nucleation

ADR-061: An ARG is any claim about support; a claim about what a source says is a THEORY

14 min read

Properties2
StatusActive
Tagsrecord

Graph View

Context

ADR-032 built ARG for two jobs: a second, independent line of support for a claim, and an objection to an inference rather than to its conclusion, which the will exchange showed was the most common real objection. To keep capture from flooding, it filed an ARG “only when something must point at the inference”.

The record holds ten ARGs and nine undercuts. All of them are dated 2026-10-04 to 2026-10-10, the two reconstructions that demonstrated the schemes. Since then about 280 claims have been filed, and no ARG. The objections did not stop: 30 counter claims are dated on or before 2026-10-10 and 31 after it. What stopped was filing them against an inference. CLAIM-163 says in its own summary that it is “the kind ADR-032 found most often real: same structure, not the same mechanism”, and it is filed as objects_to against CLAIM-162. CLAIM-254 says it “does not say relational identity is false, only that Saussure, read correctly, does not support its transport”, which is an undercut by definition, and it is filed as objects_to too.

The cause is the filing rule. The person who needs to point at the inference is the objector, and at that moment there is no ARG to point at. Undercutting means first writing out someone else’s inference as an ARG, then the objection. objects_to on the conclusion is one file and passes the lint. Filed that way, an objection that succeeds reads as a defeat of the conclusion, when what it shows is that one route to the conclusion fails.

The classification. On 2026-10-26 three readers classified all 453 claims then filed (485 by 2026-10-28; the 32 since are unclassified), one third each, against one rubric. A claim was SUPPORT when its subject is whether something supports, establishes or shows something else in the record. It was SOURCE when its subject is what a named work says, shows or proves, settled by reading the text. It was MIXED when it bundled either with a claim about the subject matter. Readers marked their own confidence. The calls were spot-checked, but not every one was read twice, and the readers differed in strictness (the third marked no SUPPORT as high), so the low-confidence counts are soft.

KindHighLowTotal
SUPPORT202242
SOURCE51621
MIXED41317
Neither373

SUPPORT falls into four groups:

SOURCE held a surprise. Of the five high-confidence cases, four ground a THEORY that already reads the same source: CLAIM-066 and THEORY-155 on Kirby, Cornish and Smith; CLAIM-069 and THEORY-168; CLAIM-072 and THEORY-159 on Saussure; and CLAIM-169 and THEORY-067 on Ainslie. Three of the four were granted objections when classified; CLAIM-069 has since been ruled Proposed. CLAIM-169, the fourth, is a thesis the record uses as a premise, and on reading it stays a claim. CLAIM-072 is not a reading the record lacks. It says that CLAIM-082 leans on Saussure for something THEORY-159’s reading does not say. That is a verdict on support. CLAIM-097, on T2I-CompBench, was classified as reading a source with no THEORY behind it; THEORY-522 reads that benchmark, so it is an objecting claim grounded in a reading, and stays. Most of the low-confidence SOURCE cases are prior-art reports (CLAIM-001, CLAIM-019, CLAIM-024, CLAIM-088, CLAIM-146) or restate textbook results the record applies (CLAIM-094, CLAIM-131, CLAIM-132, CLAIM-148, CLAIM-302). CLAIM-028 reads the exchange’s own drafts, not literature, and stays a claim.

MIXED: CLAIM-035, CLAIM-140, CLAIM-234 and CLAIM-412 with high confidence; CLAIM-130, CLAIM-134, CLAIM-211, CLAIM-219, CLAIM-225, CLAIM-235, CLAIM-253, CLAIM-309, CLAIM-343, CLAIM-393, CLAIM-411, CLAIM-432 and CLAIM-433 with low. The remaining SOURCE cases at low confidence are CLAIM-013, CLAIM-018, CLAIM-159, CLAIM-325 and CLAIM-360.

Decision

The boundary between CLAIM and ARG is drawn by subject.

  • CLAIM holds what the record says about its subject matter.
  • ARG holds what the record says about support between its entries: that these premises do, or do not, establish this conclusion. Its status is the verdict, which is what ADR-032 already made it (“whether the INFERENCE holds, which is not whether its conclusion is true”), and what ADR-059 since stated exactly: an ARG’s status is credence in its inference, within the limits its granted undercutters set.

So a negative verdict is an ARG too. “X does not show Y” is filed as the ARG whose premises are X and whose conclusion is Y. It is Rejected when the record holds that the support fails, and Proposed while that is open. Its body answers the form’s critical questions, and the question that fails is the objection. CLAIM-163 becomes the ARG from CLAIM-169, THEORY-069 (phase synchrony among coupled oscillators, the analogy’s source) and THEORY-070 to CLAIM-162, form analogy, Proposed. Its body records that the critical question “is the mechanism the same?” is answered no. That takes one file, and no one has to write out the inference first, because stating the inference is the objection.

The rest follows from that:

  • The filing rule changes. An ARG is filed when the record says something about a support relation: it argues one, doubts one, concedes one fails, or needs one as a second line. A claim’s own rests_on still needs no ARG (ADR-032’s throttle stands). The new rule moves the decision to where the objection is written, which is where it was being lost.
  • undercuts narrows. It is kept for a claim about the subject matter that defeats an inference: the non-invertibility of translation operations in CLAIM-035, which defeats the relativity analogy of ARG-004. A claim whose content is only “that does not follow” is not an undercutting claim. It is the ARG’s status and body.
  • One ARG per inference. Doubts about one inference, and replies to them, are revisions of its ARG, recorded by version and history:. They are not a chain of claims about it. A doubt that needs a code of its own is a claim about the subject matter, and it undercuts.
  • A test design is a premise. “A win on this test would not support the thesis” is the ARG from the CASE that describes the design to the thesis, Rejected. CLAIM-203 already grounds CASE-040, so its premise exists.
  • Either-or support is unchanged. Two ARGs that conclude one claim are two independent lines (ARG-005 and ARG-009 for CLAIM-027).

A claim whose subject is what a source says is a THEORY, since THEORY is the record’s reading of the literature. Where the record already holds a THEORY reading that source, any new content is a revision of that THEORY. Where the claim objects that another claim misuses the reading, the objection is the ARG from the THEORY to that claim, Rejected. CLAIM-072 becomes a revision to THEORY-159, if it adds anything to the reading, and the ARG from THEORY-159 to CLAIM-082, Rejected. A claim that reads a source and finds the target false stays an objecting claim, grounded in the THEORY. Prior-art reports and restated textbook results stay claims. They are the record’s own use of known results, and a THEORY for the sheaf gluing property would be a reading nobody did.

This is a rule about a claim whose subject is the reading. It leaves ADR-058 as it is: a reading that is evidence for a claim the claim does not stand on is corroborated_by, not grounds, and that edge is between a claim about the subject matter and a THEORY, which this decision does not move.

Two argument forms are added, because the support claims need forms the closed list lacks:

  • evidence: an observation or result supports a hypothesis. Critical questions: would the evidence be expected if the hypothesis were false? Does it tell the hypothesis apart from its rivals? Does the measure measure what it is said to?
  • source: a reading of a work grounds a claim. Critical questions: does the work say this, read in context? Is it said within the scope the claim needs? Does the claim lean on the reading for more than the reading holds?

A MIXED claim is split. The part about the subject matter stays a claim. The part about support becomes an ARG, which the claim undercuts when that part defeats it.

Alternatives considered

  • Fold ARG into CLAIM (the owner’s first framing). An inference is a claim, and a bridge claim (“the analogy carries the mechanism”) gives an undercut a target. But rests_on is one list of premises that are all needed, so it cannot say that a claim rests on either A and B or D and E without a node per route, and a CLAIM in rests_on reads as one more premise. Its status would also mean something different from its neighbours’, which is ADR-032’s own test for a separate scheme.
  • A voice field on ARG (role, as on CLAIM). The status already says what the record holds about the one proposition an ARG states. disclaimed, an inference nobody draws, is the case that tempts it. ADR-032 found most “X does not establish Y” sentences were of this kind. Such an inference is filed only when the second-mention rule says so, and its body says nobody draws it. Add the field when a check needs it.
  • More CLAIM schemes, split by role or by warrant. ADR-032 already rejected a split by role: a counter’s status means what a thesis’s does. Of the 88 demonstrated theses, 83 are Active, and their defeated_if amounts to “a counterexample or an error in the proof”, so credence barely applies to them. But they have the same relations and required fields as other claims, and ADR-043 made warrant a field on purpose. A field still suffices.
  • Keep the boundary, and add a prompt to the counter template. That treats the filing cost as a habit. The cost is structural: an undercut needs two files where a rebuttal needs one. The prompt is still worth adding under this decision, as the question that sends an objection to the right scheme.
  • Status quo. About 60 claims would keep saying something about support in a scheme whose status and relations are about the subject matter, and an objection that succeeds would keep reading as a defeat of its target.

Consequences

  • A Rejected ARG draws no step in the argument chain (the argued_from pattern already excludes it), so a negative verdict never lengthens a line. It appears under its conclusion as argued_by and on the ARG index.
  • Undercuts join the dialectic page, through a pattern. The page walked only objects_to, so a claim that undercuts an ARG was in no thread: CLAIM-032, CLAIM-035, CLAIM-085, CLAIM-141 and CLAIM-144 were on no line of it. The undercuts_support_of pattern in luria.yaml, (u:CLAIM)-[:undercuts]->(g:ARG)-[:concludes]->(c:CLAIM), puts the undercutting claim in the thread of the claim the argument concludes, and names the argument it went through. It keeps Rejected arguments, since an undercut that succeeded is the history the thread is for. It was fired on 2026-10-27 by rendering the page: 46 threads became 48, and all nine undercuts appear under their conclusions, “through” their ARGs. This part does not wait on acceptance, and is in this contribution.
  • A bare verdict gets a page through luria’s pattern pages, not a multi-scheme chain. An ARG with no undercutting claim is on no dialectic thread, and that is most of what the migration files. The draft of 2026-10-26 waited for chains to list several schemes (dmarx/luria#380). Since then luria 1.5.0 renders a pattern page as an outline, nested like a chain, and lets patterns that share an output merge (dmarx/luria#383). A trial on this record on 2026-10-28, with a bare pattern (g:ARG)-[:concludes]->(c:CLAIM) and the undercut pattern sharing one page, listed all ten ARGs under their nine conclusions with each undercut nested beneath its argument. Adopting it is the pin bump from 1.3.0, in its own contribution, and #380 is not needed for it. Rejected on the way: a pattern nesting a premise under the conclusion through a non-Active ARG, which would show the premise’s voice and credence as the objection’s, and a separate view of contested arguments, which would split the thread across two pages.
  • ADR-032 takes a history note. Its ARG paragraph and the filing rule are changed by this decision. So is its consequence “Lines of argument stop at an ARG”, which the argued_from pattern in luria.yaml stopped being true of, though no decision recorded it.
  • The schemes need small changes. The claim template’s counter section asks the routing question: does this say the conclusion is false, or that this does not show it? The ARG template drops “only when something needs to point at an inference”. The undercuts blurb narrows, and the arg-forms vocabulary gains its two forms. A CLAIM’s superseded_by widens to ARG and THEORY, so a moved claim can be retired by status.
  • The ARG scheme gains uses and works. A claim moved to an ARG can be stated in a TERM and belong to a manuscript, and an inference is stated in the same terms and argued in the same work, so both are fields on ARG (a TERM’s used_by names claims and arguments). What a moved claim carried in complements, illustrated_by or defeated_if goes to the ARG’s body. objects_to cannot name an ARG, so a reply to a moved objection is a premise of the ARG’s next version, which “one ARG per inference” already says.
  • One ARG may conclude several theses. When the same premises bear on several theses in the same way, the inference is one and its verdict is filed once: CLAIM-232 and CLAIM-233 attack one test design whose failure bears alike on four theses, and their ARG concludes all four.
  • The migration was previewed before acceptance. The workbench entry of 2026-10-28, “What accepting ADR-061 would change”, reads each of the 65 classified claims and names the ARG each move files. On that reading 30 move (10 to ARGs outright, 18 split, 2 into a THEORY’s reading plus a Rejected ARG) and 35 stay, filing 28 new ARGs and superseding 12 claims. The corrections it found are made in this text.

Migration

Done on acceptance, 2026-10-28, in the same contribution, by reading each entry (dmarx/nucleation#347, merged into this decision’s branch). It filed 29 ARGs (15 Rejected, 13 Proposed, 1 Active), superseded 14 claims by them, split 15 more to their subject matter, revised ARG-003, ARG-006 and ARG-008, and filed two CASEs the moves needed as premises. The rules it followed are these.

  • Each SUPPORT claim becomes an ARG, and the claim is retired by status, Superseded, never deleted. A CLAIM’s superseded_by names only claims today, so it widens to name an ARG or a THEORY. Whether luria writes the converse across schemes is checked, and fired once, before the first move.
  • A claim that an Active claim rests on is split, not superseded, whether it was classified SUPPORT, SOURCE or MIXED. rests_on names only claims, and inverted_premise (an Active claim resting on one not held) fails the lint since ADR-057’s forcing functions. So a dependent cannot be repointed to the ARG, and superseding its premise would break it. The dependence is the sign that the claim also says something about the subject matter: that part stays a claim, revised, and the support verdict becomes the ARG, as for MIXED. On 2026-10-28 this covers CLAIM-085 (rested on by CLAIM-243, CLAIM-248 and CLAIM-252), CLAIM-233 (CLAIM-232), CLAIM-053 (CLAIM-229) and CLAIM-066 (CLAIM-234). CLAIM-011, which Active claims also rest on, has no support clause to split; its verdict is already ARG-006’s.
  • The rule is applied after the moves, not before. CLAIM-053 is forced to split only because CLAIM-229 rests on it, and CLAIM-229 itself becomes ARGs; so does CLAIM-234, which rests on CLAIM-066. A move that removes a claim’s last Active dependent frees that claim to move, and two moves that file the same inference file it once.
  • Each SOURCE claim is read against the THEORY it grounds, and split as the Decision says.
  • Each MIXED claim is split.
  • The claims about defeat conditions stay: the four named above, and on reading CLAIM-206, CLAIM-208, CLAIM-222, CLAIM-226 and CLAIM-244.
  • References to a moved claim are repointed in the moving contribution; the pull request’s docs-check repair spells their targets (ADR-054).

Backlinks