🤖 AI

Determinization in Structure Theories: A Unified Framework via Closure, Comparability, and Joint Admissibility

This paper establishes a unified formal framework for constructing canonical interpretations from plural structure theories by classifying non-determinism into epistemic and structural types and demonstrating how closure stabilization, global completion, and canonical selection mechanisms can achieve determinization under specific structural conditions.

Original authors: Hai Hai Fu

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hai Hai Fu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Detective's Dilemma: When Clues Don't Add Up

Imagine you are a detective trying to solve a mystery. You have a bag of clues (the data) and a rulebook (the theory) that tells you how those clues fit together to form a story. Sometimes, the clues are so clear that the rulebook points to only one single, perfect solution. But often, the clues are messy. Maybe two different suspects could have committed the crime, or maybe the timeline is blurry. In these moments, a bad detective might just guess a solution and say, "It was definitely Suspect A!" without any proof. In the world of Artificial Intelligence, this kind of guessing is called "hallucination."

This paper lives in the intersection of computer science and logic, specifically looking at how AI systems can reason about complex structures without making things up. The core idea is simple: an AI should only give a single, definite answer if the rules of the game actually force that answer to be the only possibility. If the rules allow for multiple valid stories, the AI should admit, "I don't know which one is right yet," rather than picking one at random. The author is building a mathematical toolkit to tell the difference between "we have enough info to be sure" and "we are just guessing."

The Paper: Building a "Truth Filter" for AI

The paper, titled "Determinization in Structure Theories," by Hai Hai Fu, is essentially a manual for building a "Truth Filter" for AI systems. The author is worried about AI systems that act like overconfident detectives, declaring a single answer when the evidence actually supports several different possibilities. They want to create a formal framework that tells an AI exactly when it is allowed to stop guessing and start declaring a winner.

To do this, the author breaks down the problem into two main types of confusion, which they call Type S and Type E.

Type E (Epistemic Plurality) is like a blurry photo.
Imagine you are looking at a picture of a car crash, but it's foggy. You can't tell if the car hit the tree or the fence. However, if you wait for the fog to clear (get more evidence), the picture becomes sharp, and suddenly, only one possibility remains. The paper shows that for these kinds of problems, you can use a "completion" method. Think of this as a machine that keeps adding more details to the story until the fog clears and the picture snaps into focus. The author proves that for certain types of theories (like the "ICT" theory they study), you can build a machine that reliably clears the fog, provided you have enough evidence. However, they are honest about a limitation: they haven't fully proven that this machine always leads to the exact same single picture every time, only that it eventually stops changing. They call this "closure stabilization."

Type S (Structural Plurality) is like a fork in the road.
Now imagine a different scenario. You have a map with two distinct paths, Path A and Path B. Both paths are perfectly valid according to the map's rules. No amount of extra evidence will ever make Path A turn into Path B; they are fundamentally different, like choosing between chocolate or vanilla ice cream. You can't "complete" the story to make them the same. This is what the author calls "Type S." For these problems, trying to use a "completion" machine is a waste of time. Instead, you need a "selector." This is like a referee who looks at the two valid paths and picks one based on a specific, pre-agreed rule (like "always pick the path with the most trees"). The paper proves that for a specific, tricky version of this problem (called "Type S-strong," exemplified by the "Wyckoff" theory), a selector is the only way to get a single answer. If you try to force a completion machine on this, it will fail.

The "Hallucination" Warning
The most important finding of the paper is a warning label for AI developers. The author shows that "hallucination" happens when an AI tries to force a single answer (canonicalization) when the rules don't allow it.

If the problem is Type E, the AI is hallucinating if it picks an answer before the "fog" has fully cleared.
If the problem is Type S, the AI is hallucinating if it tries to use a "completion" machine instead of a "selector."

The paper provides a checklist (a set of mathematical conditions) to see which tool you need. If your system meets the "comparability" and "admissibility" checks, you can build a machine that gives a single, safe answer. If it doesn't, the system is not licensed to pick a winner, and it should stay silent or show all the options.

What They Didn't Solve (Yet)
The author is very careful not to claim they have solved everything. They explicitly state that for the "Type E" problems (the blurry photos), they have built a machine that stops changing (stabilizes), but they haven't proven that it always leads to the exact same unique answer for every single starting point. They call this an "open question." They also admit that for some "Type S" problems that aren't the "strong" version, they aren't sure if a completion machine could work or if a selector is strictly required.

The "Non-Commutative" Twist
Finally, the paper discovers a weird quirk in how these machines work when you stack them. Imagine you have two filters: one that sorts by color and one that sorts by size. If you sort by color first and then size, you get a different result than if you sort by size first and then color. The author proves that for their specific AI theories, the order matters. You can't just swap the steps around; if you do, you might end up with a result that breaks the rules entirely. They found that there is only one specific order (High-Timeframe first, then Low-Timeframe) that keeps the story safe and valid.

In Summary
This paper doesn't just say "AI shouldn't lie." It builds a mathematical map that tells you exactly when an AI is allowed to speak with certainty. It distinguishes between problems that just need more data (Type E) and problems that need a tie-breaking rule (Type S). It warns that using the wrong tool for the job leads to hallucinations, and it proves that for some complex, multi-layered problems, the order in which you apply your rules is critical. While they haven't solved every single puzzle in the universe, they have provided the first rigorous blueprint for knowing when an AI is ready to give a single, truthful answer.

Technical Summary: Determinization in Structure Theories

1. Problem Statement

The paper addresses the formal problem of canonicalization within plural structure theories, specifically motivated by failures in Large Language Model (LLM)-assisted reasoning and decision systems. In these systems, a language model proposes candidate interpretations while a symbolic structure theory defines the set of admissible interpretations.

The core issue is hallucination, defined here as "unsupported canonicalization": the system emitting a determinate answer when the underlying set of admissible interpretations has not collapsed to a singleton. The paper asks: Under what structural conditions is a structure theory licensed to output a unique interpretation?

Most market structure frameworks (and by extension, many formal theories) are non-deterministic, allowing multiple valid interpretations to coexist at the same position. The paper seeks to formalize sufficient structural conditions under which a canonical reading can be constructed and to classify the mechanisms by which this occurs.

2. Methodology and Framework

2.1 Structure Theories

The author formalizes a Structure Theory as a triple $T = (\Sigma, A, I)$ , where:

$\Sigma$ is a signature.
$A$ is a set of axioms.
$I = I_{core} \cup I_{ext}$ is an inference policy (the paper focuses on $I_{core}$ ).

The theory operates on inputs $(N, \Theta, \Pi)$ representing market realizations, parameters, and structural evidence. The admissible interpretation family $P_T$ is the set of all assignments consistent with axioms $A$ and the inference policy.

Constraints on $I_{core}$ :
The inference policy is restricted to positive, non-retractive inference:

(R1) Positive Rule Form: Rules only add conclusions; none remove them.
(R2) No Negation-as-Failure: Premises are positive conjunctions; no rule relies on the absence of a derived predicate.
(R3) No Retraction: No rule withdraws previously derived conclusions.

2.2 Taxonomy of Non-Determinism

The paper classifies non-determinism into two types based on how plurality behaves under evidence refinement:

Type E (Epistemic Plurality): Plurality vanishes as evidence is refined. If $|P_T| > 1$ , there exists a refined evidence set $\Pi'$ such that $|P_T(\Pi')| = 1$ . (Example: ICT theory).
Type S (Structural Plurality): Plurality persists even at maximal evidence ( $\Pi_{full}$ $Π_{f u l l}$ ).
- Type S-strong: A subclass of Type S where distinct admissible interpretations at $\Pi_{full}$ have no common upper bound under the specification order $\preceq_{spec}$ . (Example: Wyckoff theory).

2.3 Canonicalization Strengths

The framework distinguishes three levels of canonicalization:

Closure Stabilization (Property 4'): Every admissible seed converges under an operator to some fixed point (seed-dependent).
Global Completion (Property 4): All seeds converge to a single unique fixed point.
Determinization (AC-6): The unique fixed point is the sole admissible interpretation.

3. Key Contributions and Results

3.1 Dual Mechanisms for Canonicalization

The paper establishes that canonicalization is not a monolithic operator-theoretic process but splits into two fundamentally different forms:

Operator-Based Completion (Type E):
- Applicable to Type E theories.
- Uses a completion operator ( $D_{comp}$ ) that iteratively adds conclusions until a fixed point is reached.
- Requires: Closure, comparability (N4-comp), and compatible extension (C5-E).
- Result: Establishes Closure Stabilization (4'). Achieving Global Completion (4) or AC-6 requires an additional global confluence result (Open Question OQ-GC-1 for ICT).
Selector-Based Construction (Type S-strong):
- Applicable to Type S-strong theories.
- Uses a canonical selector ( $\phi$ ) that picks a unique element from local sets based on a theory-intrinsic order.
- Requires: Theory-intrinsic comparability (N4-sel) and global compatibility of selected choices (C5-S).
- Result: Directly achieves AC-6.
- Impossibility Result: For Type S-strong theories, Global Completion (Property 4) via an extensive operator is provably impossible because the incomparable interpretations lack a common upper bound.

3.2 Construction Lemma (Theorem 6b)

The paper provides a construction lemma establishing sufficient conditions for the existence of these mechanisms:

Completion Branch: $N0 + N3 + N4\text{-comp} + C5\text{-E} \implies D_{comp}$ exists at strength (4').
Selection Branch: $N0 + N1 + N4\text{-sel} + C5\text{-S} \implies PhaseClassify$ exists at AC-6.

A critical finding is that theory-intrinsic comparability (N4) is necessary. A minimal counterexample ( $T_{ce}$ ) demonstrates that closure alone (without comparability) is insufficient for canonicalization.

3.3 Classification of Mechanisms (Theorem U3)

Conditional on a realization-coverage assumption (OQ-Realization), the paper classifies every theory-intrinsic mechanism realized within a specific grammar into:

Class C (Completion): Realized via saturated closure of elementary rules (3a) or guarded saturated closure (3c). These satisfy properties (1)–(4').
Class S (Selection): Realized via axiom-induced selection (3b).

Key Distinction on Monotonicity:

Pure $I_{core}$ -computable mechanisms (Class C, 3a) are monotone under the specification order $\preceq_{spec}$ (Lemma C).
Guarded mechanisms (Class C, 3c) are generically non-monotone. Adding a predicate can activate a guard that suppresses a derivation, blocking a conclusion that would have fired in a smaller state. The paper provides a concrete witness ( $D_{comp}^{HTF}$ ) where monotonicity fails.

3.4 Multi-Timeframe and Staged Operators

The paper extends the framework to multi-timeframe hierarchies using a staged operator $D_{comp}^{HTF} = D_L \circ D_H$ .

Non-Commutativity: The paper proves that staged completion is structurally non-commutative. Specifically, for the raw two-operator architecture, the ordering $D_L \circ D_H$ (applying $D_H$ [HTF] first, then $D_L$ [LTF]) is the only admissibility-preserving linearization. The reverse order can produce inadmissible states.
Phase 2 (Multi-Operator): The paper summarizes existence conditions for multi-operator canonicalization, defining a "safety gap" $\Phi^*$ . Feasibility is equivalent to $\Phi^* = 0$ . It establishes a hierarchy of compatibility conditions: $GCC \implies RUCS \implies PLCS$ .

3.5 Categorical Structure

The paper defines Determinization as a partial pseudofunctor ($Det$) mapping pairs of theories and comparability structures to their canonical extensions. This provides a categorical template for recursive theories (e.g., Elliott Wave).

4. Significance and Claims

The paper claims to provide a unified framework for understanding when and how structural theories can be made deterministic, distinguishing between:

Epistemic uncertainty (resolvable by evidence accumulation via completion).
Structural incomparability (requiring a selection mechanism).

Modest Claims and Open Questions:

The paper does not establish that ICT (Type E) achieves AC-6 determinization; it only establishes Closure Stabilization (4'). Global completion for ICT remains an open question (OQ-GC-1).
The classification of mechanisms (Theorem U3) is conditional on the assumption that all theory-intrinsic operations fall within the (3a)/(3b)/(3c) grammar (OQ-Realization).
The "Type S $\to$ Selection" inference is proven only for Type S-strong theories. The generic Type S case remains open (OQ-TypeS-Imp).
The non-commutativity result is scoped to the specific "two-operator raw staged architecture"; other architectures with repair operators or joint fixed-points are not excluded.

Conclusion:
The primary contribution is the formal separation of closure-based and selection-based canonicalization, the identification of non-monotonicity in guarded systems, and the rigorous definition of the structural conditions (N4, C5) required to license a unique output. This framework offers a formal basis for diagnosing "hallucination" in LLM-assisted systems as the emission of a determinate answer when the underlying admissible set has not collapsed to a singleton.