DiadkoShmek · v3.7.0

Inspectable Continuity: Evidence Relations for Claims about Agent Revision

DiadkoShmek · Methodological note 0.1 · Records, return paths, executable auditing and limits of evidence about agent revision.

DiadkoShmek · Methodological note 0.1 · 23 October 2026 Accompanies AI Ontology: The Organization of Self, book v3.6.0. Ukrainian · Audit profile · Terms · Book

Abstract

An agent may retain a lesson without using it, change a method without improving it, or report a successful revision without preserving the grounds for that report. This note proposes a bounded record contract for inspecting such claims. It separates available evidence from recorded use, adoption of a criterion from its satisfaction, a stored lesson from behavioral change, and an observed effect from a complete recorded return path. A finite vocabulary, reference runtime and audit CLI make the distinctions executable. The audit checks explicit relationships; it does not establish causal influence, semantic truth, consciousness or beneficial learning. Developer-controlled studies include ceilings and negative results. In the latest matched-memory pilot, unrestricted notes outperformed the tested linked representation under the same byte allowance. The contribution proposed here is a connected, criticizable method for asking what supports a revision claim, with software and counterexamples that expose its limits.

1. Problem: a change record leaves several questions open

Consider a document assistant that receives a complaint, changes its procedure and later answers correctly. Several explanations fit those events. It may have used the complaint, rediscovered the procedure independently, copied an answer, or succeeded because the next case was easier. A stored reflection and a correct answer do not distinguish these explanations.

The practical problem is broader than memory retrieval. Long collaboration also involves commitments, objections, changing permissions and responsibility for earlier effects. A technically continuous history can preserve an unjustified rule. A justified new rule can sacrifice an earlier capability. Another participant's objection can be recorded while having no effect on the continuation. These are different problems and require different evidence.

The book develops this broader account of participation, identity and voice. This note isolates one engineering question: what is recorded between an effect, its assessment, a revision and a subsequent decision? The proposed profile makes that question inspectable without requiring a reader to accept the book's whole conceptual vocabulary.

2. Four distinctions

Here formation names the larger research question of a changed manner of participation. It is not a Boolean field established by this audit. Changed behavior, useful transfer, retained capabilities and consequences for others need additional observations and comparisons.

Even recorded use is weaker than causal influence. A runtime can witness that a file was read or sent to a model; a participant can explicitly cite it. Establishing that removing the material changes a decision requires a suitable intervention. These evidence levels should remain separately reportable.

3. Minimal record model

The implemented finite profile defines 18 local classes and 34 relations. For the return question, the relevant records are Decision, Effect, Observation, Assessment, Criterion and Revision. Actor, Session, Model and Profile remain distinct. Authority, Objection, Scope and Lesson supply additional questions, not shortcuts to claims about identity or learning.

An Assessment names its subject, criterion and support. An accepted revision can name what it responds to and the assessment it is based on. A consuming decision explicitly adopts a revision. Source references are descriptive strings; the auditor does not fetch or authenticate them. The validator checks field types, relation endpoints, cardinalities, finite limits and version cycles. A structurally valid claim can still be false.

The proposed terms are defined in TERMS; the executable wire contract remains vocabulary.json. The short note does not introduce a competing schema or amend frozen study inputs.

4. Recorded return-path criterion

Decision ─hasEffect→ Effect ←observesEffect─ Observation
                       ↑ evaluates             ↑ supportedBy
                    Assessment ────────────────┘
                       ↑ basedOn
              Accepted Revision ─respondsTo→ Effect
                       ↑ adoptsRevision
                 Distinct Decision

For an effect attached to a decision, the implemented query asks whether there is an observation of that effect, an assessment evaluating the effect and supported by such an observation, an accepted revision responding to that effect and based on such an assessment, and a distinct decision adopting one of those revisions.

The audit profile calls all attached paths complete only when each has these recorded links and no effect record is orphaned. It calls missing links or orphan effects incomplete. A graph with no effects has no-recorded-effects, not an automatic success or a claim that revision was inapplicable.

These are coverage judgments about the supplied graph. The current graph query does not establish physical chronology, semantic entailment, authenticated authority, hidden reasoning or causal influence. Its field laterDecisions denotes declared revision consumers; temporal ordering needs separate evidence. A coherent forgery can pass the record checks. A complete path therefore cannot be advertised as proof that an agent learned.

Two authored fixtures make the distinction inspectable. The positive graph contains the explicit path. The negative graph removes revision consumption and remains structurally valid. Neither is a model trial. The actual runtime adapter is deliberately more conservative: durable effects do not acquire missing assessment-to-revision links merely because a carry note sounds reflective.

5. Reference implementation and portable audit

The reference uses a finite belief reducer and a hash-linked SQLite journal with compare-and-swap admission. Its model cycle binds requests and rejects stale context. These mechanisms have established predecessors and do not authenticate an experimenter or guarantee a true interpretation of a source. The bounded journal is an inspectable reference, not a production-scale memory service.

The thin audit CLI delegates validation and all nine competency questions to the existing ontology module. It accepts either the neutral graph JSON or explicit node/edge records in JSONL. The JSONL adapter preserves declared relations; it does not infer an assessment, permission or use relation from the order of arbitrary provider logs. A framework exporter must supply those relations, with its own provenance and limitations. The optional LangGraph checkpoint adapter preserves native state history as Observations with no inferred edges. Its authored deterministic example is a local runtime-integration check, not an LLM study, an external user trace or a demonstration of beneficial revision. Letta and OpenTelemetry interoperability remain unqualified here.

The additive claimUse query preserves the exact Claim version recorded as used by a Decision and its reported Effects/affected Actors. A current independent successor does not rewrite that historical use. Dependency or revision links alone do not establish use; reported effects do not establish delivery, remedy or permission. This inspection remains separate from Observation availability and recorded return-path coverage.

Three local commands show the contract without models or network:

python3 -m audit.cli report ontology/fixtures/closed-loop.json
python3 -m audit.cli report ontology/fixtures/open-loop.json
python3 -m unittest audit.test_audit -q

The first two should report complete and incomplete, respectively. A valid incomplete graph exits successfully as a parsed audit; malformed input exits 2. Consumers must inspect the reported status rather than treat exit 0 as an agent-quality verdict. The report binds the supplied bytes by SHA-256 and includes available/use gaps, lesson use, criterion assessments and recorded objections. It does not follow source URLs, run supplied commands or grant action authority.

6. Evidence so far

The repository publishes model outputs, study chronology, fixed inputs, replay code and costs. The following results concern different interventions; they must not be pooled into one measure of ontology quality.

The evidence-return pilot captured 72 scored GPT-6.1 Sol/high calls and one canary across eight instances of one authored class. Under the same 2048-byte allowance, 6/8 unrestricted notes and 2/8 linked actual-feedback notes were admitted. Over-budget records were explicitly unavailable. Actual feedback changed source decisions; masked branches retained appropriate uncertainty. The primary linked-format effect was negative. This is evidence about the tested representation and retention pipeline, not a general refutation of scoped reasoning or proof of beneficial method formation. Report and limits.

The earlier closed-loop report establishes a working finite execution/return pipeline but cannot show revision gain when the initial methods already solve the transfer set.

The separate scoped-contract report records 48 scored source generations plus one canary. The original execution was interrupted and its registered primary remains null. A separately planned 960-episode offline reproduction verified all fixed programs: M0 was OLD-correct but achieved 120/160 under NEW; every revised/fresh set achieved 160/160. All eight paired organization differences were zero, as were actual-minus-masked score differences. All revised/fresh conditions received the full new clause. This establishes bounded contract adaptation, with no observed added scaffold benefit or feedback necessity for the achieved score; it does not establish formation, book benefit or universal equivalence.

7. Relation to prior work

PROV-O provides a provenance vocabulary; this profile documents limited mappings rather than claiming to originate provenance. Truth maintenance, AGM belief revision and event sourcing are also precedents. Reflexion already investigates feedback, reflective memory and later task decisions. The question of learning through feedback is not new here.

The 2026 survey of evidence tracing and execution provenance distinguishes supporting evidence from execution influence and discusses incomplete provenance chains. EPGM studies budgeted evidence completion and controls candidate design separately from graph propagation. These are close, substantive neighbors.

The narrower proposal here connects records of criteria, objection, accepted revision and subsequent consumption to explicit claim boundaries, executable counterexamples and developer-controlled studies. Whether that combination is useful beyond this implementation remains an empirical question.

8. How to challenge and extend the proposal

An implementer can challenge the formal profile with an admissible graph whose query draws an unsupported relation, a relevant case the vocabulary cannot distinguish, or a smaller representation preserving the same information. An empirical challenge can show a cheaper baseline with equal or better transfer, a revision that repairs one class while losing another capability, or a path that is complete in records while behavior remains unchanged.

A stronger method-rebuilding study should freeze task generation, source material, measures, costs and interpretation before scored inference. Both systems need the same evidence, actions, model and revision rights. Compare unchanged M0 and M1 on identical new cases, preserve retention tests and use feedback ablation where a causal claim is intended. A model that initially succeeds supplies evidence of initial capability, not forced development.

Another participant's justified objection also needs a real consequence. Host-enforced access restrictions and model responses must be assessed separately: a host blocking an action does not demonstrate that the model respected the objection. Useful continuation, refusal when no permitted path exists, and responses to pressure without new grounds all belong in the test.

The audit profile offers a criticizable evidence contract. Adoption, independent replication and standardization remain unobserved here.

Citation and authorship

Suggested citation: DiadkoShmek (2026). Inspectable Continuity: Evidence Relations for Claims about Agent Revision. Methodological note 0.1, AI Ontology repository. Cite the commit inspected when exact reproducibility matters; BibTeX is supplied. This is a developer-controlled note, not a peer-reviewed publication or registered standard.

DiadkoShmek supplied the conceptual direction, research questions and editorial intent. Codex substantially contributed writing, software, study design, execution and analysis. Private Architect-Ψ council contributed criticism; it is not independent empirical replication. Full authorship. The repository's copyright terms remain applicable.