Loading TACITUS
Loading TACITUS
METHOD · DEFINITION
The inference pattern of the knowledge layer. Typed. Grounded. Contested. Temporally honest.
OAG is the generation contract that runs over the typed graph TACITUS builds for policy and political files. It is independent of the specialization — the same inference pattern handles a regulatory consultation, a peace process, an HR mediation, or a policy options memo, because the kernel underneath is the same.
DEFINITION
Ontology-Augmented Generation (OAG) is a generation pattern in which a language model produces output that is (i) typed — every produced object is an instance of a kernel primitive or a validated extension; (ii) grounded — every claim cites the source span it was extracted from; (iii) contested — counter-claims and contradictions are first-class objects in the graph the LLM consults, not noise to be averaged away; (iv) temporally honest — every claim carries valid-time and transaction-time stamps the LLM is required to respect.
HOW OAG DIFFERS FROM ITS NEIGHBORS
| Pattern | Schema | Provenance | Contestation | Temporality |
|---|---|---|---|---|
| RAGLewis et al., 2020 | None | Document-level | Averaged | Flattened |
| GraphRAGEdge et al., 2024 (arXiv:2404.16130) | Auto-induced, flat | Entity-level | Averaged | Mostly flattened |
| OG-RAGSharma et al., EMNLP 2025 | Domain ontology, fixed | Entity-level | Not modelled | Not modelled |
| Think-on-GraphSun et al., ICLR 2024 | Domain KG, fixed | Path-level | Not modelled | Not modelled |
| Agentic deep graph reasoningBuehler, 2025 (arXiv:2502.13025) | Schema-free, self-organising | Path-level | Not modelled | Not modelled |
| OAG (TACITUS)TACITUS Research, 2026 | Kernel + dynamic extensions | Span-level | First-class objects | Bi-temporal |
RAG
Lewis et al., 2020
GraphRAG
Edge et al., 2024 (arXiv:2404.16130)
OG-RAG
Sharma et al., EMNLP 2025
Think-on-Graph
Sun et al., ICLR 2024
Agentic deep graph reasoning
Buehler, 2025 (arXiv:2502.13025)
OAG (TACITUS)
TACITUS Research, 2026
Each row is the right architecture for a different question. OAG is optimal for typed institutional analysis with contested provenance and bi-temporal honesty — the question form policy desks, mediation teams, regulatory bodies, and political-affairs offices ask by default [, , , , ].
PROPERTIES
Typed
Every produced object is an instance of a kernel primitive or a validated dynamic extension. Untyped output is an error, not a fallback.
Grounded
Every claim cites the source span it was extracted from. No claim, no span, no claim.
Contested
Counter-claims, contradictions, and disputes are first-class objects in the graph the LLM consults. The model is not asked to resolve them; it is asked to surface them.
Temporally honest
Every claim carries valid-time and transaction-time stamps. The model is required to respect both.
WORKED EXAMPLE
Sam asserts a Commitment on Monday. Alex acknowledges ambiguously the same morning. On Thursday, Alex denies scope. OAG types each speech act, stamps the bi-temporal intervals, and creates a contestation edge — instead of averaging the three turns.
{
"primitives": [
{ "id": "Sam", "type": "actor", "provenance": "msg1" },
{ "id": "Alex", "type": "actor", "provenance": "msg2" },
{ "id": "cm1", "type": "commitment",
"subject": "Q4 launch deck content", "deadline": "Thursday",
"valid_time": "2026-04-13/Thursday",
"transaction_time": "2026-04-13",
"status": "contested",
"provenance": ["msg1","msg2","msg3"] }
],
"edges": [
{ "from": "Sam", "to": "cm1", "type": "ASSERTED",
"provenance": "msg1" },
{ "from": "Alex", "to": "cm1", "type": "ACKNOWLEDGED_AMBIGUOUSLY",
"provenance": "msg2" },
{ "from": "Alex", "to": "cm1", "type": "DENIES_SCOPE",
"transaction_time": "2026-04-16",
"provenance": "msg3" }
]
}GENERATION PROCEDURE
OAG runs as a four-stage pipeline. Click a stage to expand its contract. Each stage either commits a typed graph operation or quarantines the operation for analyst review — there is no “best effort” pass through this pipeline [].
Stage 1 · Retrieval
Span-aware retrieval over the corpus returns ranked passages with character-level spans. Retrieval is typed: the retriever knows which kernel primitive the downstream extraction is targeting and prioritises spans containing canonical surface forms (commitment-language verbs for Commitment extraction, framing-language for Narrative extraction).
VALIDATION PIPELINE
Schema conformance
Required fields present and well-typed against the kernel schema or the extension subclass declared for the case. A missing field is a rejection at this stage, not a downstream surprise.
Kernel-invariant validation
Every parent-primitive invariant must hold. A Commitment cannot transition Active → Fulfilled without intervening evidence. A Leverage holder cannot equal its target. The TCGC kernel-invariant-validation task measures how often LLM-driven extraction violates these.
Extension-invariant validation
Subclass-specific invariants. HR-Commitment.review_period > 0; CeasefireCommitment.geographic_scope references a valid geographic primitive; RegulatoryAct.consultation_window is a closed interval. The extension log records every violation.
Provenance binding
Every field not derived from a kernel field binds to at least one source span. An emitted Claim without a character-level span pointer is rejected. Cryptographic-style binding between Base and Evidence is in progress.
COUNTERFACTUAL
Question: “What did the parties commit to regarding the Q4 deck content?” Inputs: the three messages from tcgc-0001. Below, the same question rendered by an RAG baseline (smoothed paraphrase) and by OAG (typed JSON with span pointers and a contestation edge).
No span pointers. Contradiction smoothed. Bi-temporal information lost.
{
"primitives": [
{ "id": "Sam", "type": "actor", "provenance": "msg1" },
{ "id": "Alex", "type": "actor", "provenance": "msg2" },
{ "id": "cm1", "type": "commitment",
"subject": "Q4 launch deck content", "deadline": "Thursday",
"valid_time": "2026-04-13/Thursday",
"transaction_time": "2026-04-13",
"status": "contested",
"provenance": ["msg1","msg2","msg3"] }
],
"edges": [
{ "from": "Sam", "to": "cm1", "type": "ASSERTED",
"provenance": "msg1" },
{ "from": "Alex", "to": "cm1", "type": "ACKNOWLEDGED_AMBIGUOUSLY",
"provenance": "msg2" },
{ "from": "Alex", "to": "cm1", "type": "DENIES_SCOPE",
"transaction_time": "2026-04-16",
"provenance": "msg3" }
]
}The OAG output preserves the three speech acts as separate typed objects, binds them to message-id spans, marks the commitment as contested rather than fulfilled, and exposes the contradiction as a first-class edge. A reviewer who only saw the RAG paragraph cannot get back to that structure.