Why TACITUS: keeping the context behind an analytical judgment

A briefing rarely starts from an empty page

A new analyst inherits a fictional file on a disputed infrastructure agreement. There are reports, draft clauses, meeting notes, and a previous options memo. One party’s public position changed after the memo was written. A deadline was extended, but the extension applies to only part of the agreement. The analyst can find the documents; understanding which assumptions remain usable is harder.

TACITUS began with that kind of problem. We want a person to reuse the useful work already done while seeing its sources, limits, and revisions. That is relevant to policy and diplomacy, but also to a business-risk assessment, a regulatory review, or research involving disputed accounts. The shared need is to explain what an assessment rests on and what would make it change.

The product and the research have different jobs

PRAXIS is the agentic workspace we are developing around Context Capsules. Its intended workflow lets a person assemble relevant material, inspect proposed records, use reviewed context in a conversation or bounded task, and examine the returned artifact. Updating the accepted context should remain a separate review decision.

A Context Capsule is the evolving knowledge format for that work. It is intended to carry source references, claims, relationships, time information, selected methods, and review history. DIALECTICA is the knowledge and review engine behind that direction. Its role is to organize proposals and produce inspectable context through defined interfaces. The complete integrated workflow remains under development.

An agentic task should have a bounded objective and authority. “Compare these accounts” is different from “update the institutional record” or “send the assessment.” NIST’s AI Agent Standards Initiative identifies interoperability, agent identity, authentication, and security evaluation as important technical work. Those are relevant requirements for our design, not capabilities conferred by describing PRAXIS as agentic. [1]

Structure has to earn its cost

We do not assume that every task needs a graph. A question answered by one clearly relevant paragraph may be handled well by search and a careful reader. Adding extraction, entity resolution, and review can introduce errors and maintenance work. Larger context windows and better models can also improve performance; our design should be evaluated against those improvements rather than depending on them failing.

GraphRAG offers one concrete research precedent. Edge and colleagues build an entity graph and community summaries to answer global questions over document collections. The paper reports improvements over its conventional retrieval baseline on comprehensiveness and diversity for the tested sensemaking tasks. That is not proof that graphs improve every retrieval task, establish factual correctness, or validate TACITUS. It shows why the question type and evaluation method matter. [2]

Anthropic’s context-engineering guidance discusses selecting and maintaining the information available to an agent across a growing interaction. We take it as implementation guidance from a model developer, not independent evidence that a particular capsule design works. It reinforces a practical concern: a larger store of material still needs a policy for what reaches the model at the next step. [3]

The distinctions worth preserving

Our design gives special attention to an assertion versus its evidence, an event date versus a reporting date, and a model proposal versus a reviewer decision. PROV supplies established concepts for describing derivation, activities, and responsibility. The capsule’s task-specific semantics and permission rules remain our responsibility. Provenance can help a reader assess a record without establishing the record’s truth. [4]

For narrative analysis, this means keeping competing accounts attributable to their sources and preserving uncertainty about intent or coordination. Similar wording is a lead to investigate, not a sufficient finding of an influence operation. For statecraft, it means distinguishing an obligation, a public position, and a possible negotiating interest. Useful software should let a practitioner inspect those distinctions instead of hiding them behind a single score.

What would count as progress

A useful evaluation needs realistic tasks, strong baselines, and people who understand the work. NIST’s ARIA planning manual combines model testing, red teaming, and user testing. We would use that distinction to test both the machinery and the decisions people make with it. An accurate extraction score alone would not show that a team produced a better assessment. [5]

We would publish where structured context helped, where it failed, and its maintenance cost. This note claims no comparative performance result. TACITUS is an independent experimental project; the goal is a workspace that makes analysis easier to examine and continue. Demonstrating that benefit is the work ahead.

A repeat brief is the first adoption test

The initial buyer hypothesis is a research director or partner at a small geopolitical-risk or policy consultancy. A senior analyst leaves, an editor repeats the same institutional correction, or a source changes before a client deadline. PRAXIS is the immediate workspace entry; the proposed expert collection follows that job. No customers, funded pilot, pricing, savings or categorical moat are asserted.

The useful next test includes a second assignment and a second analyst, equivalent source access and an independent reviewer. If preparation and upkeep outweigh repeat savings, or caveats disappear in handover, the hypothesis is weakened. [6]