Pages
On this page
Truth Packs · Coming soon
An Axioms.ai Truth Pack is a structured, AI-searchable body of source-checked knowledge on a subject. Drop one into your favorite AI assistant, research workspace, editorial desk, or product workflow so AI stops guessing from generic web training data and works from verified domain facts.
What makes a Truth Pack different
Most people try to ground AI in one of two flawed ways, pasting a messy folder of raw PDFs and scraped articles full of noise, or hoping a short prompt snippet covers every edge case.
Truth Packs fill the space in between. Unlike our Developer Gateway, which is designed to return compact packets (capped at 1,250 words) from our knowledge graph (with live web search only in select cases) for narrow single-turn agent calls, a Truth Pack can span thousands of words across an entire domain. Every mechanism, rule, exception, and quantitative finding is distilled through our five-step verification protocol, organized under clear question headers, and tagged with primary sources so an AI model or local search index can jump straight to the exact evidence a task requires.
Upload a Truth Pack straight into Claude Projects, ChatGPT, Gemini, or NotebookLM to draft, fact-check, and explore complex domains with primary citations attached.
Give internal AI assistants and team workspaces a vetted domain spine on growth, compliance, category science, or technical workflows without maintaining a custom RAG stack.
Index clean, pre-reconciled Markdown sections with explicit entity tags, scope conditions, and live-lookup boundary markers instead of noisy scraped HTML.
Inside a Truth Pack
Every Truth Pack combines four building blocks, so AI can answer with fewer errors and better coverage.
A topic index opens the pack and points each question to its axiom ID, so a model finds the right passage fast.
Each axiom carries its own scope, figures, why it holds, and primary source links, so any excerpt stands alone.
Verified axioms sit apart from graded mixed items, so an AI never states a partly supported claim as settled.
Each pack is dated and names the facts a static file cannot hold, such as recalls or local code changes.
indoor-air-quality-sample-truth-pack.md · 25 axioms · ADP v4.4
This 25-axiom pack covers indoor air quality: CO2, airflow, MERV filters, gas stoves, mold, and radon. Drop the Markdown file into Claude, ChatGPT, Gemini, or NotebookLM, or paste it with an indoor air quality question.
Questions to try
Preliminary results from internal testing
| 25 / 25verified axioms | Confirmed against primary sources by two blind cross-family AI judges, 0 contradicted; 4 mixed items sit in their own block. |
|---|---|
| -54%small-model errors | Wrong claims per answer fell 0.212 to 0.098 across Claude Haiku 4.5, GPT-5.4 mini, and Gemini 3.1 Flash-Lite (-84% on Flash-Lite). |
| -78%Gemini 4 Argon errors | Wrong claims fell 0.087 to 0.019 (-32% pooled frontier). Claude Opus 5.5 gained 1.5 web-verified claims per answer, with 0/57 false safety blocks vs. 7/57. |
| +9.4 ptsadjacent-topic coverage | Required-fact coverage rose 53.1% to 62.5% on related indoor air quality prompts outside the pack (+8.3 points overall, all 6 models). |
Source: Preliminary internal results on the Indoor Air Quality sample pack (October 2026; an internal test across 57 indoor air quality prompts and 1,949 answers on 6 models, each answer scored blind by 2 cross-family AI judges). Pre-registered studies are in progress on our Research page.
State of the research
Recent retrieval and long-context research shows that the format, structure, and signal density of external knowledge matter just as much as whether external data is provided at all.
fewer top-20 retrieval failures when contextualized units combine hybrid search and reranking.
Anthropic’s Contextual Retrieval study (2024) measured 35% fewer misses from contextual embeddings alone, 49% with contextual BM25, and 67% with reranking.
Contextual units
Traditional document chunking slices pages into fixed token windows, severing claims from the subject, timeframe, or condition they belong to. Anthropic’s Contextual Retrieval research showed that prepending explicit chunk context before indexing reduced top-20 retrieval failures by 35% with contextual embeddings alone, by 49% when paired with contextual BM25, and by 67% with reranking. Every axiom in a Truth Pack is written as a self-contained, context-tagged unit from the start.
Structure vs. chunks
When a question requires connecting mechanisms across multiple sub-topics, flat text chunks leave the model guessing how pieces fit together. Microsoft’s GraphRAG and OSU’s HippoRAG (NeurIPS 2024) demonstrate major gains on multi-hop synthesis when domain knowledge is organized explicitly around entities, relationships, and topic summaries rather than arbitrary paragraphs.
Signal density
Dropping unedited articles or raw transcripts into a prompt carries a steep cognitive tax. Peer-reviewed studies show accuracy falling 20+ points on facts buried in unindexed middle text (Lost in the Middle), reasoning degrading as noisy tokens accumulate (Same Task, More Tokens), and irrelevant sentences derailing logic (Shi et al.). Distilling sources into concise, indexed statements lets a Truth Pack cover a broad domain while keeping irrelevant noise near zero.
Pre-reconciliation
When raw retrieved passages contradict one another or include outdated numbers, language models fall back on pre-trained priors or attach post-hoc citations that do not support their claims (ClashEval; Astute RAG; Wallat et al.). Reconciling primary sources and marking settled versus unsettled claims before packaging gives AI models clean evidence they can follow reliably.
Formal pre-registered Truth Pack studies across non-commerce and commerce domains are in progress on our Research page, building on the preliminary sample-pack test above.
Waitlist · Coming soon
Truth Packs are currently in preparation, starting with an open community release on early-stage founder audience growth alongside consumer and professional domain packs. Join the waitlist below to get notified as packs release and tell us which topics would help your work most.
No spam. We will reach out as initial open Truth Packs become available.
Common questions
An Axioms.ai Truth Pack is a portable, structured Markdown reference file covering the core mechanisms, quantitative rules, exceptions, and primary sources of an entire domain. Each Truth Pack combines 4 elements: a routing taxonomy and question index, self-contained statements with source URLs, explicit settled versus not-settled status marks, and scheduled recheck cadences with live-lookup flags.
Dropping unedited PDFs or fixed-window text chunks into an AI workspace severs numbers from their scope conditions and buries critical facts in mid-prompt noise. Anthropic’s Contextual Retrieval benchmark found that self-contained, context-tagged units cut top-20 retrieval misses by 35-67%. Axioms.ai writes every Truth Pack statement as a pre-reconciled, context-carrying unit from the start.
Axioms.ai designs Truth Packs and the Developer Gateway for 2 different context budgets. A Truth Pack can span thousands of words to equip Claude Projects, ChatGPT, NotebookLM, or an internal team index with comprehensive domain coverage. By contrast, our planned developer gateway queries our structured knowledge graph (using live web search only in select cases) to return compact packets (capped at 1,250 words) sized for single low-latency agent turns.
Axioms.ai Truth Packs are currently preparing for initial release, starting with an open community pack on early-stage founder audience growth alongside consumer and professional topics via the waitlist on this page. If you want consumer product guidance today, you can browse open buying guides built from the same Truth Graph directly on Product.ai.