Roles and relations
What may be said about a claim, and what may be said between two of them. This is the reference behind two corpus-scope layers — claim-format, which asks what a claim is, and relation-vocab, which asks what each relation asserts and which of them are oppositions. That second question is open, so everything downstream of it is provisional.
Counts are generated from the claim files by scripts/corpus_facts.py. The denominator is the
10 published papers; the two method-example papers are counted separately, which is why
a relation the examples use can still show zero here.
The roles
A role is the rhetorical function a claim serves in the paper's argument — what work it is doing. It is distinct from claim-type, which is the epistemic character of the proposition itself. The role governs what edges a claim can carry and how a reconstruction reads the argument from the graph alone.
Nine roles, and 9 of them appear in the corpus. The commonest is
empirical, at 125 of 289 claims.
hypothesis — 37 claims
The paper's organising bet: the proposition the rest of the paper defends or tests.
Signals: "we hypothesize", "we propose that", "we asked whether", "we sought to test", "the central question is whether".
Carries entails: to its predictions. It holds no empirical content itself and does not
requires: empirical claims — predictions do that. Panel is normally null; a hypothesis is
paper-level.
Headley.
hypothesis-distinct-compartmental-roles— "Perisomatic and distal dendritic inhibition serve distinct computational roles in regulating neuronal output." Entails four predictions.
prediction — 37 claims
The deductive consequence of a hypothesis under stated conditions: what we should observe if the hypothesis holds.
Signals: "if X, then we should observe Y", "this predicts that", "the model predicts", "is predicted to".
derived-from: its hypothesis, and is the target of tests: from the empirical claim that
closes the loop.
empirical — 125 claims
A measured or computed result, panel-grounded. The largest bucket by a wide margin.
Signals: "we found that", "we observed", "we measured", "Figure N shows", "doubling X reduced Y from A to B".
Carries the most edges of any role: tests: predictions, dissociates-with: complementary
results, requires: methodological and scope claims, supports: synthesis, rules-out:
alternatives. Panel is required — an empirical claim is panel-grounded by definition.
control — 26 claims
An empirical result whose primary work is to eliminate a rival explanation. Functional
rather than contentful: the same proposition could be empirical in another paper.
Signals: "rules out", "excludes", "shows the effect is not due to", "no significant effect of [confound]", "control condition".
The alternative needs a node. A control exists to kill a rival, so the rival must exist for
the rules-out: edge to reach it — authored as an ordinary claim, usually hypothesis, marked
as one the paper does not hold by the stance on its assertion rather than by its role or its
slug. A control whose edges name only the claim it defends, and never the alternative it
eliminates, has not recorded the point of the control. The
stance layer found five eliminative relations aiming at
claims their own paper asserts, and 9 still name prose that resolves to no
claim at all.
scope — 20 claims
A boundary condition on the empirical claims, often global: scopes: ["*"] qualifies every
empirical claim in the paper.
Signals: "all results come from", "results are restricted to", "the model assumes", "is not a physiological pattern".
Headley.
l5-model-single-cell-scope— "All results come from a single-cell compartmental model — no network dynamics, no recurrent excitation, no population effects."
methodological — 21 claims
A procedural or analytical capability that warrants a downstream interpretation. Not a claim about the world; a claim about the apparatus.
Signals: "spike-sorting uses Kilosort 2.5", "rhythmic inhibition is modeled as sinusoidal rate modulation", "the analysis uses [method]".
Carries enables-method: and scopes: to the empirical claims that depend on it.
synthesis — 8 claims
A higher-order proposition integrating several empirical claims into one statement, staying inside the paper's own evidence.
Signals: "taken together", "in summary", "these results show", "this dissociation establishes" — typically at section breaks.
derived-from: where the synthesis is a deductive consequence, supports: where it integrates
results. The distinction is load-bearing:
Meijer R1.
orthogonality-derived-from-additivity— "Under a linear readout, additive modulation entails orthogonality of the stim and choice axes." Carriesderived-from: hypothesis-additive-modulation, which demotes the empirical orthogonality finding from independent evidence to geometric corollary. An edge choice that changes what the paper has shown.
interpretation — 11 claims
A reframing of empirical results through a theoretical lens — an act of mapping to broader
theory, as against synthesis, which stays inside the paper.
Signals: "may provide a functional interpretation", "suggests a role for", "points to a mechanism whereby", "may explain".
Carries interprets: to the empirical claims being reframed.
literature-context — 8 claims
A cited prior result treated as a first-class node: the paper depends on it, but did not produce it.
Signals: "as shown by Author (Year)", "previous work has established", and the slug pattern
interprets-<author>-<year>-<topic>.
The top-level doi: is the cited paper's DOI — the schema's exception to the doi: ~
placeholder every other role carries. All 5 of the
8 literature-context claims in the corpus put it there, and
3 put it in assertions[0].doi. The
reference-check layer resolves them against
CrossRef, which is the anti-hallucination check: a cited DOI that resolves to a different paper
than the claim names is the failure a reader would never catch.
The relations
A relation is a proposition about logical structure between two claims — not a citation. A claim
file declares each as a top-level YAML key whose value is a list of target slugs. Reciprocals
(predicts / confirms) are populated symmetrically at build.
21 are defined and 17 appear in the corpus, carrying 916 relations in total.
| Relation | Reasoning form | What it asserts | Count |
|---|---|---|---|
requires |
dependency | A would be invalid if B were false | 104 |
supports |
abduction | A is evidence for B; several supports drive the abductive loop | 148 |
entails |
deduction | A, typically a hypothesis, deductively implies B | 54 |
derived-from |
deduction | A is the deductive consequence of B; reciprocal of entails |
48 |
tests |
deduction → empirical | Empirical A tests prediction B, closing the loop | 75 |
refutes |
negative abduction | A's evidence is incompatible with B | 1 |
rules-out |
elimination | A's evidence eliminates alternative B | 22 |
dissociates-with |
dissociation (symmetric) | A and B jointly establish a dissociation | 62 |
validates |
disconfirmation control | A is a control whose specific result strengthens B's warrant | 44 |
predicts |
predictive validation | A predicts B, typically model to experiment | 0 |
confirms |
predictive validation | Reciprocal of predicts, populated at build |
70 |
interprets |
reframing | A reframes empirical B through a theoretical lens | 36 |
enables-method |
methodological warrant | A is the capability that warrants B's interpretability | 76 |
scopes |
scope qualification | A is a boundary condition on B, or on every empirical claim if ["*"] |
144 |
extends |
extension | A extends B beyond its original conditions | 26 |
qualifies |
qualification | A narrows or conditions B without contradicting it | 1 |
contradicts |
opposition | A and B cannot both hold | 0 |
opposes |
opposition | A stands against B | 0 |
replicates |
replication | A is an independent finding of the same result as B | 0 |
Four sit at zero, and the zeros do not all mean the same thing.
replicates is zero because no claim in this corpus replicates another. These are ten
unrelated papers; replication needs an independent study of the same result, which nothing here
has. The replication layer is declared with no
runner precisely so that this reads as an empty column rather than as an omission.
contradicts and opposes are zero because the eliminative work in this corpus is done by
rules-out — and whether those three are distinct relations or one relation with three names is
an open question rather than a settled
distinction.
predicts is zero while its reciprocal confirms is not, which is an artefact rather than a
finding: confirms is populated at build from the forward edge, and the corpus authors have
written the confirming direction directly.
refutes at 1 is worth its own note. Until recently it was zero on
every page, not because the corpus lacked one but because the vocabulary was written down in
four scripts and none of them included it — so the single refutes: edge in the published
corpus was uncounted by the figures and dropped by every export. The vocabulary now lives in
scripts/relations.py and the four import it.
That episode is also why refutes is not treated as an opposition for checking purposes. A
paper that refutes its own prediction is closing the hypothetico-deductive loop, which is what
predictions are for; a paper that rules-out a claim it asserts has aimed an edge at the wrong
node. Collapsing the two reported eight legitimate refutations as errors.
The reasoning forms
Deduction. entails and derived-from carry the hypothesis-to-prediction direction.
Headley's hypothesis-distinct-compartmental-roles entails four predictions; each is
derived-from the same hypothesis.
The empirical loop. tests runs the other way — from a measured result back to the
prediction it settles. entails and tests are not interchangeable and not reciprocal: one is
what the hypothesis commits you to, the other is what the experiment found.
Elimination. rules-out is the move a control makes. It needs a target that the paper does
not assert, which is why alternatives are authored as claims.
Dissociation. dissociates-with is symmetric — two results that jointly establish that two
things come apart. Neither is prior to the other, which is why it is the one relation where
direction carries no meaning.
Reframing. interprets maps a result onto theory. It is an act of mapping, not a
derivation, and reading it as a derivation is the commonest way a reconstruction overstates what
a paper showed.
Warrant. enables-method and scopes say what a result depends on and where it stops.
The choice is consequential rather than cosmetic. The synthesis pipeline reads these edges as cues for the rhetorical move it should articulate, so the edge type governs the reconstruction — which is the substance of the open question on relation-vocab.