THE REASONING OPERATORS TOUR
Twelve papers, one continuous argument. Each stop links forward to the next paper and back to this page; each paper page carries a small tour badge near the top and a prev/tour/next strip at the bottom, so you can always tell where you are and step to either side without hunting through the nav.
The throughline: start with reasoning operators that are perfectly deterministic. Relax that to operators you can still invoke — call once, for one real answer — but which aren't deterministic anymore, and combine them by pooling. Take the dual combinator, consensus, and see how much of the pooling paper's structure survives by symmetry alone. Then step back and ask what happens to a real-valued correctness measurement laid across all of it, as those operators are applied. Then close the arc: iterate those same operators under a certification gate — the DOER/CHECKER loop — first with a perfect gate, then with a noisy one. Each stop only needs the ones before it.
Edwardian Algebra
The foundation. Reasoning operators are just functions ρ : S → S on a semantic domain — deterministic, always giving the same output for the same input. Composing them sequentially gives a monoid; combining them in parallel (pooling, via ∨) gives a commutative monoid with no inverses. Together they form a right-semiring — not a ring, and the paper proves exactly why not. The invertible operators form a group G, and idempotent operators turn out to be exactly retractions.
- Start here if you want the cleanest possible version of the algebra, before any non-determinism enters at all.
- Everything downstream either embeds this paper's results directly or refines them.
Concrete Inferential Algebra
What happens when a reasoning operator isn't deterministic, but you can still call it and get one real, checkable answer? This paper builds that algebra on Markov kernels, embeds Stop 1's deterministic operators as the zero-variance special case, and proves the central result of the whole series: right-distributivity, unconditional in the deterministic case, breaks the instant an operator becomes genuinely invocable — because reusing one sample and drawing two fresh ones stop being the same act. It also shows invertibility never gets richer just because things became probabilistic: the same group G from Stop 1 is still the only thing that's ever reversible.
- The paper's one open question (is determinism necessary, not just sufficient, for universal distributivity) is stated honestly as unresolved.
- Read this after Stop 1 — it cites Stop 1's theorems directly rather than re-deriving them.
Concrete Consensus Algebra
Stop 2 combined independent invocations by pooling — either conclusion is enough. This paper takes the opposite combinator: consensus, where only what both branches confirm survives. Because Boolean complementation swaps ∨↔∧, almost everything transfers from Stop 2 by the Duality Principle alone, symbol for symbol — the paper's discipline is deriving nothing that duality already gives for free, and independently verifying the one thing it doesn't (whether the sampling procedure itself transfers, not just the algebra).
- Resolves, in passing, which definition of "deductive" is more true — neither; they're exact duals of each other, matched to their own combinator.
- Shortest paper in the series on purpose — it borrows preliminaries from Stops 1 and 2 rather than repeating them.
Correctness Gradient Algebra
Stops 1–3 studied what reasoning operators do to semantic content. This stop asks what they do to a real-valued correctness coordinate laid across the same domain — without ever saying what "correctness" actually measures, since every theorem holds for any such coordinate compatible with the domain's own branching order. The headline result is an asymmetry: pooling a correctness-safe operator with anything else is still safe, no matter what the other operator does; consensus, by contrast, can never increase expected correctness beyond either input, regardless of either one's properties. Also connects idempotent (closure) operators to the classical Knaster–Tarski fixed-point theorem.
- Partially answers a question Stop 1 explicitly reserved for later work.
- Its closure-operator remark becomes its own full section at Stop 5.
Reasoning Dynamics Algebra
Closes the three bench items left open after Stops 1–4. The group G has exactly four normal subgroups in the countable case (cited from the classical Schreier–Ulam–Baer theorem) — giving "finitary reasoning" a precise, conjugation-invariant meaning. Every reasoning trajectory either stabilizes, cycles forever without stabilizing, or never repeats at all — a proven trichotomy, not just an observed involution. And closure operators (from Stop 4's remark) turn out to correspond exactly, and order-reversingly, to finite meet-closed families of semantic states.
- Closes the tour's original stated bench — no candidate topics from that list remain queued.
- Cites classical group theory rather than re-proving it; proves the trajectory and closure results directly from definitions already fixed in earlier stops.
Narrative Operators
The first purely applied stop, and the first with no new abstract theorem in it. Instead of proving new structure, it gives a second concretization of Edwardian Algebra — narrative-canon facts and authorial writing moves, independent of whether the author is a person or an AI — and maps every structure from Stop 5 onto a recognizable narrative-craft phenomenon: the four normal subgroups onto inert prose, bounded local revisions, and fully nonlinear restructuring (with plot-twist parity as a genuinely precise consequence, not a loose analogy); the trajectory trichotomy onto converging revision, revision-thrash, and open-ended serialized/branching narrative; closure operators onto continuity-bible canon-locking.
- The aperiodic trajectory case is identified as a literal instance of this project's own branching story structure, not a metaphor for it.
- Closes with two undeveloped candidate domains for a future applied companion: legal reasoning and game design.
Correctness Geometry Algebra
A direct sequel to Stop 4, not to Stop 6's applied turn. Where Stop 4 studied what a correctness coordinate does under reasoning operators, this stop asks what it means to score an operator itself, and what it means for two operators' effects on correctness to agree, be independent, or contradict each other — correcting a natural but mistaken intuition that orthogonal should mean contradictory, and closing off a competing reading of "0" that turns out to be provably impossible to build on this series' own domain before settling on the one that works. Closes by proving pooling and consensus obey clean unconditional bounds once operators are scored in aggregate, sharpening Stop 4's original one-sided pooling guarantee into a fully symmetric statement.
- Its central negative result — ruling out a plausible-looking construction rather than adopting it — is reported alongside the positive ones, not smoothed over.
- Read this after Stop 4; it cites Stop 4's theorems directly.
Quantum Inferential Algebra
A direct sequel to Stop 2, not to Stop 7's measurement thread — though it cites and extends that thread too. Stop 2 built C on classical Markov kernels and, only in passing, compared the R ⊂ C inclusion to a classical/quantum split in physics — pure commentary at the time, doing no theorem-level work. This stop makes that line literal: a new operator class Q, where invocation resolves a genuine Hilbert-space state via the Born rule rather than an already-fixed classical distribution, with a real, hand-verified example of what changes once amplitudes — not just outcomes — are allowed to vary.
- Most of the series survives the move intact, recovered exactly once every stage of a quantum process is fully measured. What doesn't survive is named precisely: a new combinator with no aggregate bound at all, and a correction to the series' own "Unifying Theorem" on invertibility.
- Neither of the two candidates this page previously left open (legal reasoning, game design) — like Stop 7 was for those same two, this is a third, unlisted direction. Both candidates remain open and unwritten.
Loop Operators
The second purely applied stop, after Stop 6 (Narrative Operators) — no new abstract theorem, just a concretization checked for fit. Where Stop 6 concretized only Stop 1's deterministic algebra, this stop draws on nearly the whole series: the distributivity obstruction of Stop 2 explains, by an already-proven theorem rather than a new argument, exactly where Agile's parallel-decomposition assumption stops holding once a worker is a genuinely non-deterministic coding agent rather than a deterministic human move. Definition-of-Done checklists turn out to already be closure operators (Stop 5); loop health turns out to already be the trajectory trichotomy (Stop 5); permissive versus unanimous code review turn out to already be the Aggregate Pooling and Consensus Bounds (Stop 7); and multi-agent brainstorming versus classical independent parallel tasking turn out to already be coherent versus decohered pooling (Stop 8).
- Two of the five mappings are independently confirmed already running in this very project's own tooling — not merely plausible analogies.
- Neither of Stop 6's still-open named candidates (legal reasoning, game design) is pursued here — this is a fourth, unlisted direction, in the spirit of Stops 7 and 8. Both candidates remain open and unwritten.
Portable Inference Algebra
A new-theorem stop, not a third applied one — though its new theorem falls directly out of a property Stop 2 already built and never named: Kleisli composition is memoryless by construction, so a composite invocation's continuation can never depend on how its current state was reached, only on the state itself. This stop names the missing piece Stop 2 never asked about — whether an intermediate witness can actually be extracted from the process that produced it, cheaply, by anyone else, at any later time — and proves that once it can (a transparent operator, in the paper's own new terminology), free relocation across both machines and elapsed time falls out of the same one theorem as two corollaries, not two separate results.
- Retroactively sharpens Stop 2's own Distributivity Obstruction: that theorem's "share one witness across two branches" was only ever a live comparison under the transparency this stop names but Stop 2 never did.
- Neither of Stop 6's still-open named candidates (legal reasoning, game design) is pursued here — this is a sixth, unlisted direction, in the spirit of Stops 7–9. Both candidates remain open and unwritten.
Concrete Loop Algebra
A new-theorem stop that opens on the question classical abstract algebra begs: if reasoning operators can't form rings or fields, what can they form? The No-Field Theorem closes the first half permanently — assembling obstructions Stops 1 and 2 already proved into one named negative result: inference is an algebra of accumulation, not conservation, and no redefinition repairs that. The second half builds the ladder that does fit: the concrete inferential loop, a DOER/CHECKER pair (δ, γ) — a concrete operator iterated under a closure-operator gate — given two independent semantics (an equational guarded star and a categorical guarded dagger) proven identical by the Agreement Theorem. Non-termination is real mass: the loop is a sub-Markov kernel whose deficit is exactly the probability of never certifying, and Stop 5's trichotomy becomes its complete termination theory.
- Closes with a two-directional negative proven by hand: pooling inside a loop's body and pooling finished loop results can each strictly dominate the other — loops do not pool.
- Released simultaneously with its companion, Stop 12, on the model of Stops 2 and 3 — this stop idealizes the checker as error-free; the companion removes the idealization.
Noisy Gate Algebra
Stop 11's checker never errs. This companion makes the checker what the DOER already was — a concrete inference operator, here a verdict kernel into the two-element subalgebra — and accounts exactly for what its two error modes (false-pass, false-fail) do to a loop. The headline is the amplification law, computed by hand: a gate that is wrong 10% of the time delivers a loop whose output is wrong 19.8% of the time, because iteration hands the gate fresh chances to fail. The two modes are provably asymmetric — only false-pass corrupts output; false-fail wastes work and feeds it. Panels of independent gates, composed with the series' own pooling and consensus, trade the two error modes with strict inequalities rather than reducing both; the first composition that shrinks both — majority of three, Condorcet's own theorem — turns out to require reusing each gate's single verdict across branches, an act licensed by Stop 10's transparency and provably distinct from fresh invocation by Stop 2's Distributivity Obstruction.
- Three stops none of which anticipated the others — the obstruction (Stop 2), transparency (Stop 10), and panel composition (this stop) — meet in one theorem about majority voting.
- Correlated gates, fresh-invocation majority, and optimal panel size are left open, named as such.
What's Next
The original three-item bench (group structure of G, iteration/trajectories, closure operators as their own algebra) closed as of Stop 5. Stop 6 opened an applied direction — concretizations of the existing algebra in domains with no computation in them — and named two undeveloped candidates:
- Legal reasoning and case-based argument (precedent as canon, an opinion as an operator, overturning as an irreversible non-inverse)
- Game design (a rule-change as an operator on the state of play)
Stop 7 didn't pursue either — it opened a third, different direction instead: measurement and comparison, rather than a new domain to concretize the algebra in. Stop 8 didn't pursue either one either — it opened a fourth direction: what invocation itself looks like once the source of non-determinism is genuinely quantum rather than classical. Stop 9 didn't pursue either one either — it opened a fifth direction: applying nearly the whole series, not just Stop 1, to software development itself. Stop 10 didn't pursue either one either — it opened a sixth direction: whether a checkpoint mid-inference can be handed off, across machines and across time, without redoing the work. Stops 11 and 12 — a simultaneous companion pair, on the model of Stops 2 and 3 — didn't pursue either one either: they opened a seventh direction, iteration itself, answering where the ring/field ladder actually ends for inference and what the DOER/CHECKER loop climbs instead, first with a perfect gate and then with a noisy one. Both of Stop 6's candidates remain open and unwritten. When one is, it joins the tour at Stop 13.