Star on GitHub ★
Academic Repository · ORCID: 0009-0002-8726-4289

Research & Foundational Papers

Agent Behavioral Contracts is rooted in stochastic process theory, formal methods, and non-parametric convex optimization. Explore the complete academic publications, mathematical proofs, and archival datasets below.

ABC I · FEBRUARY 2026 arXiv:2602.22302 [cs.AI]

Agent Behavioral Contracts: Formal Specification and Runtime Enforcement for Reliable Autonomous AI Agents

Varun Pratap Bhardwaj · Independent Researcher · Qualixar

Abstract: Autonomous AI agents operate without formal behavioral guarantees, leading to compounding stochastic drift, regulatory violations, and catastrophic failures in critical domains. We introduce Agent Behavioral Contracts (ABC), a formal runtime enforcement framework combining design-by-contract principles with stochastic process theory.

We prove the Drift Bounds Theorem: for any agent with stochastic perturbation rate α and contract recovery rate γ > α, the expected behavioral drift is bounded by D = α / γ. We define (p, δ, k)-satisfaction to quantify non-deterministic reliability and demonstrate Θ = 0.9541 aggregate compliance across 200 benchmark scenarios on 7 frontier LLMs with < 10 ms per-action latency.

Length: 71 pages
Theorems: 4 formal proofs
Sessions: 1,980 evaluated
DOI: 10.5281/zenodo.18775393
ABC I Drift Bounds Theorem Diagram
Figure 1 · Drift Bounds Theorem Geometric Proof Shows the Lyapunov stability basin and stochastic perturbation convergence to invariant zone D = α/γ under contract recovery force.
ABC II · AUGUST 2026 · NEW RELEASE arXiv:2608.12895 [cs.AI]

Certifying Compositional Reliability Without Assuming Independence

Varun Pratap Bhardwaj · Independent Researcher · Qualixar

Abstract: Composing multi-agent workflows routinely relies on multiplying component reliabilities (RA · RB), an operation valid only under independence. In an 18,000 preregistered mission study evaluated by deterministic test oracles with zero LLM judges, we discover that paired agents from the same model family co-fail on 90.0% of missions where either fails (log OR = 6.66, φ = 0.916).

We prove that fitting parametric dependence models and bootstrapping confidence intervals creates a catastrophic failure mode: the identification gap $O(1)$ dominates the $O(n^-0.5)$ haircut, causing intervals to shrink around false parameters. We establish a non-parametric linear programming certificate over empirical moment functionals $M_10 \dots M_14$ that makes zero distributional assumptions, narrowing uncertainty bounds by 85.7% while guaranteeing anytime-valid Type-I error ($\alpha \le 0.0471$).

Length: 49 pages
Theorems: 18 formal theorems
Definitions: 25 definitions
Missions: 18,000 preregistered
ABC II Linear Programming Polytope Diagram
Figure 2 · Convex Moment Polytope Relaxation Demonstrates non-parametric linear programming certification over moment functionals M₁₀ ... M₁₄, achieving an 85.7% interval reduction without parametric assumptions.

The Qualixar AI Reliability Research Suite (9 Papers)

AgentAssert is part of Qualixar's broader 9-paper formal AI reliability research initiative authored by Varun Pratap Bhardwaj.

arXiv:2608.08253 · NEW
SuperLocalMemory 4.0

The Governed Memory Operating System for multi-agent autonomous architectures with policy boundaries.

Read arXiv:2608.08253 ↗
arXiv:2604.06392
Qualixar OS (QOS)

A Universal Operating System for AI Agent Orchestration with Type-C protocol and resource sandboxing.

Read arXiv:2604.06392 ↗
arXiv:2603.00195
SkillFortify

Formal analysis and supply chain security framework for autonomous agentic tool and skill ASTs.

Read arXiv:2603.00195 ↗
arXiv:2603.02601
AgentAssay

Token-efficient regression testing framework with deterministic test oracles for non-deterministic AI.

Read arXiv:2603.02601 ↗
arXiv:2604.04514
SuperLocalMemory V3.3

The Living Brain: Biologically-inspired forgetting, cognitive quantization, and multi-channel retrieval.

Read arXiv:2604.04514 ↗
arXiv:2603.02240
SuperLocalMemory V1

Privacy-preserving multi-agent memory with Bayesian trust defense against adversarial memory poisoning.

Read arXiv:2603.02240 ↗