Skip to content

About

Bilingual book and reproducible studies of agent memory, evidence, and decision revision.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

33 Commits

Folders and files

Repository files navigation

AI Ontology: The Organization of Self

AI Ontology — memory, evidence and revision

DiadkoShmek · Book v3.7.0 · English and Ukrainian

A research book, proposed audit profile and executable reference for inspecting claims about agent continuity, evidence and revision. Which grounds reached a decision? What followed from it? Did a consequence enter a subsequent method, and what evidence supports that claim?

Start here

Український README

Four distinctions

Keep separate Question
Available evidence / recorded use Was material accessible, or is its use actually recorded?
Criterion adoption / satisfaction Was this version accepted, and was it met?
Stored lesson / formation Did anything change beyond the memory record, and did that help?
Recorded effect / complete recorded return path Are the intervening assessment, revision and consumer links present?
Decision → Effect → Observation → Assessment → Accepted revision → Distinct consuming decision

The diagram summarizes required relations; it does not prove chronology or causality. The graph's recordedPathComplete checks explicit links. A complete path does not establish beneficial learning, truth, authenticated authority or subjective experience. The book's broader questions of identity, voice and participation remain distinct from this finite audit.

Try it without a model

Python 3.11+ standard library; run from the repository root:

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

Expected coverage: complete then incomplete. Both authored fixtures are valid and exit 0; invalid input exits 2. They are explanatory cases, not LLM trials. Reports include source SHA-256, missing links and nine competency queries. Projected JSONL is supported; arbitrary raw logs and framework-specific exports require an explicit adapter. Contract and limits.

An optional native LangGraph example captures actual checkpoint history from authored deterministic nodes. Its adapter preserves Observations and leaves participation links unrecorded. No model or external effect is implied.

Inspect exact historical Claim use: a new conclusion can have independent grounds while an earlier recorded decision still names the old version. The query preserves its reported consequence and affected actor without inferring delivery, repair or permission.

What is implemented

  • Finite ontology: 18 classes, 34 relations, nine queries, neutral JSON and JSON-LD/Turtle exports. No general OWL reasoner or SHACL processor.
  • Python reference: scoped claims, provenance, bounded SQLite/CAS durability and a model cycle. The actual-store adapter preserves missing evidence.
  • Book: sixteen parts on identity, voice, memory, relationships and revision, with separate Ukrainian and English editions.
  • Concept-to-code map and review guide.

This is a developer-controlled research reference and proposed profile, not a production memory service, adopted standard or independently replicated finding. Structural validity is not semantic entailment. Persistent records do not detect consciousness or prove a soul.

Results remain visible

Study Observed result Boundary
Scoped-memory presentation 384 scored local-model calls; compiled context did not generally beat an information-matched index Narrow representation comparison
Reading-mediated transfer Ordinary optimal plans 12/16, book 8/16; fully traced completion 0/16 in both Registered reading intervention did not improve the primary
Exact handoff repair Fully traced optimal completion 8/8 exact versus 0/8 freeform One known class; an accessible generic file restores the same data
Executable-method cycle Both arms reached the ceiling; M0 and M1 each 192/192 optimal transfer per arm; zero changed plans No beneficial revision or added-view advantage observed
Scoped evidence return 72 scored GPT calls plus one canary; free notes 77/80, linked notes 71/80 Same 2048-byte allowance; 6/8 versus 2/8 actual-feedback notes admitted
Scoped-contract adaptation 48 scored generations plus one canary; separate 960-execution reproduction: M0 120/160, every revision/fresh arm 160/160 Original interrupted, primary null; explicit new rule sufficed, no added scaffold or diagnostic-feedback score
Own-source induction One M0 trajectory; transfer Brier ordinary 0.25174, linked 0.25249, fresh 0.00869; lower is better Informative feedback, no linked gain; diagnostic leaves actual M0 alternatives unresolved, not proof of irrational anchoring
Same-information forecast calibration 9 model calls, 133 saved executions; all 96 revised distributions match unchanged M0(E1); D=0 Information conditioning sufficed; organized input +3.3551% across four calls per format; finite authored packet

The scoped evidence-return result is negative: over-budget notes were unavailable, and actual feedback changed source decisions while masked branches retained uncertainty. The separate contract-adaptation result is a ceiling: all revised and fresh programs solved the explicit new rule. Source changes and complete records do not establish beneficial formation or general ontology gain. Ukrainian reports: evidence return · contract adaptation.

Protocols, raw outputs, costs and offline reanalysis are available. Book v3.7.0 integrates the completed forecast calibration into both manuscript editions; the original study records retain their originating edition. Claim IDs include C03 unsupported, C06 refuted and C10 not-claimed; see the exact ledger and history.

The latest forecast-calibration readback separates information conditioning, source revision and added organization. The earlier induction readback retains its distinct unresolved diagnostic. Both selected archives are regraded offline without executing candidate code.

Reproduce and inspect

python3 -m unittest discover -s examples -q
python3 -m unittest discover -s ontology -p 'test_*.py' -q
python3 ontology/export.py --check
python3 -O research/epistemic_return/analyze.py

Book sources: book/ua/, book/en/; ordering: edition.json. Build both web editions with Node.js 22+: npm ci --ignore-scripts, npm run build, npm run check. PDF build and full research instructions.

Authorship and participation

Conceptual direction, research questions and editorial intent: DiadkoShmek. Codex substantially contributed writing, software, study design, execution, analysis and translation. Private Architect-Ψ council informed the work; it is not independent empirical replication. Authorship · Reader note · Changelog · Book citation · Note citation.

The proposed connected methodology has precedents in truth maintenance, AGM belief revision, provenance, event sourcing and reflective agents. World-first novelty is not claimed. Reproducible issues, objections and permitted failure cases are welcome through GitHub Issues.

© 2026 DiadkoShmek. All rights reserved. Public availability does not itself grant an open-source license.

About

Bilingual book and reproducible studies of agent memory, evidence, and decision revision.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages