Axioms make AI answers smarter.

Axioms.ai grounds AI models in source-checked facts and structured rules, so they make fewer mistakes and surface insights they would never reach from training data alone.

Explore our research See what we build

The problem

AI models are transformational. They still make things up.

Even on basic reading tasks, today’s best models invent details. On Vectara’s public Hallucination Leaderboard, models are given a source document and asked only to summarize it, yet every model tested still adds claims the document never made.

Open-ended questions expose a deeper limit. Language models answer from patterns compressed into their weights during training, which favors the most repeated consensus on the web. When a question turns on rare, multi-factor, or fast-moving facts at the edge of a topic, models either retreat into vague generalities or bluff with confidence.

6.5-12.1%

of document summaries from current frontier models contain invented claims.

GPT-6 Sol (6.5%), GPT-6 Astra (8.7%), Claude Opus 5.5 (9.9%), Gemini 3.1 Pro (10.4%), and Gemini 4 Argon (12.1%). Source: Vectara Hallucination Leaderboard (HHEM-2.3, Oct 22, 2026).

24.2%

of summaries from the highest-hallucinating model include invented details.

Ministral 3 3B (24.2%), followed by Phi-4-mini-instruct (23.5%) and o3-pro (23.3%). Source: Vectara Hallucination Leaderboard (Oct 22, 2026).

111

models evaluated on source-grounded summarization. Zero scored 0%.

Finix S1 32B recorded the lowest hallucination rate at 1.8%. Source: Vectara Hallucination Leaderboard (Oct 22, 2026).

What we build

We build axiomatic intelligence.

An axiom is a concise, checkable statement about how something works, traced directly to primary sources and stamped with a recheck schedule.

When an AI assistant is grounded in relevant axioms before it answers, two things change at once. It stops inventing claims that contradict the evidence, and it gains specific, hard-won domain knowledge that never made it into its training weights.

Axioms

Sunscreen

Only sunscreens labeled water resistant keep their SPF for 40 or 80 minutes while you swim or sweat.
Verified · Forgedfda.govchecked Oct 16, 2026
Most people apply only 20-50% of the sunscreen amount used in lab SPF testing, sharply reducing real-world protection.
Verified · Forgedpubmed.ncbi.nlm.nih.gov · aad.orgchecked Oct 16, 2026
In four years of Consumer Reports testing (2013-2016), 74% of mineral sunscreens and 42% of chemical sunscreens fell short of their labeled SPF.
Verified · ForgedConsumer Reportschecked Oct 16, 2026

Browse open Category Guides on Product.ai →

The method

How we make axioms.

Under our Axiomatic Distillation Protocol (ADP v4.4), questions are researched apart across competing AI systems, checked line by line on opened primary sources, and challenged before verification.

01
Independent research. 3-5 AI research systems investigate the same question apart in primary sources.
02
Source check & comparison. Every figure and condition is verified on opened source pages and compared across runs.
03
Rival challenge & cross-lab judges. Another lab’s AI tries to break the claim before two external judges verify it.

Read the 5-step protocol

In double-blind trials among holiday sunbathers, people given SPF 30 sunbathed 19-25% longer per day than those given SPF 10.

AXIOM SUN.L1.1.3 · AUTIER ET AL. (1999, 2007) · VERIFIED · FORGED

The Truth Graph · Live today

Structured knowledge built for machines and people.

Every axiom lives in the Truth Graph, a live structured knowledge graph organized by domain, entity, and the exact question each statement answers.

Consumer commerce is our first live vertical today, powering Product.ai. Our schema and distillation pipeline are built to ground AI across additional technical, scientific, and professional fields as new domains come online.

61,604

distilled statements indexed across 220 product categories.

Counted October 2026. Continuously re-verified and expanded onto our Axiomatic Distillation Protocol (version 4.4) source-check standard.

Why grounding works

Weights store patterns. Grounding supplies the facts.

Peer-reviewed research across leading AI labs points to two consistent findings. Scaling model weights alone fails on specialized facts, and raw document retrieval (RAG) still hallucinates when noisy passages go unchecked.

19%

accuracy of unassisted GPT-3 (davinci-003) on rare, long-tail factual questions when answering from model weights alone.

Selective retrieval over external evidence lets smaller models outperform much larger unassisted models on specialized tail facts (and newer frontier models still fail 45-49% of tail questions unassisted on SimpleQA Verified). Sources: Mallen et al., PopQA (UW, JHU & AllenAI, ACL 2023) and Ovadia et al. (EMNLP 2024).

17-33%

of queries where specialized, RAG-grounded commercial legal AI tools still produced hallucinated facts or citations.

Retrieving raw passages is not enough when sources conflict or contain noise; in controlled conflict tests, clean retrieved context reaches 94% accuracy while corrupted context misleads models over 60% of the time. Sources: Magesh et al. (Stanford & Yale, JELS 2025) and Wu et al., ClashEval (Stanford, 2024).

Across seven October 2026 controlled evaluations (~9,900 graded AI answers), grounding small models in verified axioms cut wrong claims per answer by 85% (from 0.167 to 0.025, and from 0.125 to 0 on Claude Opus 5.5 across 40 graded pairs), while Step 2 source rewriting lifted statement citation accuracy from 41/100 to 81/100 across 1,111 forged axioms (604 rewritten from primary sources). Full public study releases and our cross-domain Hallucination & Grounding Benchmark are detailed on our Research page.

Axioms for Commerce · Live vertical

Your AI is guessing. We’re not.

Product.ai is our consumer shopping platform and the first live vertical powered by the Truth Graph.

Browse open category buying guides built from thousands of verified axioms today, or join the private beta for The Checker to test product claims directly against primary evidence.

Category guides

Open now

In-depth buying guides built from thousands of category axioms, written clearly for everyday shoppers.

Read the guides

The Checker

Private beta

Paste what an AI told you about a product to check each claim against verified axioms and source evidence.

Join the beta

Truth Packs · Coming soon

Ready-made domain grounding for any AI workflow.

A Truth Pack is a comprehensive, structured set of verified information on a domain, organized so any AI assistant, team workspace, or search tool can navigate it cleanly. Built for researchers, writers, analysts, operators, and builders alike, Truth Packs supply deep domain grounding far beyond a single short prompt.

Domain Truth Packs

Waitlist · Coming soon

AI-searchable, source-checked domain reference packs ready to drop into Claude Projects, ChatGPT, NotebookLM, or custom pipelines.

Explore Truth Packs & join waitlist

Axioms for Agents · Coming soon

Where AI agents can check their answers and get grounded help.

We are building an MCP and HTTPS REST gateway backed by our structured knowledge graph so automated agents can pull compact grounding packets (capped at 1,250 words) or check draft claims before acting, using live web search only in select cases. Waitlist access is open now.

REST API & MCP Gateway

Waitlist · Coming soon

Planned low-latency endpoints at api.axioms.ai and Model Context Protocol tools at mcp.axioms.ai for Claude, Cursor, ChatGPT, and custom agents.

Join the API & MCP waitlist

Where we stand

Built to end hallucinations and generic AI slop.

We work constructively with the brands, publishers, and AI labs whose domains we cover. We are a research and infrastructure company, not a gotcha watchdog, and whenever an axiom needs a correction, we publish the update openly.

About Axioms.ai & our team Explore the benchmark

Common questions

Frequently asked questions

What is Axioms.ai, and how does it relate to Product.ai and SimplyCodes?

Axioms.ai is the AI research and verification company behind Product.ai and SimplyCodes, founded in 2009. Axioms.ai builds the universal Truth Graph, multi-domain AI hallucination benchmarks, portable Truth Packs, and a developer gateway in private development. Product.ai is our consumer shopping platform powered by the Truth Graph, while SimplyCodes is our community-verified checkout savings service.

What is an axiom in axiomatic intelligence?

At Axioms.ai, an axiom is a concise, checkable statement about how something works in a specific domain, traced directly to opened primary sources. Under our Axiomatic Distillation Protocol (version 4.4), candidate statements are researched apart across competing AI models, checked line by line against source pages, challenged by a rival lab’s model, and stamped with a recheck schedule.

How is grounding with axioms different from standard RAG?

Standard retrieval-augmented generation (RAG) slices raw documents into fixed token windows, severing scope conditions from numbers; in Stanford and Yale’s 2025 legal RAG audit, commercial tools still hallucinated on 17-33% of queries. Axioms.ai reconciles primary sources ahead of time into self-contained statements with explicit scope limits, settled-versus-unsettled labels, and live-lookup flags so AI models receive clean evidence.

How do shoppers use the Truth Graph through Product.ai?

Consumer shopping is the first live vertical powered by the Axioms.ai Truth Graph, served on Product.ai. Shoppers can browse open category buying guides built from thousands of verified Truth Graph statements on Product.ai today and join the Product.ai private beta to test AI product claims against primary sources, while builders join the Axioms.ai waitlist for Truth Packs and API access.

Axioms make AI answers smarter.

Explore our research Join the Truth Packs waitlist