# A deterministic knowledge layer for Policy / Political AI

Tacitus fixes the representation and checking of declared knowledge: concepts, typed statements, source versions, conditions, evidence requirements and output fields. The same records and rule version produce the same checks and canonical context. A deterministic representation does not make uncertain evidence true or political interpretation unambiguous. Model extraction, retrieval involving probabilistic models and generated prose require separate evaluation.

## What the TrustGraph research contributes

Official material was checked on 9 November 2026. TrustGraph's current homepage describes an ontology-grounded context graph plus a policy runtime that records allow/pause/deny decisions. Its retrieval documentation separately describes LLM-based concept extraction and answer synthesis, and ontology retrieval that selects a relevant subset instead of loading an entire ontology. Context Cores preserve reusable retrieval information. These are product/documentation claims, not independent performance evidence.

The useful architectural lessons are explicit meaning, versioned reusable knowledge, provenance, selection of a relevant subset, and checks separated from generation. Tacitus implements its own bounded policy-concept checks and evidence-first compiler. It does not install TrustGraph, claim API compatibility, reproduce its full agent runtime, or imply a partnership.

- TrustGraph product: https://trustgraph.ai/
- Ontologies/context graphs: https://trustgraph.ai/guides/key-concepts/ontologies-and-context-graphs/
- Retrieval: https://docs.trustgraph.ai/architecture/retrieval.html
- Context Cores: https://docs.trustgraph.ai/guides/context-cores/
- Source repository: https://github.com/trustgraph-ai/trustgraph

## Concepts and diplomatic language

The versioned concept records distinguish participation/consent, visibility/mandate, requests/undertakings, signature/ratification, adoption/application, allegations/facts and logistics stages. The statement schema records speaker, forum, date, addressee, exact source, declared speech act, modality, conditions and review state. It does not infer legal force from a word such as shall, should or may. Context and authoritative interpretation remain necessary.

Application checks ask whether a claimed interpretation has the declared evidence field. The result is evidence recorded or evidence needed, with repeatable input/rule hashes. This is metadata validation, not independent verification of the source or permission for an agent to act. These are Tacitus application definitions informed by the linked references, not newly discovered universal meanings.

Relevant standards: W3C SHACL validates declared graph conditions; PROV-O records provenance; SKOS represents controlled concepts; OASIS LegalRuleML represents legal-rule structures and context-dependent modality. The existing 16-paper register supplies additional research on retrieval, graphs, context and evaluation.

- SHACL: https://www.w3.org/TR/shacl/
- PROV-O: https://www.w3.org/TR/prov-o/
- SKOS: https://www.w3.org/TR/skos-reference/
- LegalRuleML: https://docs.oasis-open.org/legalruleml/legalruleml-core-spec/v1.0/legalruleml-core-spec-v1.0.html
- UN mediation guidance: https://peacemaker.un.org/en/documents/united-nations-guidance-effective-mediation
- UN Treaty Collection glossary: https://treaties.un.org/Pages/Overview.aspx?path=overview/glossary/page1_en.xml

## Saliency and efficiency

Saliency means preserving the evidence, conditions and distinctions that change the answer. Efficiency means reducing unnecessary context and review work while retaining those essentials. The compiler now retains required quotations/paraphrases and artifact content before supplemental text, shares duplicate source/evidence records and excludes unreviewed candidate paragraphs by default. It fails when required content cannot fit.

Compiler measurements cover bytes, named tokenizer estimates, exact evidence retention and repeatable hashes. They do not establish general improvements in model quality, provider cost or analyst time. The previous 48-response evaluation remains visible as mixed evidence; a new compiler needs its own held-out, matched-evidence model and practitioner evaluation before broader claims.

## Start with a real job

The database-backed start guide maps six recurring pressures to specific inputs, structures, capsules and artifacts. The assessment example uses clearly illustrative municipal-system inputs and real original adopted-law passages. Its note, crosswalk, missing inputs and dependency changes are generated by fixed rules without a model call. This demonstrates a reproducible structure; it is not an actual client result, a legal applicability decision or deployment approval.