Sage-wiki synthesis

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

Bounded synthesis

From sources to retained knowledge

Sage-wiki is a knowledge-state producer and consumer for humans and external agents. It retains documents and agent-supplied knowledge, compiles wiki articles and graph representations, and selects retained material for later searches and answers. Go code owns progression; models supply candidate content and some judgments, not the next action in an autonomous tool loop. This analysis covers the whole shipped knowledge-state loop at commit d3ef3a1346405c82eac01b4d7c493c94c1c59169, with an applicability cutoff of 2026-10-15. Its sole source, SRC-1, supplies inspected implementation and attributed design. No execution trace, target test or causal experiment establishes operation, accuracy or benefit. External providers, services and enclosing agents remain outside the frozen repository; included but uninspected alternatives remain uncertainties, not absences. See the boundary and Source register.

The runtime has several entrances. CLI, embedded Go, MCP, REST and UI callers reach overlapping but nonidentical operations. Source capture can persist supplied bytes without a model. MCP capture instead asks a model to extract items and admits structurally valid results, falling back to raw text on extraction failure. Direct article/summary writes bypass compiler generation and answer-output trust. Caller identity does not establish human authorship or per-content approval. These differences prevent treating compilation as a universal knowledge-admission gate. See RT-RTE-capture-source, RT-RTE-capture-knowledge and RT-RTE-write-article.

Standard tier-3 compilation is the representative multi-step workflow. Code diffs sources, claims queued work, summarizes successful inputs, extracts concepts, selects article inputs and prepares articles. Independent completions receive different selections rather than a shared conversation: concept extraction sees at most 1,000 bytes of each summary; article writing rereads selected source sections and an existing article. Preparations can run concurrently, while effects are applied sequentially. Hash-matching persisted summaries can be reused after interruption. The workflow account and RT-RTE-compile-sources establish these handoffs and their recovery limits.

Delivery means retained files, attempted derived indexes/graph updates, manifest finalization and returned counts/errors—not universal success. Article files are atomically written before later index and manifest effects, which do not form one transaction. Conditional post-publication scoring can trigger one fresh rewrite; a still-low article can remain published. A failed rewrite after deletion can leave the original file absent, and pass completion can coexist with individual errors. Resumability therefore does not guarantee rollback or automatic repair of each failed article. See the workflow annotation, EPI-RTE-quality-score and EPI-RTE-quality-rewrite.

The retained representations have distinct jobs. Summaries support extraction and resume; articles supply prose and later updates; entries/chunks and dense vectors support retrieval; manifest and queue state preserve lineage and scheduling. Ontology definitions, typed relations and temporal state support traversal and graph answering. Alias decisions preserve applied identities and rejected pairs that constrain later resolution. Community summaries support a separate answer route. Prose, symbolic controls and numerical similarity representations must not be conflated with provider weights or a complete claim-by-claim justification. See MEM-OBJ-summaries, MEM-OBJ-articles, MEM-OBJ-entries, MEM-OBJ-vectors, MEM-OBJ-manifest, RT-OBJ-compile-items and MEM-RTE-resolve-aliases.

From retained state to answers and reuse

Requested search and explicit file reads can serve external agents without native answer generation. Truncated search displays can point to a full read. Within compilation, code also automatically supplies saved summaries, existing articles and relevant source sections to later completions. Thus request-led external retrieval coexists with automatic internal read-back; it does not establish unsolicited injection into a host agent's conversation. Generated host instructions prescribe search before work and capture afterward, but writing those instructions establishes only an affordance conditional on host loading and precedence. Internal prompt overrides have a separately wired completion-consumer path. See RT-RTE-search-wiki, MEM-OBJ-summaries, MEM-OBJ-articles, RT-OBJ-prompts and RT-BAP-generated-skill.

Three native answer routes consume different evidence:

• Document Q&A retrieves identities, rereads permitted article prefixes, adds graph-related material, synthesizes an answer and attempts filing. Per-article truncation and approximate byte-based budgets govern context. Returned chunk IDs can include material omitted from final context; fallback-neighbor content can lack a returned source path. Metadata is therefore not a complete snapshot of what synthesis saw. See RT-RTE-query-answer and its memory annotation.
• Local graph Q&A selects hybrid-search seeds and a bounded subgraph, optionally at a historical time, then requests an answer with edge citations. Code filters citation integers, not semantic support for the prose. Its inspected handler does not file the answer. See MEM-RTE-graph-query.
• Global Q&A selects cached community summaries, maps the question over each, discards empty or model-declared irrelevant partials, then reduces the survivors. Citations identify community origins, not verified support for each final claim. Communities must be enabled; global mode rejects historical as_of requests and does not use document-answer filing. See MEM-RTE-global-query and MEM-OBJ-communities.

For native document Q&A, returning an answer and admitting it to future retrieval are different events. The strongest inspected evidence-consuming acceptance is conditional automatic promotion for future answer reuse. A model-derived passage-support score must reach its configured threshold; confirmation count must pass its threshold; chunk-set distance must exceed 0.3 or confirmations reach twice the count threshold. These are proxy criteria, not independent verification of source truth. Agreement can come from embeddings or a model; distinct chunk identities do not prove independent sources. Grounding loads currently recorded source files, skips unreadable paths and assigns a score of one when no claims are extracted. See EPI-RTE-answer-agreement, EPI-RTE-grounding-check and EPI-RTE-verify-promote.

Promotion can move/index an accepted answer for later reuse, but this is not universal quarantine. Unrestricted filing admits directly; explicit caller promotion has different grounds; non-output articles pass the shared output predicate; known-path file reads bypass it. Automatic hooks can write confirmed state before post-transaction promotion, unlike the explicit promotion function's final state write. Source changes and rejection can withdraw later retrieval eligibility, not retract already delivered answers or guarantee erasure. See RT-RTE-admit-output, RT-CLM-review-order, MEM-OBJ-pending-outputs and EPI-RTE-output-withdrawal.

Other checks have narrower licenses. Source-count gating counts declared paths, not independent support. Triple admission checks quoted-span occurrence in a compiled summary, not relation entailment or original-source truth. Functional temporal supersession depends on configured predicates and model confidence; ambiguous and other conflicts can remain live. Consequently the design claims of graph-only grounding, quarantine until verification and contradiction-free default answers exceed what the inspected controls universally enforce. Direct ontology admission further bypasses compiler-specific graph gates, as the report verification limit records. See EPI-RTE-source-count, EPI-RTE-span-check, EPI-RTE-edge-supersession, EPI-CLM-graph-grounding, EPI-CLM-output-quarantine and EPI-CLM-edge-contradictions.

Trace-fed change and assessment

The strongest supported trace-fed contribution is a wired session-to-ontology chain: scribing filters recorded user/assistant turns, automatically extracts durable definitions/relations, and later explicit ontology traversal can return them. Tool records and thinking are removed. This is a concrete retained update with a later consumer, not demonstrated improved capacity. Pathless entities need not enter article-based Q&A. Separately, the learning-entry table implements storage, deduplication, recall and pruning, but bounded caller searches and the unfilled article-template field do not establish native lesson-to-article delivery. See RT-RTE-scribe-session, its memory annotation, MEM-OBJ-entities and MEM-OBJ-learnings.

Quality-triggered rewriting is also a supported local response to evaluation. It can replace content later consumed by the system, but supplies no content-directed criticism identifying a challenged theory and placing blame. The accepted records assess no actual formulated theory instance: condition 1 (localized theory), condition 2 (decisions depending on its content), condition 3 (content-directed criticism) and condition 4 (a next round shaped by that criticism) are not individually established. Opaque provider processing is not evidence that these are absent. Improved future capacity through criticism of a consumed theory, reflection, and autonomous theory building likewise remain unestablished independently; software-owned pipeline progression alone settles none of them. See EPI-RTE-quality-rewrite, RT-CMP-llm-client and the epistemic bounded conclusion.

At this frozen boundary, the article-rewrite route supports at most a dispositional improvement-directed pathway: quality evidence can select an update, retained replacement prose can enter later synthesis, but no subsequent operation is shown to causally depend on that change. Exercised self-improvement and successful improvement are therefore not attributed. Pack installation and prompt changes offer separate means to alter future behavior, not proof of evidence-responsive improvement. See EPI-RTE-quality-rewrite and RT-RTE-apply-pack.

For a scenario requiring persistent multi-representation knowledge and later retrieval, the source establishes concrete production, maintenance and consumer routes. For one requiring certified factual knowledge, complete provenance, guaranteed recovery or demonstrated autonomous improvement, it does not establish the required guarantees. The shipped LOCOMO evaluator uses a separate answerer and reference-answer judge over retrieval; its existence is not a native-Q&A result or a causal comparison. Traces connecting retained updates to later decisions would strengthen activation findings; controlled comparisons would be needed for benefit attribution. Complete context/dependency capture, edge-aware cache invalidation, semantic support checks and audited alternate admissions would change particular provenance, freshness and warrant conclusions. See EPI-RTE-benchmark-judge, MEM-OBJ-communities and RT-RTE-query-answer.

Limitations

All rows concern SRC-1 at the reviewed commit. No unresolved cross-report conflict was identified in the supplied reconciliation. No additional source reads or evidence were introduced for this synthesis. The profile verification retains partial inventory coverage on all ten axes and declares no additional limit.

limitation	affected SRC IDs and record citations	inspected boundary	conclusion prevented	evidence that would resolve it
Implementation is not observed operation or benefit	SRC-1; RT-RTE-compile-sources, RT-RTE-query-answer, EPI-RTE-benchmark-judge	Frozen source and attributed design; no target execution or supplied causal experiment	Reliability, activation, measured benefit and attribution of benchmark results to native answering	Inspectable invocation/state/consumer traces and appropriately controlled comparisons
Direct graph admission is omitted from the detailed route inventory; carried report-verification limit	SRC-1; MEM-OBJ-entities, MEM-OBJ-relations, EPI-RTE-span-check, EPI-RTE-edge-supersession, epistemic coverage, report limit	Compiler graph controls plus verifier inspection of direct wiki_add_ontology admission	Complete graph-admission enumeration or extension of compiler span/confidence gates to caller-authored graph writes; direct functional supersession can bypass those gates	Canonical records tracing direct entity/relation admission, temporal effects and consumers
Context metadata is incomplete and budgets are approximate	SRC-1; RT-RTE-query-answer, MEM-OBJ-entries, MEM-OBJ-pending-outputs	Article-prefix assembly, pre-assembly chunk IDs and omitted fallback-neighbor source paths	Exact model-visible provenance, complete source-change protection or a strict token ceiling	Complete delivered-context/dependency records and exact budget enforcement across branches
Checks do not establish source truth or generic semantic preservation	SRC-1; EPI-RTE-article-cleanup, EPI-RTE-quality-score, EPI-RTE-span-check, EPI-RTE-grounding-check	Deterministic cleanup/proxies and model-estimated passage support	Faithful transformation of every claim, entailed new knowledge, observed accepted ampliative claims or factual certification	Paired source/candidate evidence, warranted premises, checked interpretation and scoped acceptance records
Trust and publication guarantees are path-local	SRC-1; RT-RTE-admit-output, RT-CLM-review-order, RT-RTE-search-wiki	Explicit versus automatic promotion, unrestricted filing, non-output and direct-read alternatives	Universal quarantine, atomic publication, access revocation or erasure	Enforced common admission/read boundary and cross-store failure/ordering evidence
Compilation recovery is narrower than rollback	SRC-1; RT-RTE-compile-sources, workflow annotation, EPI-RTE-quality-rewrite	Standard pipeline, file/index/manifest sequence and single retry	All-artifact consistency or automatic repair of every completed-pass error	Failure/recovery traces and explicit repair/rollback contracts for partial writes and failed retries
Community freshness and contradiction handling are bounded	SRC-1; MEM-OBJ-communities, EPI-RTE-community-summary, EPI-RTE-edge-supersession	Membership/model cache and selected functional temporal conflicts; backend details not fully audited	Complete edge/evidence-change invalidation or contradiction-free default answers	Cache dependency/invalidation evidence and comprehensive conflict-policy/consumer tracing
Learning labels and updates do not establish improved capacity	SRC-1; RT-RTE-scribe-session, MEM-OBJ-entities, MEM-OBJ-learnings, EPI-RTE-quality-rewrite	Scribe-to-traversal wiring, bounded lesson-consumer search and proxy rewrite	General prose-Q&A delivery of pathless definitions, native lesson feedback, improved future capacity or exercised self-improvement	Wired lesson consumers; retained-update/later-operation traces; independent outcome comparisons and theory/criticism/iteration evidence where claimed
External host and provider contracts remain unresolved	SRC-1; RT-BAP-generated-skill, RT-CMP-llm-client, RT-CMP-embedder, RT-RTE-write-article	Instruction-file generation, inference interfaces and generic caller-controlled admission	Host compliance, human per-content approval, complete initial provenance, immutable model identity or provider-internal parameter behavior	Host loading/precedence and consumer evidence, caller decision records and provider version/update contracts
Batch, backend, restore and ancillary consumers are incompletely assessed	SRC-1; RT-RTE-compile-sources, MEM-OBJ-vectors, RT-RTE-hydrate-workspace, EPI-RTE-style-check, EPI-RTE-benchmark-judge	Selected standard pipeline, SQLite vector path, restore gates, style pass and LOCOMO consumer	Complete memory inventory, adapter/checkpoint equivalence, deep durability, full lint/evaluator feedback or system-wide absence findings	Records of consumed alternative payloads, admissions, later consumers, replay/backend behavior and evaluator protocols/results
Runtime safety and spending controls are not deployment guarantees	SRC-1; RT-RTE-run-parser, RT-RTE-compile-sources, RT-RTE-apply-pack	Granted Linux parser execution, between-pass cost checks and pack validation	Parser filesystem/network isolation, strict system-wide cost ceiling or semantically trustworthy installed guidance	Enforced isolation and deployment contracts, complete cost accounting across routes and substantive guidance review