nucleation

ADR-032: Argument development: QUESTION, CLAIM, CASE, TERM and ARG

9 min read

Properties3
StatusActive
Version4
Tagsrecord

Graph View

Context

The record documents other people’s arguments well: LIT for works, NOTE for readings, THEORY for accounts with their evidence, and extends, rivals, corrects and presupposes between them. The owner wanted it to hold their own argument as it develops: synthesis, objections anticipated and answered, positions taken and given up.

The design was worked out in a session on 2026-10-06 and 2026-10-07. It began from established structures: IBIS (Kunz and Rittel, LIT-758), discourse graphs, Toulmin, Dung’s abstract argumentation and the AIF, ASPIC+‘s three kinds of attack, Walton’s argumentation schemes with critical questions, Carneades and IRAC from legal argument, and competitive debate. The design-rationale literature on why such systems get abandoned (LIT-759, LIT-761, LIT-756) entered the record alongside it (ADR-030).

The decisive evidence was a real exchange: the owner’s three-day development of an essay with a chat assistant, 95 owner turns and some 300 replies, read in full by seven parallel readers, which coded every argumentative move. What it showed:

  • What went wrong was identity, not missing structure. Concepts were redefined silently: will five or more times, concordance five times, the perspective and level schemes swapped turn to turn. A worked case changed identity six times. The permissiveness objection was re-derived six or more times as though new. Readings the record already held were recommended three times. Commitments were dropped and never retired. Each of these is what a code fixes.
  • Inference attacks are common, and mostly on analogies. Roughly half to two-thirds of objections during development attacked an inference rather than a premise or conclusion, and nearly all the real ones were “same structure, not the same mechanism”. In the final manuscript, 7 of about 30 claims had more than one independent line of support.
  • Most “X does not establish Y” sentences deny inferences nobody makes. There were more than 40 in the audit stretch alone. They are claims the record disowns, not attacks on arguments.
  • Credence and inclusion come apart. The exchange’s own audit rated the paper’s central thesis “promising, needs direct defense” and made it the main contribution.
  • A line of argument had to be retracted whole. The owner tried to fork the conversation upstream of where authenticity entered; the fork did not hold, and the retraction was done in-band, by audit.

Decision

Five schemes, each with merge allocation and an index view.

  • QUESTION. An open question. Statuses Open, Answered, Dissolved, Deferred and Superseded. refines another question. Claims and THEORY readings name it in answers, so its page holds both: the literature’s accounts beside the record’s own.

  • CLAIM. The record’s own proposition. status is credence, with the same words and meanings as on a THEORY. role is its voice: thesis, counter (a steelman or anticipated objection), granted (accepted, and against a thesis), disclaimed (explicitly not claimed) or supposed. A thesis must say defeated_if, and every claim names its line. Relations:

    • rests_on another CLAIM: the record’s own derivation, and the spine a chain walks.
    • grounds a THEORY, CASE or LIT: evidence from outside the record’s own argument. A LIT target means the work is cited without a reading, which is the weakest footing. The split from rests_on is THEORY’s own, between presupposes and source.
    • objects_to a CLAIM (a rebuttal, or an undermining if it is a premise), and undercuts an ARG. The field says the kind of attack, so there is no strikes: field.
    • superseded_by a CLAIM, declared with a converse supersedes.
    • complements another CLAIM.
    • uses a TERM.
    • answers a QUESTION.

    An objection is a claim, so replies and counter-replies come free, and an answered objection is a counter gone Rejected.

  • CASE. An example, thought experiment or historical episode, filed once so every claim it serves points at the same thing. standing is documented, stipulated or fictional. Relations: illustrates a claim (bears no weight), counterexample_to a claim, variant_of another case, and substitutes_for another case (a fictional company standing in for a real one).

  • TERM. A concept or distinction. No credence: it has a definition, and a changed definition is a new version or a superseding TERM, so the claims stated in it (uses) can be found.

  • ARG. An inference: premises in rests_on (claims) and grounds (readings, cases, works), at least one; concludes a claim; and form from a closed vocabulary whose blurbs carry Walton’s critical questions. Filed only when something must point at the inference — a claim’s second independent line of support, or an undercut. Premises of one ARG are linked; two ARGs concluding one claim are independent lines.

Alongside the schemes:

  • lines. A named line of inquiry, so a line can be set aside as a whole and its dependents found through rests_on. It is the axis of the CLAIM and QUESTION indexes, which therefore group by line rather than by topic; the topic pages remain.
  • works. The manuscripts a claim is marshalled into. Membership is independent of credence.
  • The workbench journal. One dated entry per working session, saying what moved and why. The claims hold where the argument stands; the workbench holds how it got there.

Chains render what no single entry holds:

  • argument walks rests_on, with line as its invariant. A step from one line into another is reported, because that is how a line set aside reaches back into one still standing: had the authenticity claims been resting under will-organization claims, the chain would have said so.
  • dialectic walks objects_to, faceted by role and status. Each objection appears with its replies, and an open objection is visibly a counter still Proposed.
  • revisions and term-revisions walk supersedes, oldest first.

The multi-scheme references this needs arrived in luria 0.36.0 (dmarx/luria#357), and CI is bumped to it. The far side of each relation is a field of the same name on every scheme it names. THEORY gains answers and supports; LIT gains supports.

Demonstration. The exchange is filed as it would have been documented, from five workbench entries, one per session. Some of what it shows:

  • the paper’s thesis CLAIM-027, Proposed but in the work;
  • its two independent arguments ARG-009 and ARG-005;
  • the permissiveness objection as one code, CLAIM-010, with its history;
  • the definition of will superseded (TERM-003 to TERM-023);
  • Northstar’s variants as separate cases;
  • the Fraenkel analogy undercut in all but its structural respect;
  • the authenticity line deferred as a whole;
  • open items as Proposed counters and unanswered critical questions, including two the final rewrite dropped.

Alternatives considered

  • An objection scheme separate from CLAIM. luria’s own rule splits schemes when two kinds of claim must be able to disagree. An objection’s status means what any claim’s does — does the record believe it — so it is one scheme with a role.
  • An ARG for every claim (AIF’s information and inference nodes throughout). Principled, and it is what the design-rationale literature says gets abandoned: the capture cost lands on the author now and the benefit on a later reader. In the final manuscript most claims had a single linked line, so an ARG would be a second file saying what rests_on already says.
  • Claims only, with no ARG. That leaves an undercut nothing to point at, and the exchange’s most frequent real objection is an undercut of an analogy.
  • A strikes: premise | inference | conclusion field on objections. With ARG available, the kind of attack follows from the kind of thing attacked, and a field would be a second copy of it.
  • Multi-scheme rests_on and objects_to — one field holding a claim’s premises whatever their scheme. This was the first draft. A chain walks a relation only when the chain’s own scheme is its sole target, so the argument’s spine and its objection threads could not be drawn as lines. Splitting by what the relation means (derivation versus evidence, rebuttal versus undercut) gives single-scheme spines without splitting one relation by target type.
  • Lines as topic tags. The topics vocabulary is closed and says what a work is about. A line of inquiry is a different axis: it is retired as a whole, and it groups the record’s own argument rather than its reading. A line field, named as the axis, groups without diluting the topics.
  • A chain over refines for question trees. Nothing uses refines yet. Second use.
  • A full argumentation ontology (AIF, Carneades). These are built to analyse finished arguments. This record develops unfinished ones.
  • Literature roles beyond premise (foil, method, analogue, boundary marker), and authorship (owner or assistant). Both recurred. Prose carries them until a check needs them.

Consequences

  • The filing rule is the record’s own: a thing gets a code the second time it turns up. The drift the exchange shows is exactly the set of things that recurred. When an agent does the filing, the risk is no longer that capture stops but that it floods, and this rule is the throttle.
  • The record now holds documents not in force on purpose. A Proposed counter is an open objection; a Rejected one is answered history. Each citation of one carries an inactive-ok saying so.
  • complements stays between claims for now. A symmetric relation that crosses schemes (CLAIM to THEORY) reads as one-sided from both ends under luria 0.36.0, and luria link --fix appends a duplicate edge on each run. That is a luria bug, fixed in dmarx/luria#358. Widen complements to THEORY once a release carries the fix. Until then a reconciliation with a reading is a rests_on plus prose.
  • The cross-line check was fired before being trusted (2026-10-07). A will-organization claim was made to rest on the authenticity claim, and the lint reported “share no line (argument; each holds: authenticity / will-organization)”. It is a report, not a failure: a cross-line step can be legitimate, and a person decides.
  • Lines of argument stop at an ARG. A claim concluded by an argument rests on its premises through the ARG, and a chain does not cross schemes, so the main result appears in no line. Giving the claim a rests_on as well would be a second copy of the ARG’s premises. A luria chain that can pass through an intermediate scheme would close this.
  • luria labels a second parent “also extends” whatever relation a chain walks. On the dialectic page, an objection to two claims reads “also extends”. Cosmetic; fixed upstream in dmarx/luria#359, which names the relation instead (“also objects to”).
  • Not yet built: a report for a thesis with no counter against it, a check that a granted claim is Active, and a rendered argument map. Each waits for a second need.

Backlinks