Funes synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Funes is a file-based library operated by an externally supplied LLM agent called the Librarian. It supplies repository structure, operating instructions, extraction helpers and mechanical audits, not an enclosing model runtime. Its claimed work is to preserve sources, compile reusable knowledge, answer from the resulting wiki and maintain its coherence. This account covers the complete shipped functional loop at commit 3fb448b8e0d1ab5f6de3c6de61d8d556dab4ee11, with an applicability cutoff of 2026-10-15. SRC-1 is the single frozen Git source. The evidence is inspected implementation for helpers and CI, doctrine/design for semantic agent work, and an attributed historical compilation failure. The starter library is empty; no target execution, populated-library trace or causal experiment was supplied. See the boundary and Source register, RT-ABS-model-driver and EPI-CLM-library-abstraction.
From a source request to a retained library
A human supplies a source to a file-capable agent. Root and library-specific entry instructions direct that agent to the shared protocol. The host must actually load those instructions; Funes does not own model scheduling, permissions or context isolation. The Librarian then owns ordinary editorial decisions. Its prescribed progression is raw persistence and registration, a source-specific summary and takeaways, concept creation or merging, reciprocal links and topic organization, followed by index, registry and changelog updates. These are successive artifacts managed by one agent, not established specialist-agent delegation or software transactions. See RT-BAP-librarian-instructions, RT-RTE-ingest-compile and the workflow account.
The important memory distinction is between preserved originals, readable derivatives and interpreted wiki content. Originals supply a recovery and verification reference. PDF sidecars and image transcriptions make otherwise awkward sources readable. Source notes reduce individual sources; concept articles merge ideas across them; topic maps and indices guide later selection. These parts reside in ordinary files, with optional Git history rather than universal automatic versioning. Their prose is interpreted by readers, while paths, links and selected text also have fixed-rule consumers in the audits. Opaque binary originals and unspecified output formats do not inherit the classification of their summaries. See MEM-OBJ-raw-originals, MEM-OBJ-text-editions, MEM-OBJ-source-notes, MEM-OBJ-concepts and RT-OBJ-library-tree.
Ordinary derived edits are admitted by the Librarian without mandatory per-article human approval. A cross-library connection instead requires a proposal and human approval before forward and reciprocal links. That is an instruction rule, not an interception of arbitrary file writes. Raw immutability and citation requirements likewise depend on instruction adherence rather than a universal write barrier. See RT-RTE-ingest-compile and RT-CLM-library-guarantees.
Extraction is an alternate production path, not automatic compilation. The PDF wrapper selects native text or OCR and can write successful pages together with failure markers, returning failure status; even --fail-fast still reaches the partial-output write. Its provenance header is optional and does not record a complete reproducible configuration. The caption helper can list tracks without fetching speech or fetch and serialize one selected video's captions. Default text loses timing and in-file track provenance that JSON retains. Neither helper automatically registers or compiles its result. Recognition accuracy, service availability and source truth remain outside what these wrappers establish. See RT-RTE-pdf-sidecar, RT-RTE-caption-fetch and EPI-RTE-caption-reshape.
Checking and delivery
The protocol rejects templated or shallow compilation and asks the Librarian to redo it rather than pad it. Mechanical tools provide narrower evidence: link existence, basename-based orphan detection, opening frontmatter markers, selected n-gram duplication and takeaway/concept echoes. Depth and compression measures are proxies; the shipped push/pull-request CI runs compile lint without strict mode, so depth advisories alone do not fail it. Scripts return diagnostics and exit status, not corrected articles. Branch protection and a merge veto are not established. See RT-RTE-health-audit, EPI-RTE-mechanical-checks and EPI-RTE-audit-disposition.
Semantic checking is positively prescribed: editorial review compares article and source for fidelity and examines contradictions, grounding and gaps. Safe fixes belong to the Librarian; other changes are proposed to the human. This is not a wired semantic veto before every write. A green mechanical audit cannot prove source reading or factual grounding: distinct fabricated prose can evade the duplicate/echo predicates. Citation links permit inspection but do not establish entailment, source truth or acceptance of a new claim. No actual summary or synthesis is available to distinguish faithful reshaping, entailed derivation and ampliative conjecture. See EPI-RTE-fidelity-review, EPI-CLM-compile-depth and EPI-OBJ-wiki.
Ingest delivery is procedural: saved raw and compiled files, updated navigation, a compiled registry row, changelog updates and a human-facing summary. It is neither atomic acceptance nor proof of quality. CI and periodic/requested editorial review are separate triggers, not mandatory stages of every ingest. Explicit commit/push belongs to new-library provisioning, not ordinary ingest. Recovery is editorial redo and available committed history, without an automatic multi-file rollback or downstream retraction mechanism. See the workflow account, MEM-OBJ-source-registry and RT-RTE-library-setup.
Later reliance and change
A later human question triggers wiki-first reading through INDEX → sources → concepts. Agent judgment selects files and useful cross-library connections. Retained knowledge reaches the answering agent as interpreted evidence, not operating commands or a gold answer. This is a prescribed pull route, with no fixed token budget, automatic top-k injection or specified model-window compaction. Host loading of newly created local guidance and later use of old reports/logs remain unresolved alternatives; the entire memory boundary cannot therefore be declared pull-only. See RT-RTE-wiki-answer, MEM-OBJ-wiki-index, MEM-BAP-wiki-evidence and the profile verification.
Answers must cite sources and acknowledge missing coverage. Substantial outputs are saved, and durable portions are offered for filing into the wiki. Saving alone does not install them into ordinary wiki-first retrieval. The output-directory README describes filing more categorically than the shared protocol's offer; neither unconditional autonomous installation nor a deployed human veto follows. Dated health reports and the changelog retain diagnoses and activity, but no specified recurrent consumer closes a log-to-method-update loop. See MEM-OBJ-outputs, MEM-OBJ-health-reports and MEM-OBJ-changelog.
The strongest supported contribution toward future capability is durable, revisable knowledge organization with a named later answer consumer, plus executable defect detection. Requested consolidation, merging and correction can change what later work reads. This is not evidence that criticism of a consumed theory improved future capacity, or that trace-fed learning occurred. Generic source origins and missing later report consumers leave trace learning undetermined rather than absent. Model parameter mutation is separately uninspected. The records do not establish a particular theory's formulation, consequential consumption, content-directed criticism and retained iterative result; no theory-builder conclusion is inferred from article labels or generic review. Nor do those records establish reflective causal feedback or autonomous performance of every internal theory-building operation. These remain independent questions, not consequences of the memory finding. See MEM-OBJ-concepts, RT-CMP-librarian-model, EPI-RTE-fidelity-review and the profile verification.
Within the shipped library lifecycle, exercised self-improvement is also unestablished. Editorial changes could enter later knowledge consumption, and setup-created guidance could direct later behavior, but the evidence does not link operational criticism to an update of Funes's own organization and subsequent behavior causally dependent on that update. The source's self-improving label does not supply that chain or demonstrate benefit. See RT-RTE-health-audit, RT-RTE-library-setup and MEM-BAP-wiki-evidence.
For a source-backed library scenario, the design supplies a coherent progression from original to reusable context and a route back for verification. For a scenario requiring certified semantic admission, reproducible extraction or demonstrated improvement, the inspected mechanisms and evidence do not establish those properties. A populated source-to-article-to-answer trace would test adherence and reliance; paired editorial criticisms, decisions and later consumption would clarify learning and theory-related questions. A controlled comparison would be needed to attribute benefit. A mandatory semantic admission path, complete extraction provenance or enforced raw protection would change the corresponding architectural assessment; none is inferred from the current protocol or audit status.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| No observed semantic workflow or causal comparison | SRC-1; RT-RTE-ingest-compile, RT-RTE-wiki-answer, EPI-RTE-fidelity-review | Frozen instructions, empty scaffold, helper implementation and CI configuration; no target execution | Successful compliance, faithful compilation, activated reliance, improved capacity and attributable benefit | Inspectable populated-library runs connecting inputs, checks, edits and later answers; controlled comparisons for benefit attribution |
| Host and dependency behavior is not frozen | SRC-1; RT-CMP-librarian-model, RT-BAP-librarian-instructions, RT-RTE-pdf-sidecar, RT-RTE-caption-fetch | Shipped entry points and wrappers, not external agent/provider, OCR/Poppler or caption-service implementations | Host instruction delivery, parameter mutation, deployed isolation, recognition accuracy and network reliability | Frozen host/dependency evidence and bounded execution traces for those properties |
| Audit-results object remains coarse; carried reconciliation conflict and report-verification limit | SRC-1; EPI-OBJ-audit-results, MEM-OBJ-health-reports, EPI-RTE-mechanical-checks, EPI-RTE-audit-disposition; report-verification limits | Inspected audit/protocol routes distinguish transient mechanical diagnostics from retained editorial prose, but no separate mechanical-result object is declared | Fully separated object inventory or one uniform storage, form or retention comparison across all audit results; function-level findings remain supported | Separate accepted operative-part declaration for mechanical diagnostics, with its consumers and retention boundary |
| Caption object combines speech and track-list metadata; carried reconciliation conflict and report-verification limit | SRC-1; EPI-OBJ-captions, MEM-OBJ-text-editions, RT-RTE-caption-fetch; report-verification limits | Inspected list/fetch/serialization branches; separate list-only object undeclared | Completely separated caption-object inventory or treatment of every helper output as speech or retained memory | Separate accepted declaration of track-availability metadata and its consumption/persistence scope |
| Output admission and later report/log consumption remain unresolved | SRC-1; MEM-OBJ-outputs, MEM-OBJ-health-reports, MEM-OBJ-changelog, RT-BAP-librarian-instructions | Protocol, output scaffold, maintenance files and setup guidance | Unconditional output installation, universal human veto, complete pull-only coverage or a closed trace-fed method-update chain | Unambiguous admission policy and inspectable triggers, selected retained inputs, updates and later consumers |
| Actual semantic relations and improvement pathways are not evidenced | SRC-1; EPI-OBJ-wiki, EPI-RTE-fidelity-review, MEM-BAP-wiki-evidence | Source-fed compilation and generic editorial review at the frozen boundary | Produced accepted new knowledge, theory-specific criticism and iteration, exercised reflective feedback, autonomous theory building or exercised self-improvement | Actual claim/input comparisons, evidence-consuming decisions, retained criticisms and updates, and later causally dependent operation; independent outcome evidence for successful improvement |