RESEARCH

We ship what we research.
And teach what we ship.

The papers, the open source, and the playbooks behind everything Cohorte builds and teaches. Every claim we make about trust and verification traces to something on this page.

Read the papers
Why this exists

The work that grounds the practice.

Cohorte publishes its methods openly: black-box reliability certification for systems you cannot see inside, a taxonomy of what makes coding agents seek a planted backdoor, and the orchestration architecture behind the open-source stack. Plausible and correct are not the same thing.

Interactive

A reliability number for a black box.

TrustGate turns a model output into a single number: at what confidence can a practitioner trust this system on this task. Pick a model and a task. Slide the target confidence. Watch the conformal cutoff move.

No model call. Every number is published.

Papers

Open research.

Black-Box Reliability Certification for AI Agents via Self-Consistency Sampling and Conformal Calibration

Charafeddine Mouzouni · Under review, TMLR (double-blind) · 2026

Given a black-box AI system and a task, at what confidence can a practitioner trust its output? We answer with a single reliability level per (system, task) pair, derived from self-consistency sampling and conformal calibration, that acts as a deployment gate with exact, finite-sample, distribution-free guarantees.

Result: GPT-4.1 reaches 94.6% on GSM8K and 96.8% on TruthfulQA. Sequential stopping cuts API cost by about 50% without losing the guarantee.
StatusUnder review at TMLR
ImplementationTrustGate (open source)
TrustGate repo Preprint by email

Mapping the Exploitation Surface: A 10,000-Trial Taxonomy of What Makes LLM Agents Exploit Vulnerabilities

Charafeddine Mouzouni · OPIT and Cohorte AI · arXiv preprint · 2026

Which features of a system prompt send a coding agent looking for a planted backdoor? We test 37 prompt conditions across 12 psychological dimensions on 7 models in real Docker sandboxes. Most hypothesized dimensions produce no detectable effect at 50 trials per cell. One works reliably: goal reframing (puzzle and CTF framing).

Result: On Claude Sonnet 4, puzzle framing sends the agent to a planted backdoor in 38–40% of trials, despite an explicit safety instruction.
StatusarXiv preprint, 2026
Code & dataPublic repository
Paper repository

Context Kubernetes: An Orchestration Architecture for Enterprise Knowledge in Agentic AI Systems

Charafeddine Mouzouni · arXiv preprint · 2026

Delivering the right knowledge, to the right agent, with the right permissions, at the right freshness, within the right cost envelope, across an organization, is structurally the container-orchestration problem Kubernetes solved a decade ago. We introduce a declarative manifest, a reconciliation loop, and a three-tier permission model where agent authority is always a strict subset of human authority.

Result: Prototype ~7,000 lines, 92 tests, 8 experiments. A negative result: in the repository's final baseline (200 queries), a simple ACL-filtered RAG leaked 0 times against 307 for the full architecture. The repository is archived as a citation.
StatusarXiv preprint, 2026
Reference implementationContext Kubernetes (open source)
Reference repo

Three Phases of Expert Routing: How Load Balance Evolves During Mixture-of-Experts Training

Charafeddine Mouzouni · arXiv preprint · April 2026

We model MoE token routing as a congestion game and track its effective congestion across training. The trajectory reveals three phases: a surge where the router learns to balance load, a stabilization where experts specialize, and a relaxation where the router trades balance for quality. This non-monotone trajectory is invisible to post-hoc analysis of converged models.

Result: Studied across OLMoE-1B-7B (20 checkpoints) and OpenMoE-8B (6 checkpoints), with bootstrap confidence intervals on every estimate.
StatusarXiv preprint, April 2026
Code & dataPublic repository
Paper repository
Open source

Reference implementations.

trustgate

Reliability and trust layer for agent outputs. Self-consistency sampling, conformal calibration, the reliability-level method from the paper under review at TMLR.

Cohorte-ai/trustgate

guardrails

Policy, safety, and compliance guardrails for agents. Input filtering, output validation.

Cohorte-ai/guardrails

context-router

Smart routing of queries to the right context and data source. The retrieval layer with permission enforcement.

Cohorte-ai/context-router

agent-monitor

Observability, logging, and behavioral monitoring for agents in production. Drift detection, runtime signals, regulator-facing reporting.

Cohorte-ai/agent-monitor

agent-auth

Permissions, roles, and authentication for agent actions. The three-tier model where agent authority is a strict subset of human authority.

Cohorte-ai/agent-auth

context-kubernetes

Archived research prototype accompanying the arXiv paper: declarative manifest and reconciliation loop. Kept public as a citation and as the record of a negative result.

Cohorte-ai/context-kubernetes
Playbooks

Between the papers and the practice.

Build playbook · owning the stack

Building a Proprietary Agentic Platform

A hands-on build: an interface, a worker, a model gateway and a persistent workspace, assembled around one modest maintenance task. Permissions checked before the model sees data, loop limits, documents that cannot authorize actions, and a private model swapped in behind a gateway. 22 chapters, three appendices and a runnable reference stack. The application is a teaching lab; container, GPU and observability runs were not executed.

Free · cohorte.co/playbooks
Market map · verification tooling

The Agent Verification Map

A documentation study of fifteen agent evaluation and verification products, placed on two axes: what each one examines, and whether it reads test cases or records from the running application. Seven measurement types, a risk grid, eight shared fixtures for a buying test, and an attached proof kit of eighteen synthetic evidence packets that runs locally. Product performance was not independently measured.

Free · cohorte.co/playbooks
Field guide · the harness

The Harness Engineering Playbook

The runtime that decides what a model's proposals become: context, tools, permissions, state, coordination, verification, and controlled evolution. Eight parts, twenty-one chapters, a recurring failure-injected case, and twenty-two copyable field artifacts. 640 pages for architects, platform leads, and risk owners.

Free · cohorte.co/playbooks
Field guide · evaluation

The Agent Eval Playbook

How to evaluate LLM agents and build reliable ones in production: the methods top labs actually use, with the math, the mechanisms, and the runnable parts. Seven parts, twenty chapters, three appendices.

Free · cohorte.co/playbooks
Playbook · the platform

The Enterprise Agentic Platform

A four-layer reference architecture for putting agents into a serious organization: local agents on each machine, a governance middleware that mediates every organisational call, a permission model where agent authority is a strict subset of human authority, and a reliability layer that proves the output. Ten chapters, four appendices, written for architects and CISOs.

Free · cohorte.co/playbooks
Playbook · accountability

From Autonomous Agents to Accountable Systems

Why autonomy is the wrong target, why accountability is the right one, and how to design a system whose outputs a human can sign for under audit.

Free · cohorte.co/playbooks
Put it to work

Bring the method to your team.

The papers are public, the code is open. The harder work is installing reliability and accountability inside a team that already exists. That is what we do.

Request a workshop
Or see how the research becomes practice.