Apodex

Meet 𝗔𝗽𝗼𝗱𝗲𝘅 𝟭.𝟬 🔭 — a heavy-duty agent team for deep research, which sets the SOTA! The team searches the web, reasons over evidence, and writes reports where every claim is backed by an explicit 𝘦𝘷𝘪𝘥𝘦𝘯𝘤𝘦 𝘤𝘩𝘢𝘪𝘯, independently audited before delivery. 🌐 apodex.ai

Predicting the future takes deep research. Great to see Apodex-1.0-mini rank first again on FutureX：futurex.live Try it here: apodex.ai

As a self‑evolving heavy‑duty solver, the Apodex family reliably converts cryptic clues into a verifiable identifier rather than a plausible‑sounding fabrication. If you work in a high-stakes domain, be our eval and throw your hardest problems at Apodex! #AIforScience

Here’s how six frontier models performed: −4: GPT‑5.4 (Zytiga), Qwen3.6‑Plus (invented “Zurnai”), GLM‑5.1 (none), MiniMax‑M2.7 (none) +4: Claude‑Opus‑4.7, Gemini‑3.1‑Pro Two baselines matched the ground truth on all four fields. The other four finished at −4 through confident mistakes, hallucination, or no answer. Apodex matches frontier-level accuracy with a 35B model and an architecture that exposes the evidence and exclusion logic you can audit. 🤏 🦾

Apodex-1.0-mini, our 35B model, currently holds the #1 spot on the FutureX. But leaderboards are sanitized: clean queries, no penalty for confident wrong answers. So we wrote one deliberately brutal question: four facts to recover, a scoring rule where confident but wrong = −4, correct = +4. Between frontier models and Apodex, here’s what happened. 🧵👇

@csaba_kissi A research team you can spin up on demand.

@Apodex_AI It's like having an entire department of researchers.

Meet 𝗔𝗽𝗼𝗱𝗲𝘅 𝟭.𝟬 🔭 — a heavy-duty agent team for deep research, which sets the SOTA! The team searches the web, reasons over evidence, and writes reports where every claim is backed by an explicit 𝘦𝘷𝘪𝘥𝘦𝘯𝘤𝘦 𝘤𝘩𝘢𝘪𝘯, independently audited before delivery. 🌐 apodex.ai

Apodex-1.0 @Apodex_AI just dropped! Self-evolving heavy-duty solver. 150 agents across 15,000 steps. Independent verification on every attempt. SOTA on BrowseComp, FrontierScience, SuperChem. Open weights available. Check it out: apodex.ai

Apodex-1.0 @Apodex_AI changed deep-research AI with a self-evolving heavy-duty solver. It coordinates 150 agents across 15,000 steps. Verifies through independent auditors. Improves each attempt by critiquing the last. Built for research problems that have no existing answer. A step-by-step guide 🧵 👇

This Delaware scenario is just one instance. Law, life sciences, clinical Medicine, finance, compliance, critical engineering: the cost of hallucination is real. We built Apodex for a simple reason: deep research agents should not only find answers, but prove why those answers are right. Link in bio. Give it a spin and tell us how it went.🦾 #LegalTech

Apodex is a Self‑Evolving Heavy‑Duty Solver built exactly for this: • Retrieval + verification agents run in parallel, routing everything through a shared evidence pool • A global verifier reasons over the assembled evidence graph, weighing the support behind each claim • A verification gate sends the answer back whenever citation coverage or evidence support falls short →Every claim in the final report traces back to a source you can open. Full audit trail.

In Mississippi, a federal judge just threw all 4 lawyers — both sides — off a case for “blindly relying on technology.” In Oregon, one team submitted AI‑hallucinated case citations. Penalty: $109,700 + the case got tossed. 🧵👇

@legalwritingpro In high‑risk domains, AI has to upgrade from “draft something for me” to: “Run a chain of reasoning and verification I can defend.”

A court just sanctioned lawyers on both sides of a case for fake citations. We ran their briefs through RealityCheck: every fabricated cite caught—and the real cases cited for things they never held. A citation can be 100% real and still sink you. #legaltech BriefCatch.com

@heyshrutimishra Yes, and that is because effective data construction can substantially improve the research ability of compact models🤏🦾 Weights are open: huggingface.co/collections/ap…

@Apodex_AI The 4B beating every 30B on BrowseComp is the quietly important number here. Worth digging into.

Prediction may be the hardest benchmark in AI. No memorization. No retrieval. No leakage. No spoilers. The answer literally doesn’t exist yet. Excited to see Apodex-1.0-mini (#1 this week) performing well in that environment.

We submitted four experimental prediction harnesses built on Apodex-1.0-mini to FutureX. They ranked #1, #2, #3, and #4🚀 One 35B model. Future prediction requires discovering, verifying, and synthesizing evidence under uncertainty.

@GavinRayDev Most thorough test we've seen so far, thank you for testing it. 🦾

I am skeptical of these new "AI Research" platforms. Decided to test: Apodex vs ChatGPT/Claude/Gemini/Grok/Meta AI (👇Links to each chat result below👇) To my surprise, it was the *best* -- but there's a catch! It runs 8 research steps and gives a final report. Each report has their OWN final answer, and all the info may not survive final synthesis. The best answer came from copy-pasting all 8 into another LLM and asking for a de-duped merge. Ranking: Apodex Merged > ChatGPT 5.5 Pro > Apodex > Gemini 3.1 Flash > Opus 4.8 > Grok 4.3 > Meta AI Apodex (Original): apodex.ai/share/81c5ce00… Apodex (8-Report Summary): gist.github.com/GavinRay97/ed9… ChatGPT 5.5 Pro (Deep Research): chatgpt.com/share/e/6a2ac7… Claude Opus 4.8 Extra (Deep Research): claude.ai/public/artifac… Gemini 3.1 Flash Extended: gemini.google.com/share/4cd74aa1… Grok 4.3 Expert: grok.com/share/bGVnYWN5… Meta AI (Contemplating): meta.ai/share/c/t0xQlS…