Operational Ontology Framework synthesis

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

Bounded synthesis

Operational Ontology Framework is a public reference for organizing AI-assisted operational work around explicit data, governing rules, action authority, evidence and durable state. This analysis covers the complete public artifact at commit 3df5d896da2a1e488e1eb325feddf190ad77990d, with an applicability cutoff of 2026-10-18. It covers only a partial operational loop: the production implementation and linked external kits are outside the frozen evidence. SRC-1 supplies doctrine/design for organizational operation and implementation evidence only for repository-publication CI. No execution trace or causal experiment was supplied, and no target or CI was executed. These distinctions define the account, rather than qualifying an otherwise demonstrated runtime (boundary, Source register).

Establish state and responsibility

The reference separates Data, Logic and Action: information available for work, rules for deciding, and authority to produce effects. Its four state roles divide continuity by purpose and intended lifetime. MEM-OBJ-pin holds stable identity, boundaries and rules; MEM-OBJ-spec holds current objectives, tasks, blockers and acceptance criteria; MEM-OBJ-handoff carries decisions, attempts, results and next action across sessions; MEM-OBJ-facts holds long-term observations with source, date and confidence. These are prescribed roles, not four demonstrated stores or published schemas. Handoff's append-only label does not establish tamper resistance, and Facts metadata does not establish accepted truth.

Standing rules are intended to direct later operators; observations and recorded results supply context; next-action guidance directs continuation. Their force is claimed, not demonstrated delivery or compliance. Storage, actual encodings, initialization and most content-admission mechanisms remain unspecified. The weekly organizational source is not established as Facts, Spec or a shared Handoff store (MEM-OBJ-authorized-source). Spec's current-work role alone does not establish later-invocation retention. No published context budget, compaction, isolation or reload mechanism establishes how these roles enter a model's working context (memory account, profile verification).

Draft, check and decide

The representative progression is a recurring weekly client review. It is selected because one non-executable example connects inputs to review and write-back, not because unrelated capabilities can be assembled into a runtime. Active clients, open commitments, deadlines, last confirmed status and a named source of truth feed an AI-produced prioritized draft. The initiating actor, context packaging, provider, model call and input-acquisition interface are unspecified. No worker delegation or separate verifier model is described in this example (RT-RTE-weekly-review, workflow account, RT-CMP-ai).

The draft rules put overdue commitments first, flag missing status instead of inferring it, require recorded confirmation before marking completion, and escalate sensitive or disputed items. The AI prepares a draft; it does not contact clients or update the source system in this example. Sorting unchanged inputs would be non-ampliative reshaping, but no actual input/output pair establishes the semantics of the whole draft. Its content relation remains indeterminate, not demonstrated derivation or established ampliative conjecture (EPI-RTE-draft-generation).

A human reviewer judges accuracy, correction needs and permission to proceed. The intended accuracy reference is confirmed organizational information; the intended use is weekly status review and authorized recording. This is an explicit content-level decision, not merely a human starting the routine. Nevertheless, no recorded evidence-consuming acceptance decision is supplied. Accuracy approval and permission to act remain separate, and neither licenses a general truth claim (EPI-RTE-source-check, EPI-RTE-review-disposition). The broader doctrine also distinguishes finding information from determining truth, and an answer from observable evidence that an action occurred (EPI-CLM-source-warrant, EPI-RTE-completion-check).

Record the approved result and continue

Approved status and remaining blockers are prescribed for recording in the organization's authorized source, with a short next-review handoff. Thus draft-only AI does not mean a write-free workflow. The physical writer, authorization interface and persistence acknowledgement are unspecified. Human approval supports claimed manual admission for the illustrated status/blocker content, not every Handoff entry, Facts update or rule amendment. The delivered-result criterion is reviewed state recorded in the authorized source plus a handoff, not the draft alone (RT-RTE-weekly-review, MEM-OBJ-authorized-source, workflow annotation).

The next operator is an explicit intended consumer, which is stronger than merely naming storage. But no selector, request initiator, delivery channel or later invocation closes the read-back chain. Neither push nor pull is established; overdue-first draft ordering is not a retained-memory selector. Recording supplies no additional truth warrant and does not demonstrate post-acceptance evidence linking or lifecycle integration (MEM-BAP-state-continuity, EPI-RTE-status-retention).

Missing or disputed information has prescribed verification, correction and escalation branches. Remaining blockers may survive recording. The source assigns no complete revision/resubmission loop or technical recovery executor. Human authority is policy owned by the organization and reviewer, dependent on authoritative sources, scoped permissions and an authorized writer. Broader conceptual Action possibilities are not current AI grants. Unpublished tool paths are neither proven mediated nor demonstrated bypasses (RT-CLM-human-authority, RT-BAP-action-authority). Stable rules require ownership and deliberate amendment, but precedence of law, security and revoked authorization is not implemented stale-copy invalidation or rollback (RT-RTE-rule-amendment).

Separate publication checks from operational improvement

The positive implementation finding is publication CI: fixed tracked-file and mode checks, credential-marker scanning, a Markdown-only implementation-pattern check and requests to three fixed URLs. Nonzero exits constrain job continuation. They do not establish mandatory merge rejection, complete credential detection, document truth or operational safety. Non-Markdown material is outside the implementation-pattern check but remains subject to other predicates; changing the checker does not preserve an immutable publication boundary. No successful CI execution is evidenced (RT-RTE-public-surface, EPI-RTE-public-check).

For the weekly-review scenario, the strongest supported contribution is a coherent prescription connecting explicit uncertainty, human correction, authorized recording and a named next consumer. The unresolved question is whether an implementation actually preserves and uses that continuity or prevents invented operational context. There is no measured comparison supporting those benefits (EPI-CLM-context-invention). Human correction improves the proposed work product by intention; it does not establish improved capacity for future action. The runtime establishes no formulated-theory revision route from which theory-builder conditions 1–4 could be affirmed separately. Reflection, autonomy, trace-fed learning and self-improvement likewise remain independently unestablished, not bundled into a system-wide negative. Rule amendment and retained outcomes alone do not establish criticism of a consumed theory, a live self-change pathway or its later exercise (RT-RTE-weekly-review, RT-RTE-rule-amendment, epistemic conclusion).

Concrete evidence that would alter this assessment includes an inspectable operational caller and permission checks; actual state/admission formats; linked input, draft, review and persistence records; a subsequent invocation receiving retained material; and evidence of the resulting behavior. A comparison measuring invented status or continuity errors could test the prevention claim. Capacity-improvement or self-improvement attribution would additionally need an identified system change, evidence shaping that change, its later behavioral consumer and the relevant outcome comparison. More state-role names alone would not close those gaps.

Limitations

Limitation	Affected SRC IDs and record citations	Inspected boundary	Conclusion prevented	Evidence that would resolve it
Operational implementation and external kits are not supplied	SRC-1; RT-ABS-agent-entry, RT-CMP-ai, RT-BAP-action-authority	Complete frozen public reference; production code and linked pages outside evidence	Actual invocation, model mutability, tool mediation, enforced authority and production security	Frozen implementation, host/provider contracts and relevant permission paths; separately captured kit evidence
State formats, initialization and maintenance remain unspecified	SRC-1; MEM-OBJ-pin, MEM-OBJ-spec, MEM-OBJ-handoff, MEM-OBJ-facts, MEM-OBJ-authorized-source, RT-RTE-rule-amendment	Public role descriptions and conceptual write-back/amendment prescriptions	Complete storage, form, provenance and admission classification; colocation, enforced append-only behavior, expiry or withdrawal propagation	Actual object schemas, provenance and admission procedures, maintained instances and invalidation paths
No concrete review, draft or completion evidence	SRC-1; EPI-RTE-data-import, EPI-RTE-draft-generation, EPI-RTE-review-disposition, EPI-RTE-completion-check	Non-executable weekly example and general evidence requirements	Source-warrant preservation, draft entailment, actual acceptance, completed effects or reliable recovery	Linked source inputs, draft, evidence-consuming reviewer decisions, action results and failure/recovery records
Continuity and improvement chains are not demonstrated	SRC-1; MEM-BAP-state-continuity, EPI-RTE-status-retention, RT-RTE-weekly-review, RT-RTE-rule-amendment	Prescribed durable recording and intended next operator; no later invocation or experiment supplied	Working read-back, selection direction/signal, lifecycle integration, trace-learning classification, improved future capacity or exercised self-improvement	Persistence and later-consumer traces, transformation/selection implementation, retained criticism or change records, and appropriately scoped comparisons
Publication checks have predicate-local force only	SRC-1; RT-RTE-public-surface, EPI-RTE-public-check	Inspected CI implementation, not executed; branch protection and external-page contents not supplied	Successful CI, mandatory merge veto, immutable publication boundary, semantic correctness or operational safety	CI run results and merge configuration for gating claims; separate semantic and operational evaluations for broader claims