MEM-RTE-annotation-edit correctly distinguish replay/staging, content update/clone and direct annotation/vocabulary changes. Historical checksum metadata is not byte rollback, and staging is not a Git commit. MEM-RTE-graph-maintenance and MEM-RTE-vector-maintenance retain unequal event coverage and distinct repair references. Weaver ignores unhandled event types even during replay; Smelter additionally follows clone/update/representation additions. Remembered declines affect later checksum/type-based work selection; unread bytes do not establish a decline. MEM-BAP-projection-control is supported without implying model training or improved capacity. SRC-1, apps/archivist/record/src/record.rs, apps/archivist/record/src/view.rs, apps/archivist/staging/src/git.rs, apps/archivist/src/commands.rs, apps/weaver/src/semiont_weaver/fold.py, apps/weaver/src/semiont_weaver/weaver.py, apps/smelter/src/semiont_smelter/smelter.py, apps/smelter/src/semiont_smelter/reconcile.py.

The selector/fallback and display qualifications in MEM-RTE-resource-gather, MEM-RTE-resource-search, MEM-RTE-reference-search and MEM-RTE-mcp-browse hold. Resource similarity uses stored resource vectors; annotation similarity embeds selected text; reference search merges several candidate sources and leaves binding to another action. Lexical hits can prevent semantic fallback. MCP list formatting can omit passages. MEM-CLM-index-stamp correctly limits compatibility to nominal provider/model/chunking checks. MEM-RTE-aggregate-writing is an afforded documented composition route, not observed autonomous synthesis. SRC-1, apps/librarian/src/semiont_librarian/resource_gather.py, apps/librarian/src/semiont_librarian/text_search.py, apps/librarian/src/semiont_librarian/reference_search.py, packages/mcp-server/src/handlers.ts, apps/smelter/src/semiont_smelter/index_stamp.py, docs/builder/skills/semiont-aggregate/SKILL.md.

Epistemic records and reconciliation

I reviewed all four epistemic objects, seven routes, three claims and each ledger function. EPI-OBJ-generated-document, EPI-OBJ-assessment-text, EPI-OBJ-gathered-prose and EPI-OBJ-reference-score distinguish generated content, judgments, request-local summaries and mixed numeric/advisory ranking results. The reports do not classify an unseen response as an established ampliative proposition merely because a model generated it.

EPI-RTE-citation-processing preserves token removal, independent ID membership and syntactic span checking without treating them as entailment. Its allowlist can exceed clipped prompt visibility. EPI-RTE-assessment-proposal establishes requested content-directed criticism; EPI-RTE-span-check establishes placement admission, with omission for not-found and warnings for doubtful matches. These are positive checks with different domains, not absent evaluation. SRC-1, apps/worker/src/semiont_worker/generation/citations.py, apps/worker/src/semiont_worker/generation/resource_generation.py, apps/worker/src/semiont_worker/detection/parsers.py, apps/worker/src/semiont_worker/detection/anchor_audit.py, apps/worker/src/semiont_worker/detection/prompts.py, apps/worker/src/semiont_worker/processors.py.

EPI-RTE-summary-generation, EPI-RTE-reference-search and EPI-RTE-knowledge-withdrawal correctly separate prose generation, candidate ranking and changed availability from factual acceptance. Summary failure and ranking fallback have different consequences; withdrawal is not recursive retraction of descendants. SRC-1, apps/librarian/src/semiont_librarian/summary.py, apps/librarian/src/semiont_librarian/resource_gather.py, apps/librarian/src/semiont_librarian/reference_search.py, apps/librarian/src/semiont_librarian/score.py, apps/archivist/src/commands.rs, apps/weaver/src/semiont_weaver/fold.py.

EPI-CLM-trusted-knowledge, EPI-CLM-assessment-concerns and EPI-CLM-conformance-scope remain attributed purpose, reuse and evaluation-design claims. Scripted inference conformance is not measured factual accuracy. The route-local conclusions about unestablished factual acceptance, theory building and exercised self-improvement are not emitted system-complete absences. SRC-1, README.md, docs/builder/skills/semiont-assess/SKILL.md, tests/conformance/librarian/README.md.

The reconciliation supports its sole supersession: the two reference-search declarations denote the same candidate-union/scoring/reply route, with compatible statuses. It does not overwrite either author's fields. I checked all four Part of: relations: citation processing/attempted retention fit inside generation; assessment proposal/span checking fit inside mark detection. Containers remain valid. Overlapping gather/summary and edit/withdrawal scopes are not exact duplicates. Production-stage objects need not be identical to their eventual stores. The anchored workflow discrepancy is disposed above; no other unresolved conflict or missing required part was identified. The reconciliation's statement about not receiving correction-answer files describes its own inputs, not a denial that this verification received answers.

Omission questions and stopping decision

I used the frozen top-level tree and producer/consumer dependencies to select challenges: whether the citation correction survives actual admission; whether another shipped caller changes the inference about human approval; whether offline restoration bypasses live command admission; and whether derived-store repair warrants complete parity. The citation trace supports the repairs. Bounded UI callback inspection shows evidence display, hints and Generate/Compose choices, not a universal post-generation approval gate. Session instructions describe event-triggered detection and warn about hearing a watcher's own writes; this affords external event-driven work, not deployed unsolicited context injection. SRC-1, packages/react-ui/src/components/modals/GatherContextStep.tsx, packages/react-ui/src/components/modals/ResourceGenerateModal.tsx, docs/builder/skills/semiont-session/SKILL.md.

Offline import restores bytes rather than resubmitting commands. I inspected its Import and extraction progression; complete restart parity was not inspected. SRC-1, apps/launcher/internal/launcher/importcmd.go:

// It untars. That is the entire mechanism. --- apps/launcher/internal/launcher/importcmd.go

The reports now acknowledge that alternative, but its admission inventory remains incomplete as stated in Limits. Projection inspection supports unequal coverage, not universal consistency. I stopped optional expansion because all selected material challenges have supported dispositions and no identified material contradiction remains unrouted. Full Browser/desktop lifecycle, remaining launcher paths, provider/backend alternatives, extractor fidelity and evaluator results remain unassessed; existing scoped claims do not require an exhaustive audit of them. Those gaps prevent broader comparisons, not the positive routes established here. The frozen target and partial-loop boundary remain faithful, so problem is not required.

Blockers

none

Limits

• The acquisition/admission inventory remains partial beyond MEM-RTE-content-revision and MEM-RTE-record-replay. Offline launcher import restores content/event-log bytes rather than resubmitting ordinary commands to vocabulary/provenance gates. SRC-1, apps/launcher/internal/launcher/importcmd.go; the minimum passage is retained above. Memory Integration issues and the reconciliation acknowledge this alternative without a complete route account. Named live-service admission and replay findings remain supported, but readers must withhold exhaustive all-acquisition/admission comparisons and complete import/restart-parity conclusions. Live command gates must not be generalized to offline restoration.