Read strict-rule antecedents as sets to stop ASPIC+ argument multiplication

[ctoth]

Implements the maintainer's decision on #102 (options A1 + A2): rule antecedents (strict and defeasible) are read as sets.

Change

Rule.__post_init__ canonicalises the antecedents of every rule. It drops repeated literals and stores the remaining ones sorted. As a result:

• A2: ~q, ~q -> p becomes ~q -> p.
• A1: permuted strict rules with the same antecedent set and consequent are equal Rule values, so frozensets of rules merge them.

Defeasible rules get the same set reading (maintainer follow-up decision, commit 17d7df7); named defeasible rules stay distinct by n(r), so undercuts still target them.

Why at rule construction: it is the one place every producer passes through. That covers hand-built ArgumentationSystems, transposition_closure (which rebuilds rules), Datalog grounding and ABA frameworks. Canonicalising at system construction would miss rules used outside a system (the ABA frameworks), and would let two differently-ordered Rule values that mean the same rule compare unequal.

Convention: M&P 2018 Def 5 (p.9) and Prakken 2010 Def 3.6 build A1..An -> φ from a rule φ1..φn -> φ and don't settle whether the antecedents are a sequence or a set. The set reading is a chosen convention, documented in the Rule docstring and recorded on #102.

Effect on the #102 theory

18,850 → 98 arguments, and about 306M → 7,226 attack pairs. The pinned operational contract (at most 200 arguments and at most 20,000 attack pairs) now passes, and its xfail(strict=True) marker is removed.

Tests (tests/structured/aspic/test_aspic_rule_sets.py)

• Set reading pinned:
  • a repeated literal is dropped;
  • permuted strict rules are one rule;
  • defeasible rules: repeated antecedents dropped, permuted unnamed rules equal, named rules distinct by name;
  • ~q, ~q -> p over two ~q arguments builds 2 arguments for p, not 4.
• A1 property: a permuted duplicate strict rule adds only clone arguments. Grounded, complete, preferred and stable extensions are unchanged up to clones, compared as sets of conclusions per extension. The generator makes the duplicated rule fire, so every example contains clones.
• Producer audit:
  • Datalog grounding merges strict rules with permuted or repeated bodies and keeps both source ids (Datalog projection loses source provenance when distinct strict rules ground identically #67).
  • transposition_closure emits canonical rules.
  • The ASPIC+ ASP encoding already counts distinct antecedents (ASPIC ASP backend fails to derive rules with repeated antecedents #64).
  • ABA ASP body counts are consistent, since ABA strict rules are now canonical.

Closes #102

🤖 Generated with Claude Code

https://claude.ai/code/session_01V4tAVyyKcs1sYcsEzL7Bzg