A statement of intent is not yet a commitment

Three sentences, three different questions

“We intend to deliver in June.” “We offer delivery by 30 June if the specification is approved by Friday.” “Both parties accept the attached schedule.” These statements may be related, but they do not establish the same thing. The first reports an intention; the second proposes a conditional arrangement; the third may provide evidence of acceptance whose scope still depends on the attached text and the parties’ authority.

The useful distinction is not that claims are unimportant and commitments are facts. A claim can report a genuine obligation, and an apparent promise can be unauthorized, conditional, disputed, or superseded. The system should retain the statement and explain the interpretation, rather than replacing the original words with a confident label.

Represent the duty before judging performance

The UNIDROIT Principles distinguish a duty to achieve a specific result from a duty of best efforts, and address how the kind of duty is determined. Their rules are a substantive reference for international commercial contracts, not a universal test for political pledges or every legal system. They illustrate why a single fulfilled/unfulfilled flag can erase essential meaning. [1]

A task to “use best efforts to secure approval” and a promise to “obtain approval by Friday” require different assessments. The absence of approval alone does not settle whether the first duty was performed. Likewise, a delivery deadline may depend on another party providing materials. If that condition is missing from the extracted record, a precise-looking dashboard can become misleading.

Reuse policy models carefully

W3C ODRL represents permissions, prohibitions, duties, parties, assets, and constraints for content and service usage. It is a useful example of why rules need subjects, actions, and conditions. It is not an off-the-shelf ontology for diplomatic obligations, nor does expressing a rule in ODRL establish that it is legally binding. [2]

For a small procurement file, a reviewed spreadsheet with source links may meet the need. A database or graph becomes attractive when obligations depend on other obligations, amendments affect multiple records, or the same evidence must support several queries. Choose the representation according to those relationships and the team’s ability to maintain it.

An update needs evidence and a reason

Suppose the supplier reports completion on 28 June, but the customer records failed acceptance tests on 2 July. Preserve both observations. Do not overwrite “active” with “fulfilled” merely because one party said the work was done. A reviewer may conclude that delivery occurred while acceptance remains disputed, or that the relevant condition was different from the one assumed.

A proposed transition record should include the old assessment, new assessment, effective time, recording time, source references, and reviewer rationale. PROV-O’s entities, activities, and agents provide useful vocabulary for recording how an assessment was produced or revised. An append-only event table can carry this history; a graph database is not required. [3]

Ask a narrower question, return the relevant qualifications

“Which delivery duties were unresolved on 1 July?” should retrieve the relevant records, amendments, conditions, and evidence—not the entire correspondence by default. That may reduce repeated prompt context. It may also omit something essential if extraction missed a side letter. Retrieval therefore needs a route back to source documents and a way to flag records whose interpretation remains unresolved.

Evaluation should include intentions mistaken for promises, conditional duties, amended deadlines, contested acceptance, and missing authority. Measure mistaken status transitions, missed exceptions, source support, reviewer correction time, and total processing cost. A change in wording or an unfulfilled obligation is not by itself proof of bad faith. The proposed PRAXIS review workflow and DIALECTICA knowledge records are under development; this note reports no automated adjudication capability or measured accuracy.

A candidate commitment needs the other party’s record

At the first “we are agreed” message, record Sam’s assertion and a candidate commitment. Alex’s acceptance has not yet been supplied. The later acknowledgment and later denial require an attributed interpretation of scope, authority, conditions and deadline; neither a graph label nor a single confident speaker settles it. The repaired fictional conversation asks for explicit acceptance and remains separate from implementation evidence. [4]