And proved it with math.

Varun Pratap Bhardwaj, Garima Singh, Arun Pratap Bhardwaj

→ Two instances of the same model failed together on 90% of missions where either failed — independence is not a safe assumption

Multi-agent reliability analysis routinely assumes component failures are conditionally independent, and almost never tests it. This paper tests it and finds strong positive dependence, which inflates joint failure rates well beyond what independence-based bounds predict. It replaces the assumption with a certificate that assumes no dependence structure at all — a linear program over the joint distribution, bounded by a Bonferroni–Clopper–Pearson box — and proves that fitting a dependence model instead degrades coverage as sample size grows.

Varun Pratap Bhardwaj, Garima Singh, Arun Pratap Bhardwaj

→ A governed, local-first memory control plane — verifiable transactions, scoped memory, auditable operations

Agent memory in production is not a storage problem, it is a governance problem: who wrote this fact, under what scope, and can the write be proven after the fact. SLM 4.0 treats memory as a control plane with verifiable transactions, scoped visibility across profiles, and a full audit trail — while staying entirely on the user's machine.

Varun Pratap Bhardwaj

→ One universal port for AI agents — 12 topologies, POMDP routing, four-tier degradation

Every agent framework reinvents orchestration, and none of them interoperate. Qualixar OS defines a universal substrate — a topology catalogue, POMDP-based routing between agents, and a four-tier degradation ladder that keeps a system answering when parts of it fail.

Varun Pratap Bhardwaj

→ Forgetting as a feature — biologically-inspired decay, cognitive quantization, multi-channel retrieval, zero LLM calls

A memory system that never forgets degrades: old facts crowd out current ones and retrieval quality falls. V3.3 borrows the brain's answer — decay curves, consolidation, and quantized cognitive states — and retrieves across six independent channels fused by reranking, without a single LLM call in the hot path.

Varun Pratap Bhardwaj

→ 74.8% on LoCoMo with zero cloud calls — the highest local-first score reported

Retrieval quality is usually bought with LLM calls. V3 buys it with geometry instead: memories live on a statistical manifold where relevance is a distance, so ranking is computed rather than generated. The result is a benchmark score competitive with cloud systems, produced entirely on-device.

Varun Pratap Bhardwaj

→ Regression testing that treats non-determinism as the subject, not the obstacle — 5–20x cost reduction

You cannot diff an agent's output against a golden file, because the output legitimately differs every run. AgentAssay tests the distribution instead of the sample: stochastic verdicts over repeated runs, with a sampling strategy that reaches a decision in a fraction of the tokens a naive repeat-and-compare would burn.

Varun Pratap Bhardwaj

→ Agent skills execute with your agent's full privilege and nobody checks them — 675 tests, 8 novel contributions

An agent skill is arbitrary code that runs with the agent's own authority, installed from a registry with no review gate. This paper gives the threat model a formal treatment, then builds static analysis, provenance verification and SBOM generation across 22 agent frameworks so a skill can be judged before it executes.

Varun Pratap Bhardwaj

→ Design-by-contract for agents — invariants that hold on every execution path. 1,980 sessions across 7 models

Prompts express intent; they do not enforce it. This paper adapts design-by-contract to autonomous agents: preconditions, postconditions and invariants stated formally, then checked at runtime, so a violation is caught at the step that caused it rather than three tool calls downstream.

Varun Pratap Bhardwaj

→ Shared memory is an attack surface — Bayesian trust scoring that degrades a poisoned writer instead of trusting it

When several agents write to one memory store, a single compromised agent can poison what every other agent later recalls. This paper models writer trust as a Bayesian posterior updated by downstream outcomes, so a source that keeps producing facts which fail verification loses influence over retrieval — without any of the data leaving the machine.