📄 The Handling of Loops in Talmudic Logic, with Application to Odd and Even Loops in Argumentation

Summary

The paper looks at how Talmudic reasoning handles cyclic attacks in abstract argumentation networks (S, R). Odd cycles are the hard case. Under standard Dung semantics they usually leave every argument undecided, yet a court still has to reach a ruling. The authors take a principle from Rabbi Shimon Shkop: if assuming an argument is accepted forces the conclusion that it is rejected, then that argument is struck out. They implement this with an added annihilator node. The test builds on Caminada's Socratic-discussion procedure, which reasons backwards along attack arrows, and a tableaux algorithm is given in an appendix. They also add time stamps to arguments and attacks, so that only the newest argument that closes a loop is tested. The motivating case is a conditional divorce that creates an odd loop between three marriage-related claims. Applying the principle rejects the final, self-undermining marriage and leaves a definite extension. The approach is positioned against CF2 and Stage semantics. It gives a principled, history-aware way out of deadlocks where standard semantics returns only "undecided."

Key points

• Basic contrast: a two-node mutual attack has three complete extensions ({a}, {b}, ∅), while a three-node cycle has only ∅.
• The paper lists three ways to use valuations V : S → U:
  • (a) build V into the extension definition (Amgoud, Bench-Capon), which Caminada criticises;
  • (b) compute extensions first, then pick among them with V (the Talmud; Wakaki 2010);
  • (c) apply V to the internal parts of arguments (Prakken 2010).
• Shkop temporal frames are N = (S, R, T), where T time-stamps both nodes and attacks. The Shkop test for an argument is run on the sub-network that exists at that argument's time.
• Worked examples:
  • the Jane/John/Terry/Frank divorce, where the March 31 marriage to Frank is the argument that gets annihilated;
  • a gift of a computer conditional on later work, where a sale that would cancel its own precondition is disallowed.
• Appendix A's tableaux can check whether several nodes can sit together in one admissible set. Appendix B models future-conditional actions with backward causality as state revision.

Interesting ideas

• I1. Self-defeat test: tentatively accept a claim, propagate the consequences, and reject the claim if it ends up contradicting itself. This yields a decision in exactly the cases where neutral semantics give only "undecided."
• I2. Time-stamped conflict resolution: when a loop appears, test only the most recently added element, because earlier steps were reasonable given what was known at the time they happened.
• I3. Incremental networks (S_n ⊆ S_{n+1}): resolve loops step by step as each new argument arrives, rather than searching a whole tangled graph for a global fix.
• I4. Conditional actions whose validity depends on a future event (backward causality) are modelled as precondition/postcondition pairs with later cancellation.
• I5. Make the structural reason for a verdict explicit, here as an annihilator node, instead of silently deleting an attack edge.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (world): When the world builder meets cyclic or self-cancelling states in a story, it could run this check and keep an explicit trace of which assumption collapsed.
• I2 (world): Contradictory facts from later in a narrative could be the ones retracted first, with event order serving as an explainable tie-breaker.
• I3 (learn): Self-induced repair rules could be added one at a time, and any new rule that creates a contradiction cycle with existing rules would be rejected.
• I4 (answer): Questions about tales involving bargains or promises ("if you do X by Saturday") could be answered by checking whether a later action makes an earlier transfer void.
• I5 (frames): When a frame is rejected for being pragmatically self-undermining, attach an explicit "defeated-by-self" marker so the explanation chain stays inspectable.