Paradox — AI Reasoning Lab

Research / 2025

Reasoning systems that show their work.

Paradox is an AI reasoning lab building transparent models that can explain every step of their inference — no black boxes, no hidden chains, just verifiable thought.

Read the research

Six directions, one question.

How do we build systems that reason in ways humans can audit? Our research program spans interpretability, formal verification, and the architecture of deliberation itself.

01 / Interpretability

Mechanistic transparency

Reverse-engineering the internal representations of large models to produce human-readable traces of every inference step. We map circuits, not vibes.

02 / Verification

Formal proof synthesis

Training models to generate machine-checkable proofs alongside their answers. Every claim arrives with its own certificate of correctness.

03 / Alignment

Scalable oversight

Methods for supervising systems that exceed human-level capability in narrow domains. Debate, recursion, and recursive reward modeling.

04 / Architecture

Deliberative reasoning

Novel architectures that separate generation from critique — models that draft, challenge, and revise before they speak. Slow thinking, by design.

05 / Safety

Deception detection

Probing techniques that detect when a model's stated reasoning diverges from its actual computational path. Honesty as a measurable property.

06 / Evaluation

Reasoning benchmarks

Open benchmarks that test multi-step inference under adversarial pressure. We publish the failures, not just the scores.

Paradox-1 — a reasoning model you can audit.

P-1

Paradox-1 / 340B / Audit Edition

A model that writes its own audit trail.

Paradox-1 is a 340-billion-parameter reasoning model trained to produce a verifiable chain-of-thought alongside every output. Each step is annotated with its evidence, its confidence, and the circuit trace that produced it.

It does not just answer. It shows you how it arrived — and flags the steps it is least sure about.

  • Parameters 340B (active 68B, MoE)
  • Context 256K tokens
  • Audit trace Native, per-step
  • Proof pass rate 94.2% (FormalBench)
  • Release Q3 2025 — API preview

A small group, on a long problem.

Twenty-three researchers across ML, formal methods, and philosophy. We publish openly, collaborate broadly, and move deliberately.

Dr. Lena Voss

Dr. Lena Voss

Chief Scientist

Dr. Arjun Mehta

Dr. Arjun Mehta

Head of Interpretability

Dr. Sara Lindqvist

Dr. Sara Lindqvist

Head of Formal Methods

Marcus Cole

Marcus Cole

Head of Safety

Recent research, in full.

We publish our methods, our data, and our failures. Every paper is accompanied by reproducible code and audit logs.

Get in touch

Build with verifiable reasoning.

We are opening a limited API preview for research institutions and safety teams. Tell us what you want to inspect.

≡ hart

Hosted on hart

intrane-landings/43-paradoxpublic
🏠 hart home🔎 Explore public👤 More from intrane-landings