Keyne Digital Assets

Docs / Knowledge Base / Concepts governance

Knowledge Base

Contents

Concepts governance

Part: Computer Science · Chapter: Concepts governance · Edition: 2026-10-06
Keywords: plan, certificate, conservativity, quarantine, OWL 2 RL, two clocks, proof-carrying retrieval, query-corpus blast radius
Glossary: blast radius, certificate (plan), conservativity, proof-carrying retrieval — see GLOSSARY

At a glance

  • A change to a rule table is a plan. Certify it. Reject it when it is forged or fails conservativity. Apply only the certified plan.
  • The published silent-policy example keeps blast_radius at 0 and risk_score at low while conservativity reads not_conservative_under_rule_table and the owl-rl rule table adds 901 new consequences.
  • On this desk, conservativity is a hard gate before insert. The ids that move are article ids and org file ids.
  • Grammar stays on a certified clock (Stanza, Universal Dependencies). Concept graphs, fill, map chat, and word–return stay on an opinion clock until they carry a rule id and a certificate id.

Online references (live)

Verified 2026-10-06. Each URL below returned a page; titles are the titles on those pages.

Offline snapshot omitted in this web edition. Offline snapshot · OWL 2 Web Ontology Language Profiles (Second Edition) · https://www.w3.org/TR/owl2-profiles/ · retrieved 2026-10-06

This chapter is the book’s account of a design. The design note in the Situation Monitor repository is docs/CONCEPTS_GOVERNANCE.md. It specifies gates. It does not add them to the writers. The loop it starts from is Open Ontologies.

1. Plan, certify, reject forged, apply

Definition (plan). A named proposal to change a rule table, or to change the graph that table produces, written down before any insert.
Example. One axiom, ex:hasParent rdfs:domain ex:Person, offered as the whole change.

Definition (certificate). A record that names the plan, the hash of the rule table that actually ran, and the closure diff that check computed.

Definition (apply). The write of a certified plan into the store a reader sees. A plan that fails certify stays in quarantine, a sibling document of the applied graph.

Definition (forged plan). A derivation whose conclusion the checker cannot obtain from the stated premises and the named rule.

Open Ontologies states the loop in four beats. Plan the change. Certify it under a named rule table. Reject a line the checker cannot derive. Apply only a plan that certificate covers.

Their README (retrieved 2026-10-06) gives the counts a run produced, and says nobody writes a count into the caption: 426 asserted, 259 certified, 1 rejected. A Lean 4 checker proved the certified edges. One certificate holds the 259. The same checker refused the forged line, exited 1, and named the rule. The stack line in that README is Oxigraph 0.5 for RDF and for SPARQL 1.1, and Lean 4 v4.33.1 for the checkers.

On this desk the same four beats are the writers that today invent and store in one step: category labels on an article, grammar frames, the news concept graph, the org concept graph, and the argument map. The plan is the document those writers produce before insert. The page and the graph deck read the applied graph. Quarantine is not a field inside the applied document and not a refused list that a later GET copies onto the map.

A changelog entry, a git commit of the map, or a cache write of the concept graph can sit beside a bare edge list. Those records are not a certificate.

2. The silent-policy bug

Definition (silent-policy bug). A plan whose shape report stays quiet while the conservativity verdict says the rule table has new consequences.
Notation. In the published transcript: blast_radius 0 triples affected, risk_score low, conservativity = not_conservative_under_rule_table, new consequences = 901, rule table = owl-rl.
Example. The whole change is one rdfs:domain axiom. Added classes 0. Removed classes 0.

The README prints that transcript and states the gap in one sentence: each number from a shape diff says the change is safe, and the change gave a new type to each individual the property already had. The new sentences are consequences under the rule table. OWL 2 RL, in the W3C Profiles recommendation, is the profile aimed at rule-based reasoning over OWL 2 and RDF graphs. RDF 1.1 is the data model those graphs use: a triple of subject, predicate, and object. A missing triple is not a negation. The quiet board is the set of triples the shape diff counted. The moving object is the set of sentences the rule table now licenses.

Nine hundred and one new consequences is the published count. It is the “hundreds” the desk design is written against. A risk score that remains low in that row does not describe the policy. The policy changed in the intension.

The same shape on this desk is a label or an implies edge whose concept board stays near a dozen keywords while a standing query gains articles. blast_radius computed as “no new nodes on this board” is that quiet board. A printed correlation magnitude that stays small while the row is still stored is that quiet board on the word–return desk. If the standing-query id set grew, the plan is rejected. The score does not get a vote, and the write does not happen.

3. Conservativity as a hard gate before insert

Definition (conservativity). With respect to a Horn rule table: the change alters no consequence over the names the store already uses. Open Ontologies states the verdict as a word, not_conservative_under_rule_table among them, rather than as a bare boolean.
Definition (hard gate). Compute the id set a standing query would return if this label or this edge were applied. When that set is a proper superset of the set under the current store, certify fails and the insert does not happen.

The lifecycle document makes conservativity the part of a plan that is about meaning. The rest of the plan is shape: classes, properties, individuals, a triple delta, a blast radius, a risk score. In that document the check is opt-in, a check that cannot run leaves conservativity.skipped, and the plan is still returned. The Situation Monitor design tightens the verdict into a gate.

The gate covers two writes, and it runs before either write:

  1. The category insert, before categories is stored on the article.
  2. The implies join, before those edges exist in the object a reader receives or an org graph stores.

A filter that drops one label after the store already counts it sees a membership that is already true. That filter is a later pass. The gate is the check that runs first.

4. Two clocks

Definition (certified clock). The grammar store: Stanza tokens (tokenize, pos, lemma, depparse), the grammar_hash, article frames, grammar occurrences, and the /ieml reads of classes, occurrences, and frames.{role}.
Definition (opinion clock). The concept graph, org fill, map chat, and the word–return rows.

Stanza’s own site describes a pipeline that tokenizes, tags, and dependency-parses in the Universal Dependencies formalism. Universal Dependencies is the cross-language annotation of parts of speech and syntactic dependencies. Those roles are the certified grammar. An implies edge, a Pearson row, or a fill turn stays on the opinion clock.

Opinion becomes a frame role, or a map edge, only after plan, certificate, and apply on that target. The two clocks do not share a write.

ClockWhat it holdsWhat reaches a reader
CertifiedStanza tokens, Universal Dependencies roles, grammar_hash, frames, occurrences/ieml classes, occurrences, frames.{role}
OpinionConcept graphs, org fill, map chat, word–returnThe applied graph, after a certificate on that target

5. Proof-carrying retrieval

Definition (proof-carrying retrieval). Each applied edge carries a rule id and a certificate id on the payload the client and the agent receive.
Example. Open Ontologies writes a certificate. The command in the README is oo-cert asserted.tsv derivations.tsv. Exit 0 is covered by the theorem OOCert.certificate_sound. A false conclusion in the derivation file exits 1 and names the rule that does not hold.

SPARQL 1.1 is the query language over the RDF graph. Oxigraph is the SPARQL graph database the Open Ontologies README names for RDF and for SPARQL 1.1 (Oxigraph 0.5). A query that returns an edge returns the two ids with it.

On this desk the slim copy of a concept graph has to keep both ids. A field that exists on the rich graph and is dropped when the edge dict is rebuilt is absent at persist and absent again on read. A certificate that lives only in a cache or a changelog, while the response is a bare edge (id, from, to, relation, label), does not meet the definition.

The agent-facing tools that wrap public briefs stay wrappers until an edge that reaches them has a certificate id on that payload. There is no side channel of uncertified implies.

6. Query-corpus blast radius

Definition (query-corpus blast radius). The growth, measured before apply, of the article ids and the org file ids a standing query would return.
Notation. Diff the id sets. A node count on the concept board is a different measurement.

The desk that moves is the article query and the org-file query. The desk that looks small is the concept graph. Measure the first.

Query
Geopolitics membership, including the window that drops the category filter when the filtered set is small
Id to diff
Article id

Category briefs and first-label grouping

Id to diff
Article id

Org window, the approved files before the graph is built

Id to diff
Org file id

Grammar desk: occurrences and frames.{role}

Id to diff
Occurrence id, and the article URL on the frame

Word–return rows that enter the board

Id to diff
The correlation row. Count this set on its own. It is not an article-id query.

A display cut that keeps a few dozen nodes is the quiet board from §2. A plan can rewrite a query inside that cut, and a plan with a dozen edges can too. The diff that certify reads is the id diff above.

7. What a certificate does not decide

Definition (policy review). A separate act from the certificate. The certificate shows entailment under the rule table that was hashed onto it. It does not show that the axiom is a good category, a good implies, or a good correlation.

Pin the rule-table hash on the certificate. A proof checked against one table and stored against another is a forged certificate. The hash covers the rules that actually ran. When those rules change, old certificates do not apply.

References

Pages fetched 2026-10-06. A URL that did not return a confirmable page is omitted; none of the candidates below failed that check.

  1. Open Ontologies (repository). Plan, apply, and roll back changes to a production ontology, with a blast radius report and a proof an auditor can re-check. https://github.com/fabio-rovai/open-ontologies
  2. Open Ontologies, README. The silent-policy transcript (blast_radius 0, risk_score low, not_conservative_under_rule_table, 901 new consequences, rule table owl-rl), the Lean 4 checker, and the stack line “Oxigraph 0.5 for RDF and for SPARQL 1.1” / “Lean 4 v4.33.1”. https://github.com/fabio-rovai/open-ontologies/blob/main/README.md
  3. Ontology Lifecycle. Plan, enforce, apply, monitor, drift. Conservativity is the part of a plan that is about meaning; in this document the check is opt-in and a skipped check still returns the plan. https://github.com/fabio-rovai/open-ontologies/blob/main/docs/lifecycle.md
  4. Motik, Cuenca Grau, Horrocks, Wu, Fokoue, Lutz (eds.). OWL 2 Web Ontology Language Profiles (Second Edition). W3C Recommendation, 11 December 2012. Section 4, OWL 2 RL, including reasoning in OWL 2 RL and RDF graphs using rules. https://www.w3.org/TR/owl2-profiles/
  5. Cyganiak, Wood, Lanthaler (eds.). RDF 1.1 Concepts and Abstract Syntax. W3C Recommendation, 25 February 2014. https://www.w3.org/TR/rdf11-concepts/
  6. Harris, Seaborne (eds.). SPARQL 1.1 Query Language. W3C Recommendation, 21 March 2013. https://www.w3.org/TR/sparql11-query/
  7. Oxigraph. A graph database and an RDF and SPARQL toolkit. README on the default branch. https://github.com/oxigraph/oxigraph
  8. Lean 4 programming language and theorem prover. Repository README: “This is the repository for Lean 4.” https://github.com/leanprover/lean4
  9. Avigad, de Moura, Kong, Ullrich, with the Lean community. Theorem Proving in Lean 4. https://lean-lang.org/theorem_proving_in_lean4/
  10. Stanza – A Python NLP Package for Many Human Languages. Pipeline for tokenization, part of speech, lemmatization, and dependency parsing under Universal Dependencies. https://stanfordnlp.github.io/stanza/
  11. Qi, Peng; Zhang, Yuhao; Zhang, Yuhui; Bolton, Jason; Manning, Christopher D. Stanza: A Python Natural Language Processing Toolkit for Many Human Languages. 2020. The Stanza site links this abstract. https://arxiv.org/abs/2003.07082
  12. Universal Dependencies. Framework for consistent annotation of parts of speech, morphological features, and syntactic dependencies across languages. https://universaldependencies.org/

Sources already cited in other chapters stay in those chapters (each chapter card’s Online references, plus the Full references / References sections in the long archive). This list does not replace them.

See also

Glossary terms introduced

blast radius, certificate (plan), conservativity, proof-carrying retrieval

Back to top ↑

All entries Glossary Foundations