Semiont synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Retained knowledge becomes input to delegated work
Semiont is a shared knowledge platform for humans and model-based agents. Its inspected architecture connects resource files and annotations to derived graph/vector stores, explicit retrieval, queued generation and subsequent knowledge writes. It is not merely an MCP adapter, nor does the traced Worker operate an unrestricted assistant tool loop. The platform's stated aim is source-grounded, trusted knowledge for accurate applications; the evidence establishes mechanisms for attribution and use, not a truth guarantee. See EPI-CLM-trusted-knowledge and RT-RTE-yield-generation.
This account uses SRC-1, the repository frozen at a920067ac4ea63a57b1855b7d263485fbd22bab4, with an applicability cutoff of 2026-10-15. Its evidence is inspected implementation and declared design. No supplied execution, model response, test result or causal comparison establishes operation or benefit. The boundary is complete artifact, partial loop: shipped clients and maintenance paths remain eligible evidence, but not all are inspected. External models, dependency implementations, corpora and enclosing assistant behavior are not established by this analysis. See the boundary and evidence and Source register.
Resource bytes and their record are distinct. Bytes occupy a mutable storage URI; sharded event logs retain accepted commands and attribution, and current views are replayed projections. An earlier checksum is not a recoverable copy of overwritten bytes. Archivist stages changed files for Git, but staging does not establish committed history or remote durability. Direct callers can upload without generation; updates, clones and annotation/body edits supply additional acquisition and revision paths. Generic caller identity does not settle whether a human made the per-content decision. See MEM-OBJ-content, MEM-OBJ-event-log, MEM-RTE-record-replay, MEM-RTE-content-revision and MEM-RTE-annotation-edit.
Weaver folds recorded events into a graph and uses saved sequence marks for catch-up and repair. Smelter reads resource bytes or annotation quotes, extracts text where needed, embeds chunks and stores vectors with readable passages and lineage metadata. Extracted text and geometry have a separate checksum-keyed cache. Neither vectors nor cache metadata authenticate the source's assertions. Repair references differ: graph replay follows events, whereas vector reconciliation compares the current catalog, checksums, media types and types. Event coverage is unequal, so reconstruction does not establish universal parity among views, graph and vectors. See MEM-RTE-graph-maintenance, MEM-RTE-vector-maintenance and MEM-OBJ-anchored-text.
Retrieval selects twice: at the service and at the model consumer
Knowledge retrieval is request-directed. Browse, text search, reference matching and resource/annotation gathering expose different retained parts. Text search uses lexical results first and permits semantic fallback only on an empty first page; an empty answer is therefore not evidence that relevant knowledge is absent. Reference matching unions name, type, neighborhood and vector candidates, optionally rescores the highest twenty with a separate model, and returns an ordering. Binding remains a separate caller action. MCP list formatting can omit the passage returned by the underlying search service. See MEM-RTE-resource-search, MEM-RTE-reference-search and MEM-RTE-mcp-browse.
Gathering assembles current source text, graph relationships and semantic neighbors, sometimes adding model-generated summaries or relationship prose. Supplementary failures can produce degraded context rather than fail the whole request. These summaries are request-local unless a caller saves or submits them. Generation then makes another selection: the annotation prompt uses the first textual comment/assessment body, hint, surrounding text and selected graph/relationship material; it does not separately render the gathered linked-target text/summary. Semantic snippets are sorted, limited to three and clipped to 240 code points each. Resource-focused contents have separate 4,000-code-point caps. These are local budgets, not a universal token bound. See RT-RTE-annotation-gather, MEM-RTE-resource-gather, EPI-RTE-summary-generation and RT-BAP-generation-prompt.
Each generation receives a selected-input frame, not the external assistant's conversation. A submitted context is a by-value snapshot without automatic refresh. Omitted originals remain available through a new explicit request; an ID in a prompt does not itself reload them. Automatic event-targeted supply also exists, but on the maintenance side: events select retained parts for folding/indexing, while startup uses progress and compatibility filters. It is not evidence of unsolicited foreground-model context injection. See RT-OBJ-job-params, MEM-RTE-graph-maintenance and MEM-RTE-vector-maintenance.
A representative workflow reaches storage before final delivery
The shipped MCP yield_from_annotation handler provides a concrete multi-step workflow, rather than a hypothetical composition. The caller names a resource/annotation and generation options. The handler gathers with a 2,000-character context-window option, then delegates the returned context. Gateway code supplies the authenticated principal; Dispatcher validates, queues and matches worker claims; Worker code rechecks filters, chooses the processor and assembles the prompt; the configured model proposes document text; Archivist owns admission. Separate configured models serve gathering, ranking, query embeddings and stored embeddings. Configuration names do not establish immutable remote weights. See the workflow account, RT-RTE-yield-generation, RT-CMP-worker-inference, MEM-CMP-matcher and MEM-CMP-smelter-embedding.
The selected wrapper requests neither citations nor an output format; its ordinary processor output defaults to Markdown. It does not include PDF repair as a routine stage. Worker uploads the artifact without an inspected per-document human approval step. Worker-role creation must cite a job assigned to that holder, while direct non-worker creation without a job remains allowed. Format, checksum, identity and assignment checks license recording, not factual reliance. File storage itself precedes command admission. See RT-RTE-resource-admission.
Upload, source linkage, citation commits and job settlement are separate effects. The citation-enabled alternative has a concrete implementation mismatch: Dispatcher records the assignment on the original focus resource, but Worker commits generated citations on the new resource. Archivist checks the commit target's own log and refuses a nonempty batch before recording. SDK refusal propagates to Worker failure handling, preventing completion without withdrawing the uploaded resource or successful source linkage. An empty batch sends no commit request and can complete, which preserves the selected wrapper's ordinary progression. The workflow's unqualified commit-then-complete sentence cannot support a nonempty-citation success claim. See EPI-RTE-citation-retention, RT-RTE-yield-generation and report verification.
Worker completion contains resource identity, name and truncation status. The MCP wrapper instead delivers success text and accumulated progress, not the document or new resource ID. A subsequent browse/gather is a different read. Cancellation noticed after model return can prevent upload, but is not immediate provider cancellation. After upload, failure is not rollback. Yield jobs have zero admission retries; mark jobs instead commit chunks before advancing checkpoints and can leave annotations after cancellation. Dispatcher sweeping and process supervision supply recovery, not exactly-once effects. See RT-RTE-yield-generation, RT-RTE-mark-detection and the workflow account.
Checks, criticism and future behavior have different scopes
There are positive checking and criticism capabilities. Assessment jobs request judgments about source passages' validity, strength and implications; anchored assessment prose can persist and later enter generation or candidate scoring. Span checks omit unplaceable proposals, but fuzzy or ambiguous placements can proceed with warnings. The assessment remains advisory, not an automatic rejection of the source. Citation-token checks can drop invalid metadata while leaving its associated sentence intact, and their allowed-ID set can exceed the text surviving prompt clipping. Neither mechanism establishes entailment. See EPI-RTE-assessment-proposal, EPI-RTE-span-check, MEM-OBJ-annotation-text and EPI-RTE-citation-processing.
PDF requests add a narrower enforcing evaluator: a local Typst compiler can reject output and feed diagnostics into at most two whole-document repairs. Compilation establishes renderability in that environment, not factual correctness or generalized learning. The subprocess has no inspected timeout argument; the async inference bound cannot be extended to it. Caller-controlled editing, deletion and archive also change later availability, but do not recursively retract previously generated descendants. See RT-RTE-pdf-repair and EPI-RTE-knowledge-withdrawal.
The strongest supported adaptation is durable outcome/progress control. Smelter remembers a bytes-read/no-usable-text decline with checksum, media type and reason; later startup skips unchanged declined inputs. Weaver sequence marks and Worker unit/cursor state similarly alter subsequent replay or resumed work. These automatic trace-fed writes qualify as local trace learning in the memory comparison, with runtime event/progress inputs. They do not demonstrate improved capacity for future action or a learned general strategy. Assessment retention likewise supplies a potential criticism-to-future-context path, not evidence that criticism of a consumed theory improved capacity. See MEM-OBJ-declined, MEM-BAP-projection-control, MEM-OBJ-weaver-checkpoint, RT-RTE-mark-detection and profile verification.
The independent properties remain separately bounded. For the inspected assessment/generation/repair routes, theory-builder condition 1 is unestablished: no supplied response identifies a formulated explanatory theory. Condition 2 is unestablished: selected content reaches prompts, but dependence on what an identified theory says is not shown. Condition 3 is unestablished: requested passage criticism does not supply an instance of blame-directed criticism with resulting revision or changed reliance. Condition 4 is unestablished: retained annotations and repeated compiler repair do not establish an iterated theory-revision result shaping the next round. These are not system-wide absences. Reflection is not established by the operational state records alone; the members do not establish the required self-representation and both causal directions. Internal job execution is computationally orchestrated, but autonomous theory building and end-to-end autonomy are not established. See EPI-RTE-assessment-proposal, RT-RTE-pdf-repair and the epistemic bounded conclusion.
For self-improvement at this partial-loop boundary, the objective-responsive decline update is a supported behavior-shaping pathway, and retained knowledge has later context consumers. Neither should be dismissed merely because its output is not code. However, no supplied subsequent operation exercises an update with causal dependence, and no comparison demonstrates successful improvement. Exercised self-improvement therefore remains unestablished, independently of the local trace-learning classification. Model/provider selection at boot is configuration-directed capability selection, not demonstrated autonomous machinery revision. See MEM-RTE-vector-maintenance, RT-BAP-worker-config and profile verification.
For the scenario of retrieving selected shared knowledge and automatically adding model-written resources or annotations, the implementation supports the service progression, local admission gates and later read-back. For a scenario requiring verified factual knowledge, complete citation provenance or atomic failure recovery, those same records do not supply the needed guarantees. Repairing the assignment/commit mismatch would alter the citation-admission assessment, not by itself the entailment assessment. Traced outputs and evidence-consuming acceptance decisions could establish what claims were checked and accepted. Subsequent update-dependent runs and a suitably controlled comparison could change the learning or self-improvement assessment. The documented conformance suite uses scripted inference/embedding stand-ins; even a passing run would primarily address protocol behavior, not factual model quality. See EPI-CLM-conformance-scope.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| Acquisition/admission inventory remains partial; offline import restores bytes rather than resubmitting live commands. | SRC-1; MEM-RTE-content-revision, MEM-RTE-record-replay, report verification limit | Named live admission/replay paths; verifier's bounded offline import inspection | Exhaustive all-acquisition/admission comparison, extension of live vocabulary/provenance gates to restoration, or complete import/restart parity | Complete restoration route and restart-state/consumer trace at the same boundary |
| Reconciliation retains an unqualified workflow citation/completion discrepancy; report verification dismisses it as material only within the no-cite wrapper scope. | SRC-1; RT-RTE-yield-generation, EPI-RTE-citation-retention, workflow storage/completion account, report verification | Ordinary wrapper omits cite; alternative nonempty citation commit targets an unassigned new resource | Combining the workflow sentence with the corrected records as evidence of nonempty citation storage followed by success | Branch-qualified workflow wording; a supported assignment-transfer/admission route would be needed to establish successful nonempty commits |
| No observed operation, actual model outputs or causal comparison was supplied. | SRC-1; RT-RTE-yield-generation, EPI-RTE-assessment-proposal, EPI-CLM-conformance-scope | Source implementation and evaluation design only | Successful end-to-end operation, output fidelity/entailment, valid criticism, accepted ampliative claims, improved capacity or exercised self-improvement | Provenance-linked input/output/check/acceptance traces, subsequent update-dependent use, and comparisons matched to the causal claim |
| Included host, launcher, provider, backend and evaluator alternatives are not exhausted. | SRC-1; boundary, RT-CMP-worker-inference, MEM-RTE-mcp-browse, MEM-RTE-vector-maintenance | Selected service paths and bounded UI/instruction reads; full Browser/desktop lifecycle, remaining launcher paths, OCR fidelity, provider alternatives and evaluator assertions unresolved | Whole-system policy/security/recovery guarantees, universal human approval or its absence, complete memory inventory, model accuracy or a system-complete negative about theory building | Relevant source and deployment-policy inspection plus scoped evaluator/output evidence |
| Projections have unequal event coverage; external models and compatibility are only nominally identified. | SRC-1; MEM-RTE-content-revision, MEM-RTE-graph-maintenance, MEM-RTE-vector-maintenance, MEM-CLM-index-stamp | Inspected folds/reconcile paths and provider/model/chunking stamps | Universal view/graph/vector parity, immutable model identity or guaranteed query/index alignment | Event-by-event producer/consumer coverage and consistency evidence; immutable model identities and verified deployed alignment |
| Durable records do not guarantee recoverable bytes, full rationale or propagated withdrawal. | SRC-1; MEM-OBJ-content, MEM-RTE-record-replay, RT-RTE-yield-generation, EPI-RTE-knowledge-withdrawal | Mutable content, event replay, optional caller instruction and current projection withdrawal | Byte rollback from checksum history, reconstruction of the full generation prompt/reasons, distributed rollback or automatic descendant retraction | Retained immutable content versions, complete derivation inputs with later readers, and explicit rollback/retraction consumer paths |
Source-read log: none. This synthesis uses the handed members and adds no source evidence.