core/docs/decisions/ADR-0203-binding-graph-acyclicity-invariant.md

ADR-0203 — Binding-Graph Acyclicity Invariant (circular_dependency refusal)

Status: Accepted (proof_chain phase 2.1 — the isolated guard; ADR-0201 §Deferred) Date: 2026-06-02 Relates to: ADR-0132 (binding-graph data model — this adds an invariant to its construction contract; see §amend-vs-additive), ADR-0201 (canonicalizer + phase plan), ADR-0202 (proposition representation contract).

---

Context

ADR-0132's SemanticSymbolicBindingGraph.__post_init__ enforces referential integrity (every BoundEquation dependency names a known symbol) but not acyclicity. A cycle in the equation dependency structure — x defined from y, y defined from x, or a self-dependency x ← x — is circular reasoning: structurally well-formed, semantically invalid. It is the proof-domain analog of the 20/5 == 4 class the arithmetic gate refuses.

proof_chain (ADR-0201) is the first consumer that can build such a structure: its phase 2.2 wiring constructs binding graphs from proofs, where a malformed proof could close a dependency loop. Per CORE's "build the defensive refusal NOW" discipline, the guard must exist before the structure where a cycle can exist — so this phase (2.1) lands it before any 2.2 wiring.

Decision

Add an acyclicity invariant to binding-graph construction:

• generate/binding_graph/acyclicity.py — a pure find_cycle(adjacency) cycle detector (deterministic three-colour DFS; sorted traversal → byte-stable reported cycle; self-edge → length-1 cycle). No model import — testable in isolation against synthetic adjacency graphs.
• Enforcement at the shared construction boundary. __post_init__ builds a {lhs_symbol_id: ⋃ dependencies} adjacency over its equations and calls find_cycle; a non-None result raises BindingGraphError("circular_dependency: equation dependency cycle …") naming the cycle. This runs on every binding graph — math and (future) proof alike — making a cyclic graph unrepresentable for all consumers.

Why the shared __post_init__ (not a proof-only check)

Putting it at the substrate boundary closes the gap universally and makes illegal states unrepresentable (the CLAUDE.md design principle), rather than leaving the math binding graph cycle-unchecked. The guard exists the instant the structure becomes constructible — strictly before the proof wiring that makes a cycle reachable.

Math-lane regression proof (the shared-constructor risk)

Because the guard runs on the existing math/algebra path too, it must refuse no existing graph. Verified:

• The only production producer is generate/binding_graph/adapter.py (bind_math_problem_graph). It mints a fresh result symbol per operation (_op_result_symbol_id(idx)) and depends only on symbols that already exist — edges point strictly backward in construction order. It is therefore acyclic by construction and cannot produce a graph this guard would refuse.
• No existing test constructs a dependency cycle (the sym_ghost fixture is the referential-integrity refusal, not a cycle).
• The full binding-graph + admissibility test surface — 392 tests — stays green with the guard in __post_init__. (Smoke: 67 passed.)

A future/in-flight consumer that did build a cycle would now be refused at construction — which is the point.

amend-vs-additive (ADR-0203 vs amending ADR-0132)

Decision: a new additive ADR (this one) that references ADR-0132 — not an amendment of the closed record.

ADR-0132 is Accepted. The acyclicity invariant was not part of its original decision; it became necessary later, when proof_chain made cycles reachable. Recording it as a new invariant preserves that history — why the guard was added and when it became load-bearing — which an in-place edit of the closed record would erase. This mirrors the history-vs-current-state discipline used elsewhere (append the new fact; don't rewrite the settled one). The code lives beside ADR-0132's referential-integrity check in __post_init__; the decision record is additive.

Honesty boundary (load-bearing — carried by every phase-2 ADR, 0203–0205)

Through phase 2.3, proof_chain is sound over its declared atoms — it does not reason over recognized input. Atoms are opaque/declared symbol ids (ADR-0202); grounding them to ADR-0144 EpistemicNode/FeatureBundle carriers is phase 2.4 (fork B). This must never be softened to "reasons over input" before 2.4 lands. This ADR is structure-only and makes no grounding claim; it is named here so the boundary travels with the work, not just with the rule ADRs.

Evidence

• tests/test_binding_graph_acyclicity.py — 17 tests: pure checker (acyclic→None; self-loop/2-/3-cycle/cycle-with-tail detected; deterministic), construction enforcement (cyclic + self-dependent refuse; acyclic + adapter-shape construct), and referential-integrity-still-fires-first.
• Mutation-verified non-vacuous: neutering find_cycle (→ always None) makes a 2-cycle construct without refusal — test_two_cycle_equation_set_refuses would fail. The guard is load-bearing, not decoration.
• Full binding-graph/admissibility surface: 392 passed. Smoke: 67 passed.

Deferred (not in this phase)

• 2.2 — proof-graph builder (proof → BoundEquations; canonical_key → rhs_canonical); the first construction that exercises this guard through the real proof path (ADR-0204).
• 2.3 — modus_ponens + the disagreement rule (ADR-0205).
• 2.4 — atom→EpistemicNode carrier grounding (ADR-0206).

Governance Cross-Reference (ADR-0225)

This late-corpus ADR is governed by ADR-0225:

• Safety boundaries: changes must preserve ADR-0027/0028/0029 identity and safety-pack boundaries; no identity, safety, or policy mutation is implied unless explicitly reviewed.
• Versor closure: runtime field paths must preserve versor_condition(F) < 1e-6; this ADR does not authorize hidden normalization or hot-path drift repair.
• Reconstruction-over-storage: evidence must remain reconstructive and content-addressed rather than duplicating opaque state.
• Replay-equivalence: serving, teaching, promotion, or checkpoint changes require a named deterministic replay / byte-equivalence gate.
• Mutation standing: any durable corpus, pack, policy, or epistemic-status mutation remains reviewed, proposal-only until accepted, or proof-carrying as applicable.