Enoch synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
A personal agent with host-owned progression
Enoch's strongest supported contribution is an implemented chain from recorded conversation and task experience to reason-bearing improvement proposals and later worker context. It also turns selected conversation history into retained task briefs. These are automatic trace-fed updates with identified later consumers, not merely archival logs. Whether they improve future capacity remains unestablished: this run supplies implementation and doctrine, not a completed change-and-benefit comparison. See MEM-RTE-evidence-update, MEM-RTE-task-snapshot and MEM-BAP-candidate-guidance.
The evidence is the Enoch repository, SRC-1, at commit 0934ad59bd253adfffb8263c65e4bf0f4e2d7a9f, with applicability cutoff 2026-10-13. The boundary covers the shipped application and reference integrations as an enclosing runtime. Its whole-system functional scope is not an exhaustive inspection claim. No target, tests, examples or services were executed; no execution trace or causal experiment was supplied. External model/harness internals, live private state and installed plugin bodies remain outside inspected evidence. The Source register fixes that evidence identity.
Enoch is a personal-agent application for conversation, registered operations and repository work. Its broader claim is governed evolution from feedback and experience into tested, reviewable changes under caller adoption control (EPI-CLM-governed-evolution). The application receives events, chooses command dispatch or a model conversation, schedules retained work, and controls validation and publication. Models propose actions, answers, briefs, edits and assessments. Repository, review, chat and service adapters execute effects; external Codex or Claude harnesses execute model requests. This division matters: host progression rules are inspectable, while model compliance and provider success are not established by those rules.
On an ordinary conversation turn, the host admits an event, persists an inbox receipt, and either dispatches a command or asks the selected runtime to answer. A structured model action enters an existing command handler; its real receipt goes back to the model. The host permits at most six executable actions. A persisted running action interrupted before receipt recording is returned as uncertain, rather than blindly replayed. Reply retry and acknowledgement are separate from external action effects. Thus this is bounded, journaled progression, not a demonstrated exactly-once effect guarantee (RT-RTE-chat-actions, RT-OBJ-conversation-journal).
What survives a conversation
Local memory separates natural-language content from symbolic control fields. Long-term facts live in JSON; IDs, confidence and timestamps govern admission and ordering, while rendered fact text supplies descriptive context to a later model. Private personal identity is a separate JSON document whose selected fields supply designation, relationships, values and care guidance. Session mappings are access metadata pointing to opaque provider context, not locally retained transcripts (MEM-OBJ-long-term, MEM-OBJ-self, MEM-OBJ-codex-sessions, MEM-OBJ-native-context).
A model's final reply can request a fact save. The host extracts the text and the store rejects empty or blocked phrases, clips fields and merges normalized-text duplicates. Default explicit/high labels do not authenticate human authorship or establish truth. No per-fact human review or evidence-consuming factual acceptance is established on this branch. Whole-store replacement and unrestricted code/plugin effects are alternative surfaces, so ordinary validation is not a universal write guarantee (MEM-RTE-memory-admission, EPI-RTE-memory-store).
Later startup supply is automatic, not query-conditioned retrieval. Facts are ordered by confidence, then ascending update time, under a default 8,000-character fact budget. Selection stops at the first over-budget line rather than skipping it. The model receives ID/type/text without source references or confidence. The descriptive-use warning requests that facts not override higher-priority instructions; it does not enforce model interpretation. Application first-session injection and Codex absent-mapping injection are separate layers, with possible duplicate delivery (MEM-RTE-startup-context, MEM-BAP-retained-facts).
Recovery does not guarantee transcript continuity. Eligible Codex resume failure drops the mapping and reconstructs startup identity/facts plus the current request. The inspected Claude unavailable-session retry resends the current message without adapter-side startup-memory reconstruction. Native compaction, eviction and exact retained payload remain opaque. Deleting a fact removes it from future store renders, not from raw logs or already supplied session text (MEM-OBJ-native-context, EPI-RTE-memory-withdrawal).
From request to reviewable work
The representative workflow is a chat task requesting new repository changes. It is supported by one shipped caller chain, not inferred by joining unrelated capabilities. The conversation model first selects prior decisions, constraints and exclusions into a brief; code normalizes and clips it to 3,000 characters. Clarification or snapshot errors can stop queueing, and unavailable runtime access can produce paused work. A separate task-session invocation receives the original request and selected brief, not the full conversation transcript (MEM-RTE-task-snapshot, workflow account).
The host claims queued work with worker/process identities and supplies deadline, cancellation and progress controls. It prepares a workspace from the authoritative revision, invokes model editing, then inspects changed paths. A separate workspace is not filesystem confinement: the Codex task setting requests full access. If nothing changed, completion can mean a no-change response rather than fulfillment of the user goal (RT-RTE-task-work).
For ordinary new changes, doctor runs configured repository and readiness checks. Failure can feed bounded model repair in the same task session and under the original deadline; the default repair limit is two. Diagnostic excerpts are evidence for repair, not additional instructions. Final validation failure preserves the workspace and blocks ordinary new-change publication; it is not automatically retried as another editing attempt. Test-preservation instructions are policy, not immutable enforcement, because model work can edit checks. Doctor's license is progression under configured predicates, not correctness of every requirement or a reference answer for arbitrary work (EPI-RTE-doctor-checks, RT-OBJ-validation-diagnostics, RT-RTE-task-work).
On a pass with remaining changes, capture checks constrain changed paths and verify a clean captured revision. Review admission checks provider receipts: when remote review is required, an unconfirmed or local-only result cannot satisfy delivery. Publication stages and revision identities support retry from captured state rather than rerunning model editing. These resume and existing-result branches do not establish fresh doctor validation on every completion. Artifact handoff, task-state completion and chat notification remain separate; none proves live user receipt. The result is reviewable work, not landed or running code (WF-RTE-review-handoff).
Experience can guide a later change without warranting it
Conversation logs and task lifecycle events feed manual or due scans. Code selects pending references or changed task markers, redacts input, and calls an isolated model to extract observations and desired outcomes. Schema and known-reference checks admit signals into journals; synthesis then retains candidate rationale, proposed change, expected benefit, risk and test plan. Later curation and approved worker context consume those reasons. The default task reader's 5,000-event cap limits old-history completeness (MEM-RTE-evidence-update, MEM-OBJ-task-events).
The prompt's phrase validated evidence refers to structural admission, not independently verified observations (EPI-CLM-validated-evidence). Reference membership does not establish that a cited record supports a claim. Confidence-derived scores rank candidates without measuring expected benefit. No concrete accepted ampliative proposition or checked derivation from warranted premises is established in the selected routes; output semantics remain indeterminate without instances. Retention and evidence linkage precede demonstrated epistemic acceptance (EPI-RTE-evidence-scan, EPI-RTE-evidence-synthesis).
Curation has a narrower real evidence gate. Resolution suggestions for body changes require authoritative-body completion evidence, not partial or unpromoted task completion. Code validates authority/reference membership; a model still judges relevance to the candidate. Those predicates do not prove candidate-specific resolution or the reliability of promotion labels. Recommendation changes proposal selection; removal is a separate disposition that changes eligibility while retaining history. Later terminal regression can displace stale completion evidence, but a reverted status is not a repository revert (EPI-RTE-candidate-curation, MEM-RTE-candidate-curation, EPI-RTE-regression-status).
Other proposal inputs remain distinct. Brainstorming can use mission/theme and completed-work summaries without being evidence-backed in every branch. Published-skill assessment imports untrusted reference and retains an adaptation candidate, not an installed skill; admin learn only displays metadata. Direct-parent assessment is advisory, with assessed status gating adaptation work. Stored adoption additionally depends on completed-task and landed-review receipts, not demonstrated restarted use. Automatically recorded skill-work artifacts offer descendant-inspection guidance, but no wired descendant reader was established (RT-RTE-evolve-admission, MEM-RTE-skill-assessment, EPI-RTE-lineage-assessment, MEM-OBJ-lineage-inbox, MEM-OBJ-learning-artifact).
Admission, activation and independent property findings
The scheduler stops at proposals awaiting approval; approval queues ordinary work. Merge and update are later operations. Update requires clean/ancestry-compatible source, activates an authoritative checkout revision, runs fresh doctor and requests restart on a pass. Failure attempts rollback and exposes rollback failure. The old process remains old code until restart. These paths wire improvement-directed own-body change, but no supplied trace closes actual change, live later authority and subsequent behavior causally dependent on a particular update (RT-RTE-evolve-admission, RT-RTE-software-update, RT-BAP-software-revision).
Human-control claims are qualified by the action path. Model-issued commands reuse handlers; a human actor label is not fresh semantic authorization. The distinction between implementing a feature and authorizing merge is prompt policy. Capability and epoch checks mediate covered calls, not arbitrary in-process extensions, raw provider calls or prior full-access shell effects. Admin update is another path without chat authorization (RT-CLM-merge-authorization, RT-BAP-capability-policy, RT-RTE-extension-loading).
The property conclusions are independent:
- Trace-fed updating: wired for retained briefs and evidence-derived proposals with later consumers. This does not require novel knowledge or demonstrated improvement.
- Learning through criticism of a consumed theory: improved future capacity is not established. Task diagnosis/repair changes implementation under an unchanged method; candidate ranking is not evidence of criticism with claim-specific blame placement.
- Theory-builder conditions 1–4: the selected accounts establish neither a particular localized explanatory theory for condition 1, decisions dependent on its content for condition 2, content-directed criticism and changed reliance for condition 3, nor retained criticism shaping a next round for condition 4. Each remains unestablished within these reads, not absent throughout the system or opaque model processing. Proposal reasons and test plans alone do not close these conditions.
- Reflection: an implemented partial feedback affordance represents aspects of Enoch's own task operation in journals/signals and can mediate later proposals and own-body work. Accurate self-representation and exercised operation in both causal directions are not demonstrated.
- Autonomy: computation performs action selection, scheduling, extraction, curation and bounded repair on the inspected branches. This is not evidence that computation performs every internal operation of a theory builder, nor that adoption is globally automatic or globally authenticated human control.
- Self-improvement: a dispositional, improvement-directed own-body pathway is afforded and its constituent proposal/task/update routes are wired, with mission/theme-directed capability or reliability objectives. Exercised self-improvement and successful improvement remain unestablished at this frozen horizon. Wiring alone does not establish subsequent operation causally dependent on a change.
These bounds follow RT-RTE-task-work, RT-RTE-evolve-admission, RT-RTE-software-update and the epistemic bounded conclusion. They neither imply fixed model weights nor exclude other retained learning: external model mutability and exact versions remain uninspected (RT-CMP-codex-model, RT-CMP-claude-model).
For the scenario of preparing repository changes for review, the implemented workflow has concrete ownership, rejection and staged recovery mechanisms. For dependable recall, it has coarse startup selection and selected summaries but no established complete transcript or system-wide forgetting. For governed self-change, it separates proposal, work, landing and activation, without an invariant authenticating every human decision. These are scenario-relative findings, not a reliability ranking.
A linked execution record from selected input through candidate, review, restart and subsequent behavior would change the exercised-path assessment. A controlled comparison against an independently specified capability/reliability objective would be needed to attribute benefit, with component claims no finer than independently varied components. Immutable validation inputs or independently authenticated authorization would strengthen the corresponding guarantees. Provider checkpoint evidence and completed refresh/import traces would change continuity conclusions. This synthesis introduced no source evidence or direct source reads.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Implementation and doctrine only; no supplied operation or causal comparison | SRC-1; RT-RTE-task-work, MEM-RTE-evidence-update, RT-RTE-software-update | Frozen application producer/consumer and admission paths; no target execution | Successful delivery, exercised self-change, improved future capacity and benefit attribution | Provenance-linked execution artifacts through later operation; objective-relative intervention/comparison with confounding limits |
| Completion-authority provenance is incomplete; carried report-verification limit | SRC-1; MEM-OBJ-completion-evidence, MEM-RTE-candidate-curation, EPI-RTE-candidate-curation, report limits | Consumer-side task/promotion journal classification and authority/reference predicates | Reliability of authoritative-body labels or proof that cited change resolved the candidate | Complete promotion/deployment provenance and candidate-specific resolution evidence |
| Refresh, withdrawal and transfer coverage is partial; carried report-verification limit | SRC-1; MEM-RTE-task-snapshot, MEM-OBJ-lineage-inbox, RT-RTE-host-migration, report limits | Brief delivery, selected lineage assessment/adoption and portable-file selection; incomplete scheduled refresh, discovery/invalidation and import/rollback | Complete stale-context withdrawal and cross-host context equivalence | Traced refresh/invalidation/import/recovery predicates and provenance-linked before/after consumer contexts |
| Epistemic inventory is selective; carried report-verification limit | SRC-1; source-and-claim boundary, RT-RTE-extension-loading, report limits | Inspected memory, evidence, assessment, validation and disposition routes; visual, identity-maintenance, conformance, library and installed callback alternatives incompletely assessed | System-complete knowledge-production, acceptance, evaluator and enforcement comparisons, including global negatives | Source-grounded producer/check/consumer accounts of consequential unassessed alternatives and installed callbacks |
| Native payload, compaction and deployed model/adapter identity remain opaque | SRC-1; MEM-OBJ-native-context, MEM-OBJ-codex-sessions, RT-CMP-codex-model, RT-CMP-claude-model, boundary | Local mappings and co-shipped adapter requests, not provider internals or installed dependency bytes | Complete continuity, checkpoint classification, fixed weights, parameter-update conclusions or deployed-version equivalence | Pinned installed inventory and provider retention/version/update contracts or inspectable payload traces within a newly declared evidence boundary |
| Admission labels and prompt policy do not authenticate every human decision or mediate every effect | SRC-1; MEM-RTE-memory-admission, RT-CLM-merge-authorization, RT-RTE-extension-loading, RT-RTE-task-work | Ordinary store/handler checks, model-issued commands, full-access tasks and raw callback interfaces | Universal human-controlled admission, semantic approval, filesystem confinement or effect safety | Independent per-action authorization, complete mediation and verified deployment isolation covering alternate paths |
| Selection and clipping do not preserve complete evidence or warrant | SRC-1; MEM-RTE-startup-context, MEM-RTE-task-snapshot, MEM-OBJ-task-events, EPI-RTE-evidence-scan | Budgeted fact display, clipped briefs and capped task histories with structural reference checks | Faithful summaries, complete old-history analysis or factual endorsement of observations | Input/output instances with scope-condition and provenance preservation checks; complete selected-history accounting |
| Skill-work artifacts have no established wired descendant update | SRC-1; MEM-OBJ-learning-artifact, EPI-RTE-learning-artifact | Recorder and completed-task caller; declared descendant-inspection guidance | Successful acquired skill or concrete later behavior-shaping inheritance from the artifact | Identified later reader, retained-input lineage, validated adaptation and later consumer evidence |
The reconciliation reports no unresolved conflict. Report verification declares the three carried limits above; profile verification declares none. Its partial inventories do not imply absent native or plugin behavior.