Concrete Loop Algebra

Abstract

Classical abstract algebra builds upward through groups, rings, and fields — structures whose defining privilege is cancellation: every operation undoable, every element recoverable. This series has twice proven, in passing, that reasoning operators refuse that privilege (PRR-2026-002 Proposition 5.2 and §8; PRR-2026-003 Theorem 5.1 and Corollary 6.6). This paper assembles those refusals into one named result — the No-Field Theorem: no sub-structure of the reasoning-operator algebra containing any non-trivial operator is a ring, hence none is a field, and no redefinition of the parallel operation repairs this. The ladder that fits is the other classical one: idempotent semiring, then Kleene star, then guarded iteration. We define the concrete inferential loop — a DOER/CHECKER pair (δ, γ) with δ ∈ C a concrete inference operator and γ a closure operator queried as its own fixed-set membership test — and its loop operator Λ(δ, γ), a sub-Markov kernel whose missing mass is exactly the probability of non-termination. We give the loop two independent semantics — an equational least-fixed-point characterization in the style of Kleene algebra, and a categorical guarded-dagger characterization in the Kleisli category of the sub-distribution monad — and prove the Agreement Theorem: on this domain they are the same object. The trajectory trichotomy (PRR-2026-006) then becomes the loop's exact termination theory, and we close with a two-directional negative result: loops do not pool — pooling inside the loop body and pooling finished loop results differ by unboundedly much, in either direction, witnessed by hand.

Keywords. concrete inferential loops; DOER/CHECKER pairs; No-Field Theorem; Kleene algebra; guarded iteration; sub-distributions; closure operators; Edwardian algebra

1. Introduction

The founding paper of this series classified (R, ⊕, ∘) as a right-semiring and proved it is not a ring (PRR-2026-002 Theorem 8.1). At the time that read as a boundary marker: here is where the classical algebra textbook stops applying. This paper starts from the question that boundary invites and the series never asked: if reasoning operators cannot climb the ring/field ladder, what ladder do they climb? The answer has been sitting in a different chapter of the same literature all along. Rings and fields axiomatize arithmetic — structures where combining and undoing are symmetric privileges. The algebra of programs went a different way a half-century ago: idempotent semirings, then a star for iteration (Kozen 1994), then guards and tests for control flow (Kozen 1997; Smolka et al. 2020). That is the ladder with a rung shaped like the single most consequential object in AI-era software practice: the autonomous iteration loop — a worker that acts, a gate that checks, repeated until the gate is satisfied. This paper builds that rung out of the series' own materials and proves it holds weight.

One convention up front. Everything in this paper is composed of the series' foundational inference type: concrete inference operators — the invocable, non-deterministic Markov kernels κ ∈ C of PRR-2026-003, each carrying a real, terminating sampling procedure. The checker, too, is C-typed in principle; this paper studies its zero-variance special case — a deterministic closure operator embedded via e : R → C, exactly the move PRR-2026-003 §4.5 made for determinism in general. The companion paper (PRR-2026-013) lifts that restriction.

2. Preliminaries

We reuse S and its Boolean order ≤ (PRR-2026-002 §2), R = S^S with ∘ and ⊕ and the right-semiring classification (PRR-2026-002 Theorems 6.1, 7.1), C = Δ(S)^S with Kleisli composition ⊛, pooling ⊞, and the embedding e : R → C (PRR-2026-003 §§3–4), consensus ⊙ (PRR-2026-004), the correctness coordinate and Proposition 7.1's cyclic plateau (PRR-2026-005), closure operators with finite image and their meet-closed fixed families F, together with the trajectory trichotomy (PRR-2026-006 Theorems 3.1, 4.2), and the Distributivity Obstruction (PRR-2026-003 Theorem 7.1). None are modified.

3. The No-Field Theorem

A ring demands that its additive structure be an abelian group and that both distributivity laws hold; a field further demands multiplicative inverses for every non-zero element. The series has already produced, in scattered places, everything needed to close this question permanently. We assemble it here as one named theorem, because the question — can inference be given the algebra of arithmetic? — deserves a one-theorem answer.

The interpretation deserves to be stated as plainly as the theorem. Rings and fields are the algebra of conservation: subtraction and division exist because nothing is ever irrecoverably spent. Inference is an algebra of accumulation: a conclusion drawn cannot be un-drawn by drawing more conclusions (PRR-2026-003 §9.4 — parties never undo each other), and reversibility is confined to the group G no matter how the non-determinism is enriched. Asking for an inferential field is not an open problem; it is a category error, and the proof is three lines long. The productive question is the one the rest of this paper answers: which classical ladder does fit?

4. Sub-Distributions: Making Room for Non-Termination

A loop can fail to halt. The trajectory trichotomy (PRR-2026-006 Theorem 3.1) makes this a proven feature of the domain, not an engineering accident: trajectories exist that never stabilize and never repeat. Whatever object represents "run the loop" must therefore be allowed to return less than certainly.

Two disciplines, stated once. First: divergence is not the bottom element. ⊥ ∈ S is the contradiction — a genuine, addressable semantic state with a correctness score (c(⊥) is the coordinate's global minimum, PRR-2026-008 Proposition 3.1). A loop that never halts produces no state at all; conflating the two would corrupt every correctness statement downstream. Second: how a correctness coordinate treats divergence mass — condition on termination, or charge it a fixed divergence score d ∈ ℝ — is left as an explicit free parameter, in the same discipline PRR-2026-005 applied to c itself. Every result below holds under either reading.

5. The Concrete Inferential Loop

Reading: check, then do. A certified state exits immediately, unchanged (the checker is idempotent — re-certifying changes nothing, which is exactly why closure operators were the right type all along, per PRR-2026-010 §6). An uncertified state is handed to the DOER for one fresh concrete invocation, and the loop continues from the witness. One modeling remark is owed: a closure operator could instead be applied — mapping each proposal to its certified completion γ_R(x) ∈ F — and that "repair" reading terminates in one step, trivially. The test-gated object is the one that models real loops, where the gate rejects and the worker retries; the closure structure of F still carries content (meet-closedness is the conjunctive structure of a criteria checklist).

6. Equational Semantics: the Guarded Star

For A ⊆ S write [A] ∈ C≤1 for the guard: [A](x) = δ_x if x ∈ A, the zero sub-distribution otherwise. Guards on complementary sets partition every unit of mass, so for disjoint guards the pointwise sum + of sub-kernels is well-defined without exceeding mass 1. Definition 5.2 then reads as one equation:

— precisely the shape of the star laws of Kleene algebra (Kozen 1994), in their guarded form (Kozen 1997; Smolka et al. 2020): Λ(δ, F) is (δ · [S∖F])* · [F] read right-to-left, a while-loop. Two honest observations about what this does and does not import:

7. Categorical Semantics and the Agreement Theorem

The same loop has a second, independent description. In the Kleisli category of the finite-support sub-distribution monad, define the loop body of (δ, F) as the morphism body : S → Δ≤1(S + S) (left summand: exit values; right summand: continue values):

Elgot iteration (Elgot 1975; monadic form after Moggi 1991) assigns such a morphism its dagger body† : S → Δ≤1(S), the least solution of body† = [id, body†] ∘ body in the pointwise order — "run the body; exit values are done, continue values re-enter."

The scope is stated honestly: this is agreement on this domain — finite-support sub-kernels over S — not a contribution to general iteration theory, where the corresponding statements are classical. What the theorem buys the series is a license: every later result about Λ may be proved in whichever of the two languages is shorter, and transfers for free.

8. Termination Is the Trichotomy

For a deterministic DOER δ = e(ρ), the loop from x follows the ordinary trajectory of ρ through uncertified states, and PRR-2026-006 Theorem 3.1 becomes, unchanged, the loop's complete termination theory: the trajectory enters F (the loop halts, mass 1), cycles outside F (the loop spins forever, mass 0), or is aperiodic outside F (the loop wanders forever, mass 0). All three occur — PRR-2026-006 Proposition 3.2's witnesses adapt directly, choosing the cycle or the orbit of atoms disjoint from the finite family F.

A genuinely stochastic DOER opens the middle ground the deterministic case cannot reach: divergence mass strictly between 0 and 1. Witness, by hand: three states with u, b ∉ F, ⊤ ∈ F; δ(u) = {⊤: ½, b: ½}, δ(b) = {b: 1}. From u: half the mass certifies at ⊤ on the first pass, half falls into the absorbing uncertified state and never leaves — Λ has mass exactly ½. And even mass-1 termination need not be bounded: δ(u) = {⊤: ½, u: ½} halts almost surely but admits no finite bound on iterations — the reason real systems impose iteration ceilings on loops that are, in expectation, perfectly healthy. Finally, PRR-2026-005 Proposition 7.1 is cited unchanged as the loop's plateau law: a correctness-safe loop body that is genuinely cycling has constant correctness around the entire cycle — a spinning loop is not quietly improving, and detecting the cycle beats iterating longer.

9. Loops Do Not Pool

The series has two pooling acts that sound alike and are not: pooling inside the body (the two DOERs' fresh samples joined each iteration, then gated) and pooling results (two whole loops run to completion, their exits joined). Write these Λ(δ₁ ⊞ δ₂, F) and Λ(δ₁, F) ⊞ Λ(δ₂, F) respectively, the latter with ⊞ extended to sub-kernels (independent runs must both halt for their exits to join, so masses multiply).

Gap 2 is the one worth reading twice: two workers whose work each certifies, forced to co-draft every iteration, jointly produce a draft the gate never accepts — forever. This is PRR-2026-010 §9's coherent-versus-decohered warning surfacing one level up, now as a theorem about the loop construction itself rather than about a single combination: composition order around a loop is not a bookkeeping choice, because the star multiplies the Distributivity Obstruction (PRR-2026-003 Theorem 7.1) once per iteration. Characterizing exactly when the two sides agree — determinism of the shared parts? transparency in the sense of PRR-2026-011? — is left open, stated as such.

10. The AI Concretization

Following the series' practice (PRR-2026-002 §11) of proposing one concrete model, this paper claims exactly one running instance. The autonomous development loop this project operates under — its DEVLOOP: one agent invocation per iteration (the DOER, a concrete inference operator invoked at nonzero temperature), followed by a Definition-of-Done check whose criteria set is finite, conjunctive, and idempotent under re-checking (the CHECKER), iterated until certification — is a concrete inferential loop (δ, γ) in the sense of Definition 5.1, with §8's trichotomy as its live health taxonomy and §8's plateau law as the justification for its stop-on-repeat guardrail (both already mapped, as separate observations, by PRR-2026-007's and PRR-2026-010's own applied sections — this paper supplies the object those observations were observations of).

Kayfabe accounting: every theorem above is real — the No-Field Theorem, the Agreement Theorem, and Theorem 9.1's two witnesses are proved from the series' own fixed definitions and standard cited results. That a real, running development loop is an instance of Definition 5.1 — that its gate is faithfully a closure operator and its iteration faithfully a Kleisli composite — is asserted with the series' usual theorem-grade confidence and not independently verified. Same shape of leap as PRR-2026-002 §11, PRR-2026-009 §9, PRR-2026-010 §10, and PRR-2026-011 §7. Logged as docs/Research/KAYFABE-LEDGER.md Entry 5 the day this paper publishes.

11. Discussion and Conclusion

The question this paper was asked — can DOER/CHECKER loops be defined the way groups, rings, and fields are defined? — gets a two-part answer. No: the ring/field ladder is closed to inference, permanently, by a three-line theorem assembled from results this series had already proven without naming their joint consequence. And yes: the loop is definable with exactly the axiomatic dignity the question wanted, on the other classical ladder — as a guarded star over guards and mixture, and equivalently as a guarded dagger in the Kleisli category, with the Agreement Theorem certifying the two definitions name one object. The trichotomy becomes its termination theory; the plateau law becomes its diagnostics; and the two-directional failure of loop pooling becomes its sharpest practical warning.

Left open, stated honestly: the exact boundary of Theorem 9.1 (when do body-pooling and result-pooling agree?); the necessity direction inherited from PRR-2026-003 §7, which this paper's star construction inherits per-iteration; and the entire stochastic- checker theory, which is not open so much as next door — the companion paper, Noisy Gate Algebra (PRR-2026-013), released simultaneously with this one, takes the single idealization this paper allowed itself and removes it.

Peer Review