AGON: extracting a claim without endorsing it

The smallest mistake can change the assessment

Consider a fictional municipal briefing: “The transport office said the bridge would reopen on Friday if the inspection passed. Residents disputed the office’s claim that the structure was safe.” A careless extraction can turn this into a scheduled reopening, an established safety finding, and a single disagreement between two uniform groups. The passage establishes none of those things. It reports a conditional statement and a challenge to a claim.

That is the problem AGON explores. The task is to propose useful records while preserving who said what, whether a statement was quoted or reported, and what conditions limit it. ODNI’s analytic standards distinguish underlying information from assumptions and judgments.[1] We borrow that discipline for a research design; the directive does not validate our software.

An extraction is a proposal for a reviewer

Our proposed record would retain the source document, a passage locator, the quoted words, the attributed speaker, and the extraction method. A model’s interpretation belongs in a separate field. The passage should remain recoverable when a reviewer disputes the interpretation. W3C PROV-O provides a useful vocabulary for derivation and responsibility, but describing provenance cannot establish that the original assertion is correct.[2]

The following records are illustrative schema sketches, not a supported API. “Residents” remains an unresolved group reference until the source supplies enough information to identify speakers. The condition on reopening must travel with the candidate commitment.

A narrow role in the proposed stack

AGON and KAIROS are archived research precursors. We are exploring how their conflict and temporal methods could feed a versioned DIALECTICA adapter. AGON would propose actors, claims, commitments, and supporting passages; KAIROS would propose dates and ordering constraints. DIALECTICA is the intended knowledge and review boundary. PRAXIS is the agentic analytical workspace in which people use reviewed capsules for questions, analysis, and drafts.

This separation matters operationally. A successful extraction job should not silently publish a capsule or turn a suggested actor match into canonical identity. A rejected or qualified record needs a review decision that downstream work can respect. This is our proposed integration contract, not a claim that the complete pipeline is deployed.

Where the method can fail

Alias resolution is particularly risky. A ministry, its spokesperson, and an anonymous official are related references, but they are not interchangeable authorities. Denials also need an object: denying responsibility for a closure differs from denying that a closure occurred. Negation, translation, quoted speech, and conditional promises can each change the record’s meaning.

Interests require another boundary. A demand may suggest several motivations. A proposed explanation should name its author and alternatives rather than become an unqualified property of the actor. In defensive narrative analysis, repeated language can justify closer examination; it cannot by itself establish coordination, intent, or manipulation.

What we would test before relying on it

An evaluation set should contain independently annotated passages with disagreements retained for adjudication. We would score speaker attribution, passage accuracy, condition preservation, and incorrect actor merges separately. A single aggregate extraction score could hide a serious failure to distinguish allegations from findings.

Challenge cases should include a quoted denial, a revised promise, a translated statement, and two people with the same surname. Reviewers should also measure correction effort and whether a corrected record changes a later brief. NIST’s work on evaluation probes supports examining behavior throughout an agent workflow.[3] It does not supply results for AGON. Until such tests are run and reported, this remains a method to investigate.

A historical precursor, not endorsement or present integration

AGON is an archived conflict-structure reference. Its public source is all rights reserved; visibility is not permission to copy or redistribute. Extracting an actor or an attributed claim does not endorse the label, prove intent or settle a contested account. DIALECTICA v3 is the current local ingestion/review direction; a proposed adapter to the other engines remains unvalidated. [4]