Edge inference

step · paper open

Declaration c0255f40352f has not been accepted by anyone. · no paper read under it

Which claims depend on which?

Mechanical: re-runnable from its declared inputs.

Part of Induction — What did the readers find, and what survived reconciliation?

How it works

Given the draft's claims — numbered, each with its role, panel, stance, full sentence, cited spans and the readers' evidence quotes — it returns typed relations in the corpus's own vocabulary, with a one-sentence why per edge. The relations, their binding directions and the confusable pairs come from contract/vocabulary.md, so the prompt and the checker read one rule. Validation is mechanical: an unknown or self relation is dropped, derived-from and confirms are synthesised as reciprocals rather than emitted, and the direction is enforced — tests from an empirical result or control to a prediction, entails from a hypothesis, scopes from a scope claim, rules-out/contradicts/opposes only at a claim the paper does not assert, part-of only where it forms no cycle and no second whole. An invented target is worse than a missing edge.

--dump-prompt writes the exact request and exits, so that an analyst or a different model answers the question this layer would have asked rather than a paraphrase written from memory, and --answer feeds that answer back through the same validation; --per-arc chunks the request into one call per hypothesis arc for weaker models and merges the answers. Edges are the part of a tree most easily lost to one failed call — every other stage can succeed and leave a paper with claims and no structure at all.

How to run it, in the reference

Rests on

Feeds — a change here disturbs these

How it is defined

A model answers this layer, so the prompt is the layer. It is reproduced below from the committed file, and it is a declared input — editing it makes every run that used it stale.

extract/prompts/edge-inference.mdthe prompt it runs undernot in the repository

The declaration names this path and the repository does not have it. An input that does not exist hashes to nothing, so it cannot make a run stale — the layer is declared to depend on something it is not in fact tracking.

extract/prompts/contract/vocabulary.mdthe prompt it runs undernot in the repository

The declaration names this path and the repository does not have it. An input that does not exist hashes to nothing, so it cannot make a run stale — the layer is declared to depend on something it is not in fact tracking.

What it produces78 edges

One paper, as the worked example — Gadeke, v10. Read from runs/gadeke-2026-guilt-insula/edge-inference.output.json · 21 KB. model supplied:runs/gadeke-2026-guilt-insula/edge-inference.answer.v10.json

# sourcetargetrelationwhy
1 responsibility-social-choice-yields-low responsibility-outcomes-generates-guilt-participant entails the guilt hypothesis deductively implies the happiness prediction (results-119)
2 responsibility-social-choice-yields-low responsibility-partner-outcomes-influences-participant entails if responsibility for outcomes generates guilt, a model with social_pRPE should fit happiness better
3 anterior-insula-neural-substrate-guilt anterior-insula-tracks-guilt-insula entails the insula-as-guilt-substrate hypothesis implies insula BOLD is higher Social>Partner with a Social:LowOutcome interaction (results-124)
4 functional-connectivity-between-guilt-responsibility-related connectivity-between-guilt-responsibility-related-outcome-ph entails the connectivity hypothesis implies a seed-to-voxel PPI reveals a prefrontal Condition-by-Choice interaction
5 neural-substrate-tracks-participant-responsibility neural-substrate-tracks-participant-responsibility-2 entails the responsibility-tracking hypothesis implies outcome-sensitive regions respond more to partner pRPE from participant choices
6 when-partner-received-low-lottery responsibility-outcomes-generates-guilt-participant tests the partner-outcome x decision-maker interaction tests the happiness prediction
7 when-partner-received-low-lottery responsibility-outcomes-generates-guilt-participant confirms the interaction was significant in both studies: happiness lower after low partner outcomes from own choice
8 insula-rois-responded-more-low anterior-insula-tracks-guilt-insula tests the insula ROI Social>Partner low-outcome response tests the insula BOLD prediction (results-127)
9 insula-rois-responded-more-low anterior-insula-tracks-guilt-insula confirms insula ROIs responded more to low partner outcomes in Social, mirroring the guilt effect
10 mass-univariate-voxel-wise-analysis-found-small anterior-insula-tracks-guilt-insula tests the mass-univariate voxel-wise search tests the insula BOLD prediction (results-129)
11 mass-univariate-voxel-wise-analysis-found-small anterior-insula-tracks-guilt-insula confirms a small left anterior insula cluster survived small-volume FWE correction (p = 0.024)
12 likelihood-ratio-test-showed-responsibility responsibility-partner-outcomes-influences-participant tests the likelihood-ratio comparison tests whether the Responsibility model (with social_pRPE) fits best
13 likelihood-ratio-test-showed-responsibility responsibility-partner-outcomes-influences-participant confirms the Responsibility model fitted better than all other models in both studies
14 partner-reward-prediction-errors-resulting responsibility-partner-outcomes-influences-participant tests the social_pRPE weight tests the prediction that partner pRPE from own choices contributes
15 partner-reward-prediction-errors-resulting responsibility-partner-outcomes-influences-participant confirms social_pRPE weights were reliably greater than zero in both studies
16 one-cluster-left-sts-responded neural-substrate-tracks-participant-responsibility-2 tests the left-STS search tests the partner-pRPE-from-own-choice prediction (fig4h)
17 one-cluster-left-sts-responded neural-substrate-tracks-participant-responsibility-2 confirms one left-STS cluster responded more to partner pRPE from participant than partner choices (pFWE = 0.022)
18 functional-connectivity-between-left-anterior connectivity-between-guilt-responsibility-related-outcome-ph tests the insula-seed PPI tests the prefrontal Condition-by-Choice interaction prediction (results-142)
19 functional-connectivity-between-left-anterior connectivity-between-guilt-responsibility-related-outcome-ph confirms a right-IFG cluster's insula connectivity varied with condition and choice and survived FWE correction
20 when-partner-received-low-lottery responsibility-social-choice-yields-low supports the behavioural guilt effect is evidence for the guilt hypothesis
21 likelihood-ratio-test-showed-responsibility responsibility-social-choice-yields-low supports the Responsibility model's superior fit supports the responsibility/guilt account
22 partner-reward-prediction-errors-resulting responsibility-social-choice-yields-low supports a positive social_pRPE weight supports that partner outcomes from own choices affect happiness
23 insula-rois-responded-more-low anterior-insula-neural-substrate-guilt supports the insula guilt-tracking response supports the insula-as-substrate hypothesis
24 one-cluster-left-sts-responded neural-substrate-tracks-participant-responsibility supports the left-STS responsibility signal supports the responsibility-tracking hypothesis
25 functional-connectivity-between-left-anterior functional-connectivity-between-guilt-responsibility-related supports condition/choice-dependent insula connectivity supports the connectivity hypothesis
26 prior-literature-documents-association-between anterior-insula-neural-substrate-guilt supports the prior insula-guilt association is why the insula was hypothesised as the substrate
27 prior-functional-connectivity-work-shown functional-connectivity-between-guilt-responsibility-related supports prior functional-connectivity work motivates the connectivity hypothesis
28 rutledge-colleagues-established-changes-momentary likelihood-ratio-test-showed-responsibility supports Rutledge's momentary-happiness framework grounds the fitted happiness models
29 participant-momentary-happiness-varied-rewards likelihood-ratio-test-showed-responsibility supports happiness varying with participant rewards grounds fitting reward-based happiness models
30 participant-momentary-happiness-varied-rewards-2 likelihood-ratio-test-showed-responsibility supports happiness varying with partner rewards grounds the partner-RPE terms of the models
31 participants-own-reward-prediction-errors likelihood-ratio-test-showed-responsibility supports the ordering of RPE weights corroborates the Responsibility model structure
32 when-partner-received-low-lottery both-studies-participants-felt-worse supports the guilt interaction is the evidence for the synthesis that participants felt worse from their own choices
33 risk-premiums-not-differ-between participants-showed-very-similar-risk supports equal risk premiums across conditions support the similar-risk-preferences synthesis
34 participants-slightly-more-risk-averse participants-showed-very-similar-risk supports only a slight Study-1 risk-aversion difference supports the 'similar with a tendency' synthesis
35 no-significant-interaction-between-difference participants-showed-very-similar-risk supports no EV-difference-by-condition interaction supports similar preferences across conditions
36 left-superior-temporal-sulcus-cluster one-cluster-left-sts-responded supports the STS response across both sessions corroborates the left-STS responsibility result
37 dot-products-between-individual-neural insula-rois-responded-more-low validates positive dot products with the published Yu GRBS give convergent validity to the insula guilt response (results-152)
38 manipulation-check-bilateral-ventral-striatum model-based-glm-entered-best-fitting-computational validates the ventral-striatum manipulation check confirms the model-based GLM recovers reward signals
39 guilt-effect-occurred-whether-participant when-partner-received-low-lottery validates the guilt effect held whether the participant's own outcome was high or low, ruling out own-outcome confounds
40 responsibility-choices-not-influence-happiness when-partner-received-low-lottery validates no responsibility effect after high partner outcomes shows the effect is specific to low outcomes (guilt)
41 risk-aversion-parameter-not-differ-between momentary-happiness-modelled-five-computational validates equal risk-aversion across gain/loss trials justifies pooling in the models
42 pre-task-icebreaker-succeeded-establishing-positive when-partner-received-low-lottery validates the icebreaker established a positive partner attitude, warranting that the effect reflects guilt toward a valued partner
43 lower-happiness-when-participant-decision-maker participant-happiness-lower-when-participant interprets responsibility aversion is offered as a reading of the agency effect (results-073)
44 behavioural-guilt-effect-larger-happiness when-partner-received-low-lottery interprets 'simple guilt' is the interpretive reading of the behavioural guilt interaction (results-080)
45 authors-suggest-left-sts-region one-cluster-left-sts-responded interprets the STS-tracks-partner-outcomes reading interprets the left-STS result (results-138)
46 connectivity-between-left-anterior-insula functional-connectivity-between-left-anterior interprets the prefrontal-sensitivity reading interprets the insula-IFG connectivity result (abstract-009)
47 during-outcome-phase-responses-low insula-rois-responded-more-low part-of the outcome-phase table is one measure of the insula guilt response the whole result states
48 difference-response-between-low-high insula-rois-responded-more-low part-of the low-minus-high difference is one measure of the same insula guilt response
49 when-partner-received-low-lottery linear-mixed-model-containing-all requires the reported guilt interaction depends on Model 5's LMM being the warranted model
50 one-cluster-left-sts-responded model-based-glm-entered-best-fitting-computational requires the STS responsibility result depends on the model-based GLM2 regressors
51 likelihood-ratio-test-showed-responsibility model-selection-among-happiness-models requires the model-comparison result depends on the likelihood-ratio selection procedure
52 insula-rois-responded-more-low during-receipt-lottery-versus-safe requires the insula guilt ROI depends on the outcome-responsive localizer regions
53 hold-partner-behaviour-constant-across when-partner-received-low-lottery enables-method simulating the partner's choices holds behaviour constant, making the guilt comparison possible
54 model-selection-among-happiness-models likelihood-ratio-test-showed-responsibility enables-method the likelihood-ratio selection method makes the model-comparison result interpretable
55 parameter-recovery-procedure-synthetic-data-generated responsibility-redux-model-incorporating-expected enables-method parameter recovery establishes the model estimates are stable enough to trust the fit
56 findings-rest-two-samples-healthy when-partner-received-low-lottery scopes the guilt effect is bounded to two samples of healthy adults
57 each-trial-participants-chose-between when-partner-received-low-lottery scopes the effect is bounded to the three-condition safe/risky choice paradigm
58 study-reproduced-study-design-inside insula-rois-responded-more-low scopes the neural results are bounded to the fMRI reproduction of the design in Study 2
59 participant-happiness-lower-when-participant when-partner-received-low-lottery dissociates-with two effects both present and distinct: agency lowers happiness regardless of outcome, guilt only after the partner's loss
60 dot-products-between-individual-neural individual-grbs-dot-product-values-not in-tension-with the neural guilt response matches the published Yu/Koban signature at the group level (60), yet individual signature scores do not track individual behavioural guilt (61); both stand and the null boun…
61 only-precuneus-tpj-showed-positive decisions-social-compared-solo-condition part-of the precuneus/TPJ Risky-Safe analysis (fig4c) is one component of the Social>Solo cluster result (fig4b)
62 decisions-social-compared-solo-condition each-trial-participants-chose-between requires the Social>Solo contrast requires the three-condition (Solo/Social/Partner) choice paradigm to have been run
63 responsibility-model-yielded-higher-values likelihood-ratio-test-showed-responsibility supports the Responsibility model's higher R2 (table1) corroborates the likelihood-ratio verdict that it fits best
64 responsibility-model-yielded-higher-values momentary-happiness-modelled-five-computational requires the R2 model comparison (table1) requires the five happiness models to have been fitted
65 among-computational-models-fitted-momentary momentary-happiness-modelled-five-computational requires the Responsibility Redux AIC ranking (table1) requires the five happiness models to have been fitted
66 responsibility-redux-model-incorporating-expected momentary-happiness-modelled-five-computational requires the Responsibility Redux model's predictions (fig3c,fig3g) require that model to have been fitted
67 parameter-recovery-procedure-synthetic-data-generated responsibility-redux-model-incorporating-expected validates the parameter-recovery procedure (fig3s1) verifies the Responsibility Redux estimates are stable, strengthening the fit
68 participants-chose-risky-option-lottery participants-showed-very-similar-risk supports the Solo>Social lottery-choice difference (fig2a,fig2d) is one of the risk-preference results the synthesis draws together
69 participants-chose-risky-option-lottery each-trial-participants-chose-between requires the Solo/Social choice-proportion result requires the three-condition choice paradigm to have been run
70 mixed-effects-regressions-choices-social-condition participants-chose-risky-option-lottery part-of the mixed-effects choice regression (app1table1) is the formal analysis behind the reported Solo/Social choice difference
71 linear-mixed-model-containing-all when-partner-received-low-lottery validates Model 5 with all two-way interactions fits best (app1table2), validating that the reported guilt interaction is drawn from the warranted model
72 both-studies-participants-felt-worse responsibility-social-choice-yields-low supports the across-both-studies guilt effect is evidence for the responsibility-guilt hypothesis; the edge task wrote edges into the synthesis and none out (#53)
73 left-ifg-cluster-showed-opposite connectivity-between-left-anterior-insula qualifies the uncorrected left-IFG cluster showed the opposite connectivity pattern, so it bounds the connectivity finding rather than confirming it (#53)
74 responsibility-outcomes-generates-guilt-participant responsibility-social-choice-yields-low derived-from reciprocal of entails (responsibility-social-choice-yields-low → responsibility-outcomes-generates-guilt-participant)
75 responsibility-partner-outcomes-influences-participant responsibility-social-choice-yields-low derived-from reciprocal of entails (responsibility-social-choice-yields-low → responsibility-partner-outcomes-influences-participant)
76 anterior-insula-tracks-guilt-insula anterior-insula-neural-substrate-guilt derived-from reciprocal of entails (anterior-insula-neural-substrate-guilt → anterior-insula-tracks-guilt-insula)
77 connectivity-between-guilt-responsibility-related-outcome-ph functional-connectivity-between-guilt-responsibility-related derived-from reciprocal of entails (functional-connectivity-between-guilt-responsibility-related → connectivity-between-guilt-responsibility-related-outcome-ph)
78 neural-substrate-tracks-participant-responsibility-2 neural-substrate-tracks-participant-responsibility derived-from reciprocal of entails (neural-substrate-tracks-participant-responsibility → neural-substrate-tracks-participant-responsibility-2)

Across the corpus

9 not run · 1 blocked upstream·a paper links to its own cell, where this layer's output for it is rendered

Inputs and outputs

Reads, besides its dependencies
Produces
  • runs/{paper}/edge-inference.output.json

One per paper — the table above links each one that exists.

Views
  • graph — on the paper page, as the claim graph
  • table — rendered above, over the 78 edges in the artifact
  • comparison — on the cell page, two versions aligned by the matcher, wherever the ledger holds more than one

Running it

The command comes from the declaration, so this text and what actually runs cannot diverge. pipeline.py run also runs the unmet dependencies first.

python3 scripts/pipeline.py run <paper> edge-inference

Underneath, that runs cd extract && python3 -m claim_graphs.cli edge-inference --paper {paper}.