🤖 AI

Local verification cannot detect non-transportability: a cohomological theory of context preservation in agentic reasoning

This paper introduces a cohomological framework demonstrating that local verification safeguards in agentic AI are structurally incapable of detecting non-transportable conclusions arising from harmonic evidence conflicts, and proposes the Ksetra procedure to identify and gate such global inconsistencies using cycle-based statistics.

Original authors: Suyash Mishra

Published 2026-08-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Suyash Mishra

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 Mapmaker's Dilemma: Why Getting Lost Isn't Always Your Fault

Imagine you are trying to navigate a massive, unfamiliar city using a set of small, overlapping street maps. You have a map for the downtown district, another for the harbor, and a third for the hills. To get from your hotel to the museum, you might hop from the downtown map to the harbor map, and then to the hill map. This is how modern AI agents work when they try to solve complex problems: they chain together bits of information from different sources, like a doctor moving from a lab result to a patient's history, or a bank moving from a local credit score to a global economic trend.

For a long time, scientists thought the only way to make sure this journey was safe was to check every single step. You'd verify that the downtown map matches the harbor map at the border, and that the harbor map matches the hill map. If every local connection looked good, you assumed the whole trip was safe. But there's a catch: sometimes, even if every single border looks perfect, the city itself might be shaped in a way that makes your final destination depend on which route you took. It's like walking in a circle on a Möbius strip; you might end up on the "other side" of the world without ever realizing you crossed a boundary. This paper asks a scary question: What if our best safety checks are blind to this kind of hidden, structural confusion?

The Paper: When "Local Checks" Miss the Big Picture

This paper, written by AI researcher Suyash Mishra, tackles a problem that sounds like a math puzzle but is actually about how AI agents make decisions in the real world. The author argues that the current way we verify AI reasoning is fundamentally incomplete. We check the "local" connections (does this piece of evidence fit with that one?), but we miss the "global" shape of the problem.

The Core Discovery: The Invisible Loop
The paper proves that an AI can pass every single local safety check and still arrive at the wrong answer, simply because it took a different path through the evidence. The author uses a branch of mathematics called cohomology (think of it as the study of holes and loops in shapes) to show that evidence can have a "twist" in it.

Imagine you are walking around a park. You check the fence between the grass and the flowers, then the fence between the flowers and the trees, and finally the fence between the trees and the grass. Every fence looks solid. But if the park is built on a giant, invisible loop (like a donut shape), you might end up back where you started but slightly shifted, or on a different "layer" of reality. The paper calls this holonomy. It's a structural glitch where the evidence doesn't quite close the loop, even though every individual piece of evidence looks fine.

What the Paper Rules Out
The author is very clear about what doesn't work. They prove that any verification system that only looks at small, local pieces (like checking one pair of maps at a time) is structurally blind to this problem. No matter how carefully you check the borders, you cannot detect this "twist" if you only look at the borders. The paper explicitly rejects the idea that better local checks or more consensus among AI panels will fix this. If the twist exists, a panel of experts arguing about the path will just be arguing about the same invisible loop; they won't find the hole.

The Three Types of Confusion
The paper breaks down evidence conflicts into three distinct types, using a mathematical tool called Hodge decomposition (which is like sorting a messy pile of laundry into three separate baskets):

The Gradient (Calibration): This is just a simple offset. Maybe one map says "North" is up, and the other says "North" is slightly tilted. This is easy to fix; you just recalibrate the compass.
The Curl (Local Inconsistency): This is a mess in a small area. Maybe the flower fence doesn't match the tree fence. This is detectable if you look at a small group of three maps together.
The Harmonic (The Invisible Twist): This is the big one. It's a conflict that exists only when you look at the whole loop. It passes every local test but makes the final answer depend on the route taken. This is the part that current AI safety checks cannot see.

The Solution: Ks.etra
To fix this, the author proposes a new method called Ks.etra (pronounced "K-setra"). Instead of just checking if the pieces fit, Ks.etra calculates the "harmonic energy" of the evidence network. If this energy is high, it means there is a structural twist that cannot be resolved by just gathering more data at the same level.

The paper suggests that when an AI detects this twist, it shouldn't just guess or try harder. It should abstain (refuse to answer) and, more importantly, tell the human operator exactly where to look. It acts like a signpost saying, "The loop doesn't close here; you need to split this group of people into two smaller groups to fix the map."

What the Simulations Show
The author tested this idea in two simulated worlds: one involving drug discovery (Pharma) and one involving credit scores (Credit).

In the Pharma simulation, using Ks.etra reduced the rate of harmful decisions by 0.032 (about 3.2%) compared to the best existing method.
In the Credit simulation, the improvement was 0.043 (about 4.3%).
The paper also found that the "twist" (harmonic energy) was a strong predictor of errors that couldn't be fixed, with a correlation of 0.37 in the simulations.

How Sure Are We?
It is important to note that these results come from simulations, not real-world data yet. The author is very honest about this. They built a mathematical model where the "twist" was generated by a specific mechanism (called effect modification, which is like a hidden variable changing the rules of the game). In these simulations, the math held up perfectly. They even created a statistical test (an F-test) that can detect if a global, consistent answer exists, and it worked well in their tests.

However, the paper concludes by saying that the next step is to test this on real data. They suggest looking at existing medical studies or financial records to see if the "twist" predicts real-world disagreements. Until then, the idea that "local checks are blind to global loops" remains a powerful theoretical insight supported by computer simulations, but not yet a proven fact of the real world.

The Takeaway
The paper's main message is a warning for the future of AI: You can't just check the steps; you have to check the shape of the journey. If an AI is trying to transport a conclusion from one context to another, and the "shape" of the contexts has a hidden loop, the AI might be confidently wrong. The solution isn't to argue more; it's to recognize the loop, stop, and redraw the map.

Technical Summary: Local Verification Cannot Detect Non-Transportability

Problem Statement

Agentic AI systems increasingly transport conclusions across diverse biological, clinical, and commercial contexts. Current state-of-the-art architectures (e.g., Medea) rely on local verification: checking at each step that entities are representable, parameters are compatible, and outputs cohere with the plan. These checks are "local" in that they are supported on bounded pieces of context space (single contexts, pairs, or triples).

The paper identifies a structural incompleteness in this approach. While local checks can detect specific types of errors, they are blind to structural non-transportability. This manifests as a situation where an agent passes every local verification check yet produces a conclusion that depends on the specific reasoning path taken. This phenomenon is formally analogous to "loop inconsistency" in network meta-analysis, where direct and indirect evidence around a closed loop are individually sound but jointly incoherent.

Methodology

Theoretical Framework

The author models the problem using algebraic topology and sheaf theory:

Context Site: The space of decision-relevant situations $X$ is covered by a finite set of contexts $\mathcal{U} = \{U_i\}$ . The nerve $N(\mathcal{U})$ is constructed where vertices are contexts, edges represent overlaps (bridging evidence), and triangles represent triple overlaps.
Evidence Cochain: Evidence is modeled as a real-valued 1-cochain $\omega \in C^1(N(\mathcal{U}); \mathbb{R})$ , assigning a measured contrast to each overlap.
Agentic Transport as Path Integration: An agent transporting a claim from anchor $a$ to target $t$ performs path integration: $\hat{\theta}_t = \theta_a + \langle \omega, \gamma \rangle$ , where $\gamma$ is the path in the nerve.
Hodge Decomposition: The space of 1-cochains is orthogonally decomposed into three components:
- Gradient ( $g = \delta_0 \beta$ ): Per-context instrument offsets. Removable by recalibration.
- Curl ( $c \in \text{im}(\delta_1)^\top$ ): Local inconsistencies detectable at triple overlaps (non-zero $\delta_1 c$ ).
- Harmonic ( $h \in H^1$ ): The component in the kernel of both $\delta_1$ and $\delta_0^\top$ . This represents the first Čech cohomology class $[\omega] \in H^1$ .

Key Theoretical Results

Path Independence: The transported estimate is path-independent if and only if the evidence cochain is exact ( $\omega \in \text{im } \delta_0$ ).
Holonomy: The disagreement between two reasoning paths is exactly the holonomy of the cohomology class $[\omega]$ .
Blindness of Local Verification: Any consistency check supported on simplices (edges, triangles) cannot distinguish between an exact cochain and one perturbed by a harmonic component. Since harmonic components satisfy $\delta_1 h = 0$ , they pass all triple-overlap checks, yet they generate non-zero disagreement between paths.
Mechanism of Obstruction: The harmonic component arises from effect modification combined with overlap-specific population composition (Simpson's paradox on a network). If effect modification is absent, the harmonic energy vanishes to machine precision.

Proposed Method: Ks.etra

The author proposes Ks.etra (Ksetra), a procedure that:

Constructs the context site and assembles the evidence cochain.
Decomposes $\omega$ into gradient, curl, and harmonic parts via least squares.
Triage:
- Dominant Gradient $\to$ Recalibrate instruments.
- Dominant Curl $\to$ Re-measure or re-run deliberation (local inconsistency).
- Dominant Harmonic $\to$ Structural obstruction; the covering must be refined (stratification).
Gating: The system abstains from making a decision if the harmonic energy $\|\omega_h\|$ exceeds a threshold. Upon abstention, it outputs the specific cycle carrying the largest holonomy as a data-collection instruction.

Results

Simulation Studies

The author conducted simulations across two domains:

Pharma RWE: Contexts defined by (CKD stage $\times$ care setting). The attribute space is contractible; $H^1$ arises only from evidence gaps.
Consumer Credit: Contexts defined by (region $\times$ macroeconomic regime). The regime axis is cyclic (expansion $\to$ recovery), creating a topological obstruction by construction.

Key Findings:

Predictive Power: In the mechanistic model where holonomy emerges from effect modification, both harmonic and curl energies are comparably predictive of the irreducible error of the optimal estimator (Spearman $\rho \approx 0.37$ for harmonic and $0.33$ for curl). This correlation arises because a single latent cause (effect modification) generates both components. However, harmonic energy retains independent predictive value after conditioning on the other components (partial $\rho \approx 0.19$ ), whereas curl does not. The operational distinction between the two is defined by their remedy (re-measurement for curl vs. refinement for harmonic), not solely by their predictive value.
Path Dependence: Both harmonic and curl components predict path-dependence (spread across reasoning paths), but only harmonic predicts the error floor independently of the other components.
Performance: Ks.etra (gating on harmonic energy) outperforms a "Residual-gated" baseline (gating on total conflict) in the Area Under the Risk-Coverage Curve (AURC), reducing it by 0.029 (pharma) and 0.039 (credit).
Drift Cost: "Silent coarsening" (collapsing contexts without verification) incurs significant error even in perfect systems.

Finance Test Case (FX and IFRS 9)

Exact F-Test: The author derives an exact F-test for the existence of a global claim, partitioning degrees of freedom into calibration, local coherence, and transport. The test is correctly sized under isotropic noise.
FX Arbitrage: In a constructed 15-currency complex with executability constraints, the harmonic component detects "loop arbitrage" (P&L on cycles not filled by executable triangles) that standard triangular arbitrage checks (curl) miss.
IFRS 9: Applied to a PD model inventory over a cyclic regime space, the method identifies 17 independent loops capable of carrying incoherent stories undetectable by pairwise or triple validation.

Significance and Claims

The paper makes a negative result as its central contribution: No family of simplex-supported consistency checks can detect the harmonic obstruction. This implies that current agentic verification strategies, regardless of implementation quality, have a fundamental blind spot.

Significance:

Theoretical: It identifies loop inconsistency in evidence synthesis as the first cohomology of a context site, extending the concept from treatment networks to arbitrary context coverings.
Operational: It reframes abstention from a heuristic confidence score to a cohomological quantity. A non-zero harmonic class is not just a warning; it is a specific instruction to refine the covering (stratify further) rather than gather more evidence at the same granularity. The distinction between curl and harmonic components is defined by their remedy (re-measurement vs. refinement), not solely by their predictive value.
Economic: The method offers a transparent mapping from selective-risk geometry to portfolio outcomes, showing that cohomological gating is economically viable only where the cost of a wrong commitment materially exceeds the gain of a correct one (e.g., regulated therapeutic and credit decisions).

Limitations and Modesty:

The paper explicitly states it contains no real-data validation. The simulations are controlled validations of the theory.
The mechanism for harmonic generation (effect modification) is modeled, not observed in real-world agent traces.
The author acknowledges that correlated errors across overlaps and the elicitation of the correct context factorisation are open problems not solved here.
The contribution is framed as a theoretical identification of a blind spot and a proposed remedy, with the call for real-data validation (e.g., using published network meta-analyses or Medea's yeast benchmarks) as the necessary next step to move results off simulation.