Questions & references
Why prepare the knowledge?
The benefits, the technical choices and the claims that still need testing.
Why prepare policy knowledge before asking AI?
A useful policy answer depends on who said something, which actor they meant, the source version and the conditions attached. A knowledge pipeline captures those relationships once, so each briefing can reuse and inspect them. The intended benefit is less repeated preparation and fewer lost distinctions. Each workflow still needs evaluation.
How can data structures improve saliency?
Saliency means keeping the evidence that changes the conclusion. A task-specific structure can retrieve the relevant actor, clause and exception together, instead of sending an undifferentiated document pile. Models can still miss or misuse that evidence. Lost in the Middle shows that evidence position matters in the tested long-context settings; it does not validate Tacitus.
Where can efficiency come from?
Reusable definitions and source records reduce repeated assembly. Deduplication can avoid spending context on the same evidence twice. Selection can keep unrelated material outside the task. Building and maintaining a graph also costs time and compute, so we measure preparation, retrieval, model usage and review together. Shorter context alone is not success.
What do identity clarity and deduplication change?
A named actor gets an explicit identity record, with aliases, dates and source support kept separately. Similar names do not automatically mean the same organization or authority. Repeated copies can point to one source record while retaining each publication reference. Different accounts, conflicting claims and original reporting stay separate. Repetition is not independent corroboration.
How can a graph preserve diplomatic language?
Keep the exact wording as an object linked to its speaker, source version, speech act, conditions and review state. Prepared to discuss, subject to instructions must remain distinguishable from agrees to establish. Our authored example shows how those records can guide a draft and expose a lost reservation. A graph does not prevent semantic averaging on its own; extraction, interpretation and generation need separate checks.
What do expert reasoning and expert ontologies mean here?
We capture an explicit analytical method: which distinctions to test, what evidence is needed, what would weaken a conclusion and who should review it. An ontology defines types and relations; a semantic layer applies a selection to a particular question. These Tacitus checklists are authored frameworks for expert review, not proof that an expert approved a case. They do not store hidden model reasoning.
What is deterministic, and what still requires judgment?
Fixed data, versions and rules can produce repeatable checks and context selection. They do not make a contested claim true or make model prose deterministic. Schema validation can check a declared structure; a reviewer still needs to assess source interpretation, relevance and political or legal implications.
What has Tacitus actually benchmarked?
We have reproducible checks for evidence retention, context size, source identity and rules. A frozen 24 KB compiler regression now keeps three exact quotations and an artifact that the old version omitted. An earlier 48-response model evaluation had mixed results. We have not established a general improvement in generation quality, inference cost or analyst time.
How should a pilot test the benefit?
Use the same source pool and held-out questions for task-only, retrieval and structured-context arms. Blind the outputs for domain review. Score attribution, actor identity, conditions, supported claims, useful uncertainty and drafting quality. Record input/output tokens, indexing and retrieval costs, latency, failures and reviewer time. Automated scores support that review; they do not replace it.
Which tools and standards support the work?
Tacitus uses source records, versioned JSON contracts, RDF vocabularies, SHACL validation, SQL storage and graph inspection. The public workbench demonstrates bounded local reuse with source and time qualifications. GraphRAG and TrustGraph inform graph retrieval design; they are separate projects, not installed integrations or endorsements. The MCP adapter exposes read-only tools for compatible clients.
Does the knowledge depend on one AI model?
The canonical records, source identities and explicit relationships are designed independently of a model provider. An adapter still has to render them for the chosen workflow, and models differ in how they use context. Model-agnostic data does not guarantee equal results across models. Each provider and task needs its own evaluation.
Can we buy this now? What does a pilot cost?
We are working with invited pilot customers and testers. Pilot scope and pricing are agreed around one task, the source collection, review requirements and delivery work. There is no published subscription price or self-serve checkout. Contact us to discuss fit; no payment or commitment is made through this website.
What is PRAXIS, and can I use it?
PRAXIS is the agentic workspace for context capsules and artifacts. Access is invitation-only and AI generation is currently paused; paid upgrades are unavailable. You can still explore this website’s shared tools. Full connections between knowledge pipelines, the workbench and hosted agent workflows remain in development.