Skip to content

Latest commit

 

History

209 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

og-caie-spec

Ontology-Grounded Contextual AI Evaluation (OG-CAIE) as an executable specification. Site: https://dynamicalsystemsgroup.github.io/og-caie-spec/

The process Humane Intelligence and Dynamical Systems Group use to evaluate a deployed AI system against the needs of a specific domain, written down as:

  • a glossary whose every term cites one canonical definition (ISO 9000:2026, SEVOCAB, NIST AI 700-2, W3C), with exactly four coined terms;
  • a SysML v2 model (OpenSysML v0.4.3) of the contracting lifecycle with the Evaluation Process Ontology nested inside it, and of the human and machine assemblage that runs it: structure only, rendered to RDF as the canonical graph; thirteen essentials, each tagged machine-verified or human-validated, live beside it in model/trace.ttl;
  • one evaluation record (PROV-O + EARL), the measles evaluation of the chatbot example, a bundle checked by SHACL shapes, with counterexamples that must fail, each the record with one change;
  • a rulings register holding every interpretive choice verbatim.

Status (2026-09-07)

  • Glossary: ratified. 69 terms, 4 coined (DSO, EPO, CAIE, OG-CAIE), settled through 51 rulings (R-01 to R-51) recorded verbatim in rulings/adjudications.ttl. 94 quotes are machine-located in content-hashed snapshots or, where the source is held locally, in its committed digest, 56 verified by Z against the ISO screenshots or the browsing platforms; none pending. Concerns C-25, C-30, C-43, C-45 and C-57 are open.
  • Both cycles are bound to the canon (R-31, R-32). Each step cites the process step it matches: SEBoK's account of the ISO/IEC/IEEE 15288 acquisition, stakeholder needs and system requirements processes; the ISO/IEC/IEEE 29119-2 test strategy and planning, execution and completion processes via SEVOCAB; ISO/IEC 17000's review, decision and attestation. The contracting lifecycle (need, propose, agree, access, deliver, accept) cites its canon step by step: SEBoK's business or mission analysis and account of proposals, ISO/IEC/IEEE 15288's agreement processes for the agreement itself, ISO 9000's contract, the 29119-2 test environment and completion report and SEVOCAB's acceptance (sheet 10-22); every item kind says which layer pins it. The sponsor's mission and obligations towards the affected populations open the record (R-37, R-38).
  • Site: a presentation layer over the model (R-34), seven pages and six appendices. The front page runs from why this counts as science to the bridge into the engineering standards; the vocabulary page shows exactly the terms the site uses, with hover definitions; four chapters (Stakeholders and contracting, Context and evaluation, A nested lifecycle, Records and reporting) each pair the specification with the measles walkthrough in five headed parts and close with the separation principle; the conclusion reads the crosswalk backwards. Figures are views from a registry (ogc/views.py), each captioned with what it brings into focus and leaves out (R-38). Appendix A is the sample report for the synthetic case (report/, R-51), a d3 dashboard rendered from the record that answers the sponsor's question and nothing else; Appendix B opens the same model as a knowledge graph explorer (explorer/, R-39), rendered from the RDF with the ogc command on every node and an in-browser SPARQL box; Appendix C's notebooks run the checks each chapter claims (notebooks/); Appendix D is the rulings, the judgments the model is grounded in (R-41); Appendix E reviews the toolchain, rendered from the lock file, the pinned digests, the gate and the workflow, with the reviewer recipe (R-43); Appendix F is the works cited, rendered from the source register as BibTeX, with each chapter closing on the sources it cites (R-48).
  • Model: revised (R-21 to R-38), draft. SysML holds structure only; the pruned RDF rendering model/og-caie.model.ttl is the canonical structure (R-22), checked by wiring shapes M1 to M5 over kinds of parts and ports: four parties, three actor categories within the testing organization, the six-step EPO nested as the fulfil step of the contracting lifecycle (one model, R-33), both process DAGs, 76 ports and 44 seams, every input wired once and every output somewhere, the sponsor's signatory as the person who signs, approves and accepts for it (R-50), and the sponsor's obligation to the affected populations as a relation. Four model counterexamples fail their shape. The essentials SCI-01 to SCI-13 live in model/trace.ttl.
  • The end-state demonstration (C-30) is built; one block awaits its rebuild. ogc/executor.py walks the process as the model states it and emits a record; the run conforms to S0 to S9, is complete against the model, has a recomputable coverage and traces fully, and twelve mutations are each caught by a named check (ogc execute, chapter Records and reporting). One finding is open as concern C-43: a record without a plan approval conforms to the shapes and is caught only by completeness. Z validated every block and wire on rulings sheet 05 but the report assembler (B8), rebuilt so that the conformance check on the record precedes the final report (sheet 10-41), and four wires new with R-50 (W15, W22, W36, W44); C-30 closes with their ticks.
  • Record and shapes: current (R-50, sheet 10). The measles evaluation (track/measles-evaluation.ttl) is a bundle every item is a member of, tagged synthetic; it names the parties and the sponsor's signatory, the two declarations, the agreement that accepts the proposal, the access with its period and instrument, the sponsor's approval of the requirement set, three requirements in the county's words with five criteria, four sessions of three turns, a paired determination, a cannot-tell resolved by a follow-up session and the attestation that supersedes it, a draft report with its gap flagged, the verdict on the record before the final report with the digests of what it ran and of the record it judged, a recommendation that states its fitness (R-51), and the acceptance; the step is derived through the model, never asserted; shapes S0 to S9 and forty-seven counterexamples in RDF, each the record with one change, follow the rulings; the traceback query reaches the parties.
  • Pending outside the repo. The w3id redirect (perma-id/w3id.org#6652) is open; IRIs resolve nowhere until it merges. The paper draft and the term contract will be aligned to this glossary once the glossary is closed.

Ask the graph

uv run -q ogc <command> navigates the vocabulary graph deterministically: ogc schema, ogc find, ogc term, ogc define, ogc quote, ogc verify (terms, sources and steps), ogc list, ogc source, ogc sources, ogc ruling, ogc rulings, ogc concern, ogc concerns, ogc sci, ogc steps, ogc epo (an EPO class, role or value: its label, superclasses, the layer it is pinned at, the term it names, the disjointness axioms and the shapes that mention it), ogc shapes, ogc shape (a node shape's target and constraints), ogc crosswalk (--popper for the seven Popper rows), ogc views, ogc view (a model view as mermaid with its perspective), ogc execute (the process executed from the model, with its checks and a VERDICT line; --planned, --sessions, --requirements, --criteria and --populations set the executor's parameters, capped at 100 criteria, 20 sessions and 100 criterion-sessions of work because the checks are quadratic), ogc record (the measles evaluation item by derived step, with who and when and the synthetic tag; ogc record <name> for one item), ogc check-word, ogc sparql (read-only, sorted, one merged graph; --model adds the model graph, --record the record, and a query that names the record or the model graph without its flag is refused), ogc doctor. Every output starts with # ogc <command> <args> @ <sha> (for sparql, the query as typed with its sha256, re-runnable); --json returns the result as one object carrying the same invocation and sha under _ogc, and every error as one object with _ogc, error, hint and candidates; ids may be typed as the tool prints them (a local name, a CURIE such as term:probe or rul:R-16, or a full IRI); exit 0 found, 1 not found, ambiguous, a bad filter value, refused or a failed verdict, 2 usage. The skill .claude/skills/ogc-glossary/SKILL.md teaches an LLM to use it instead of grepping the Turtle. It follows the navigation tool the authors built for their earlier glossaries (ruling R-29).

Authors and citation

A collaboration between Dynamical Systems Group and Humane Intelligence, co-authored by Michael Zargham (DSG) and Julie Hollek (HI). Cite it as such; CITATION.cff carries the machine-readable form (GitHub shows it under "Cite this repository").

Run the gate

uv sync && bash checks/run-checks.sh

The only verdict is the CHECKS: PASS|FAIL line; details in checks/out/.

IRIs

Ontology and record IRIs are authored under https://w3id.org/og-caie/. Until the w3id redirect is merged they resolve nowhere and are used purely as identifiers; the site above is the fallback.

Licence

CC BY-SA 4.0 for the whole work. The licence does not extend to the quoted source materials, which remain their publishers' property. This is an open standards activity under development with a computational implementation pathway. See LICENSE.md.

About

Ontology-Grounded Contextual AI Evaluation (OG-CAIE) as an executable specification

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages