Eva-brain synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Eva-brain is a desktop knowledge-maintenance system. It asks an external coding agent to maintain linked Markdown between original documents and later questions, rather than retrieving only from raw documents at answer time. The evidence is SRC-1, the repository at commit 706fb91f4ee851cc355785f4e7c6eab37146938f, with an applicability cutoff of 2026-11-09. The whole-system boundary includes the desktop caller, Rust orchestration, vault contract, MCP navigation and structural evaluator. It excludes external runtime/model internals and actual user vaults. Inspected implementation and doctrine support the account; no supplied execution trace or causal experiment establishes successful operation or benefit (boundary, EPI-CLM-durable-knowledge).
From source selection to an admitted wiki
A person creates or imports a brain and supplies source files. Rust copies new filenames into raw/, commits the originals, and queues eligible sources. It also discovers files already deposited there. Eligibility uses source-name occurrence anywhere in the retained operation log, not content hashes or parsed success status. A rejected source can therefore remain excluded from later ordinary enqueue, and incidental substring matches can suppress another candidate (MEM-OBJ-raw, MEM-OBJ-log).
For each job, Rust captures the current Git revision and baseline lint issues, creates a branch/worktree, and dispatches Codex CLI, Claude Code or OpenCode. The external runtime owns the model's successive reads, tool calls and edits. Eva supplies the source, profile and instructions to read the vault contract, search existing pages, merge rather than duplicate, preserve provenance and update links. The parent receives a textual summary and the worktree's actual patch, not a verified reasoning trace. The durable products distinguish entity/concept prose, single-source digests, reusable analyses and symbolic page metadata (RT-RTE-source-ingest, RT-OBJ-ingest-worktree, MEM-OBJ-knowledge-pages, MEM-OBJ-summaries, MEM-OBJ-analyses, MEM-OBJ-page-metadata).
Rust, not the proposing model, owns ingest admission. It rejects protected-path changes found either in uncommitted status or in committed differences from the base. It then requires uncommitted edits, commits the proposal and an Eva-owned log entry, and compares deletions and new lint issue strings. No deletion and no new issue permits automatic merge. Otherwise a person receives a held diff and can accept or discard it, including accepting listed issues. Renames count as deletions; inherited lint issue strings alone do not trigger review. These are gate-local controls over this admission route, not physical prevention of all external-tool effects or controls over every Eva writer (RT-RTE-source-ingest, RT-CLM-structural-gate).
The evaluator licenses structural conclusions only. For example, summary provenance checking requires a nonempty source with a raw/ prefix, not source existence or entailment. The inbound-link test does not prove index reachability. Thus structurally admissible content is not thereby verified knowledge. Source copying preserves documents without authenticating their assertions; model transformation and human merge decisions do not establish evidence-consuming semantic acceptance against a recorded criterion (EPI-RTE-structural-check, epistemic conclusion).
The representative workflow ends in a successful merge followed by desktop reload of Markdown, graph and lint views—not merely the agent's summary. An ordinary hold makes the current worker return pending review. That is not a system-wide scheduling barrier: another enqueue can start a worker without checking held entries, and the shipped dev caller uses an empty-source scan that can discover other eligible files. This is a source-derived alternate route, not an observed concurrent run or proof of a production-UI click sequence (workflow delivery, MEM-RTE-desktop-reload, report verification limits).
Rejection discards the proposal but retains raw input and a rejection log. Agent/gate errors take a different failure path without automatically scheduling retry. Recovery is not a durable transaction: queue/review metadata is process-held, cleanup can ignore errors, a post-merge error need not imply rollback, and held metadata removed before a failed merge is not restored. Git history and staging support inspection, not demonstrated crash recovery (workflow recovery, RT-RTE-source-ingest).
Later consumption and optional return to memory
Accepted pages become evidence for later ingest, query, Health and profile tasks. MCP reloads the wiki per request and exposes search, full-page read, neighbors and shortest path. Search ranks lexical title/alias/body matches and returns at most ten excerpts; full reads have no server token cap. Graph results express resolved link connectivity, not evidential support. Raw and infrastructure exclusions apply to this loader, not every native filesystem tool. Retained text has both an informational role at the model and a ranking role at search (MEM-RTE-mcp-navigation, MEM-BAP-wiki-evidence).
Wiki retrieval is requested, while task startup automatically supplies the active retained profile/configuration. Settings can change later guidance and adapter/model selection; bootstrap preserves existing operating files. Each task has a separate subprocess and generated prompt, but Eva does not establish clean external histories or control host compaction. A nonempty working set changes instructions only: MCP and native tools retain wider access, and citations are not checked against the set. Empty scope permits the whole brain. Neither strict exclusion nor an actual bypass is demonstrated (RT-RTE-brain-settings, RT-BAP-vault-contract, RT-RTE-vault-query, profile verification).
Query returns a parsed, nonempty cited answer without automatically retaining it. Health requests evidence-backed diagnoses but its next steps remain advice. Seven profile tools produce such products as evidence maps, reflections, decision frames and study material; a generated answer key is not an independent correctness oracle. These requests establish capacities and intended discipline, not faithful execution (RT-RTE-advisory-tools, EPI-OBJ-profile-result).
A person can save query or profile output through a distinct route that always requires review. Rust formats the supplied answer as an analysis, index entry and log, then merges or discards it after a decision. The command does not authenticate the answer as a prior model result. Reference filtering can remove lineage while leaving answer prose intact. Later reuse can therefore be synthesis-on-synthesis without an enforced independent-source check. Saving establishes retention, not semantic acceptance or preservation of full reasoning (RT-RTE-analysis-save, EPI-RTE-analysis-format, MEM-OBJ-analyses).
What changes through use—and what remains unestablished
The strongest supported adaptive contribution is narrow but concrete: automatic operation-log writes retain task outcomes/rejection activity, and later enqueue consumes filename occurrence as an admission veto. This is wired trace-fed behavior-shaping memory under the comparison definition, not demonstrated improved capacity. Source-fed page integration also affords digesting, revision and duplicate merging, but actual semantic transformations remain unresolved without source/output pairs. Arbitrary raw content leaves possible additional trace-fed routes unknown (MEM-OBJ-log, RT-RTE-source-ingest, profile verification).
The research profile goes beyond requesting compression: it asks for competing explanations to be updated when evidence challenges a thesis. At that requested pathway, theory-builder condition 1—localized theory content—is afforded by editable thesis/explanation pages; condition 2—content-dependent consumption—is afforded by instructions to use them; condition 3—content-directed criticism with blame and revision—is claimed in guidance rather than evidenced as completed; condition 4—criticism shaping another round—is afforded by retention, not observed iteration. Persistence across tasks is wired. These independent findings do not establish a completed theory-building episode (epistemic conclusion, EPI-OBJ-profile-result).
No retained criticism/result pair establishes that criticizing a consumed theory improved future capacity. Reflection is likewise unestablished: Health's diagnosis of a wiki does not itself establish a causally connected representation of Eva's own operating machinery. End-to-end autonomous theory building is unestablished; opaque model work and human-held admission do not supply role-by-role computational completion. At the bounded desktop/vault system across later tasks, exercised self-improvement is not established either. Log-conditioned scheduling and human settings changes show behavior-determining inputs, but no objective-linked update and subsequent operation demonstrate improvement-directed self-change causally exercised. These are not claims that all such capacities are absent (epistemic conclusion, RT-RTE-brain-settings, MEM-OBJ-log).
For repeated source integration and later evidence-grounded questions, the supported result is an inspectable, editable intermediary with explicit proposal admission and reload. For verified research conclusions, strict evidence-scope exclusion or reliable unattended recovery, the inspected mechanisms establish less. Source/claim pairs with recorded semantic criteria and decisions would change the warrant assessment; tool-level scope filtering would change the exclusion assessment; durable queue/review reconstruction would change the recovery account. Traces linking retained criticism to changed reliance and another round would change the theory-building assessment. Controlled comparisons tracing retained updates to later behavior and outcomes would be needed to attribute improved capacity or successful self-improvement. No product ranking follows from this boundary.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied execution or causal comparison | SRC-1; RT-RTE-source-ingest, MEM-BAP-wiki-evidence, epistemic conclusion | Frozen implementation, doctrine and attributed reports; no target execution | Successful end-to-end operation, exercised compliance, improved future capacity or causal benefit | Provenanced runs linking sources, proposals, decisions and later consumption; objective-specific intervention/comparison |
| External runtime and model state is outside inspection | SRC-1; RT-CMP-codex, RT-CMP-claude, RT-CMP-eva-qwen, RT-RTE-runtime-setup | Eva adapters, requested permissions and mutable selectors/tags; not provider internals or installed distributions | Whole-machine containment/privacy, clean model history, immutable model identity or internal parameter-update conclusions | Frozen runtime/configuration and model identities plus inspected enforcement and run evidence |
| Imported provenance and arbitrary raw payloads remain unresolved | SRC-1; MEM-OBJ-raw, MEM-OBJ-knowledge-pages, MEM-OBJ-page-metadata, RT-RTE-brain-settings, profile verification | Known later copying, authoring and formatting; preserved initial files and generic callers not authenticated | Uniform content authorship/control, complete encoding/lineage or trace-origin classification | Provenanced input corpus and metadata derivations, with recorded admission principals |
| Structural admission and saved references do not establish semantic warrant | SRC-1; EPI-RTE-structural-check, EPI-RTE-analysis-format, RT-RTE-analysis-save | Parser predicates, gate comparison, formatter and review boolean | Source entailment, accepted ampliative knowledge, full retained rationale or independent-source support | Actual source/output pairs, retained evidence-consuming semantic decisions and post-format lineage checks |
| Working-set scope is model-directed | SRC-1; RT-RTE-vault-query, MEM-RTE-mcp-navigation | Prompt restriction, full-root tools and output parsing | Hard out-of-set exclusion; also no observed bypass conclusion | Scope-aware enforcement across all read channels and corresponding traces |
| Ordinary review pause is not universal scheduling exclusion | SRC-1; RT-RTE-source-ingest, workflow handoff, report verification limit | Hold returns current worker; enqueue can restart without checking held entries, including dev empty-source scans | “Nothing else can run until review”; observed concurrency or production-UI sequence also unestablished | Enforcement over enqueue/restart and held state, or bounded traces establishing actual interleaving |
| Recovery is process-local and nontransactional | SRC-1; workflow recovery, RT-RTE-source-ingest | Cleanup, merge/lint ordering, in-memory queues and held-item removal | Error-as-rollback or durable crash-recovery guarantee | Persistent recovery protocol and failure/restart traces verifying retained-state reconciliation |
| Requested criticism and retained state do not establish completed theory building or self-improvement | SRC-1; EPI-OBJ-profile-result, epistemic conclusion, MEM-OBJ-log | Research instructions, advisory Health, later-consumer wiring and settings | Completed criticism/iteration, reflective or autonomous theory building, improved capacity, or exercised self-improvement | Linked theories, criticisms, blame/revision and next-round decisions; self-representation evidence; objective-linked update-to-later-behavior attribution |
No additional source reads were needed. The account uses accepted records; the profile verification declares no additional publication limit.