Pages
On this page
At Axioms.ai, an axiom is a single checkable statement about how something works. Under our Axiomatic Distillation Protocol (version 4.4), statements go through 5 independent checks across competing AI systems as new categories are distilled and existing Truth Graph shelves are re-verified.
How an axiom is tested · ADP v4.4
Our Axiomatic Distillation Protocol (ADP v4.4) decomposes a domain question into concrete mechanisms and runs four adversarial research and source-verification steps, followed by an independent cross-lab judge gate outside the pipeline. Because competing AI systems handle adjacent stages, single-model blind spots and hallucinations get caught as categories are distilled and upgraded.
Starting from a concrete question about how something in a domain works, 3-5 deep-research systems across competing providers investigate primary sources on their own, including refute-first passes.
Every specific figure, date, limit, and mechanism in the candidate answers is checked directly on the opened primary source pages, and unsupported claims are cut.
Candidate answers and underlying studies are evaluated side by side across runs to record where independent models converge, where studies split, and what scope limits apply.
Another company’s AI tries to prove the surviving statement wrong, attacking weak evidence, missing caveats, and scope overreach before the final wording is locked to confirmed passages.
Outside the distillation pipeline, two cross-lab AI judges test the finished statement against an independent source check before it earns a Verified standing (tiered as Forged or Probable; single-run Signal leads stay unverified until corroborated).
Inside the source check
First-draft AI research syntheses frequently blend real citations with unverified training-weight priors. Peer-reviewed retrieval audits show that up to 57% of post-hoc RAG citations in tested setups fail to causally support the specific claim they are attached to (Wallat et al., ACM SIGIR ICTIR 2025).
During our source check, every quantitative figure, date, scope condition, and mechanism must be read directly on an opened source page (if a paywalled paper cannot be opened in full, the record logs that explicitly). Unmatched details are cut, and the surviving statement is composed strictly from confirmed passages.
Independent verification
Once an axiom finishes the four distillation steps, a separate web-grounded model from a third lab compiles a fresh source check on the claim. Two frontier judges from competing model families then evaluate the statement independently (Panickssery et al., NeurIPS 2024; Li et al., ICLR 2026). Under ADP v4.4, a statement earns Verified standing only when both judges agree.
When credible primary studies split or a candidate statement fails either cross-lab judge, it never receives a Verified label; unsettled findings are marked Not settled with the competing evidence recorded alongside them. In our upcoming Axioms.ai research studies, blind human readers also spot-check sampled outputs alongside the automated judges.
The axiom record
Each record in the Truth Graph stores its standing chip, the opened primary URLs, the four-step test trail (Research, Source check, Compared, Challenge), the date checked, its scheduled recheck clock, and What would change it, the specific evidence that would revise or overturn the finding.
Every published axiom record also includes an inline Ask us to recheck this control so readers, brands, and researchers can flag new primary evidence directly on the statement. Whenever new evidence changes an axiom, we update the record and preserve the revision history.
Sunscreen protects only the skin it covers. In UV-imaging and fluorescence clinical studies, people missed a median 20% of intended body skin in a single application, including 14% of the eyelid region and frequent 0 mg/cm² gaps at the ears, temples, and neck.
How this axiom was tested
Under our Axiomatic Distillation Protocol (version 4.4), statements go through these 4 checks before 2 external AI judges confirm support, ending with a challenge where another company’s AI tries to prove the claim wrong.
An axiom is a concise, checkable statement about how something works, traced directly to primary sources.
Learn how we make axiomsSelect a step to see what it found.
3 or more AI systems researched this apart
Research systems across competing providers each searched primary sources and wrote up what they found, without seeing each other’s work.
All 3 clinical studies confirmed
We opened each clinical study record and verified the exact missed-coverage figures quoted in the statement.
The UV-imaging and fluorescence studies aligned
Independent clinical application studies point the same way, documenting a median 20% of body skin missed overall, 14% missed around the eyelids, and frequent 0 mg/cm² gaps on the ears and neck.
Held up to a challenge
Under ADP v4.4, a model from a competing lab argued against this statement and couldn’t break the confirmed clinical findings.
What would change it
Multi-site UV-imaging trials showing everyday single-pass application consistently covers over 95% of exposed skin, including eyelids, ears, and neck.
Checked October 16, 2026 · Stable, rechecked within 12 months.
Email draft opened to contact@axioms.ai.
Comparison across runs
During the Compared step, we track how many independent research runs surfaced the same finding on their own and whether underlying primary studies agree or split. Verified axioms carry either a Forged tier (3+ runs aligned and primary sources confirmed) or a Probable tier (2 runs aligned with qualified boundary conditions), while single-provider Signal leads remain unverified candidates until corroborated.
| Category | Definition |
|---|---|
| Broad convergence · Forged | Three or more independent research runs surfaced the finding, primary studies align, and both cross-lab judges confirm Verified standing. |
| Qualified scope · Probable | Two runs converged while a third surfaced boundary conditions or population limits that narrow the Verified statement. |
| Single-run novelty · Signal (unverified) | Surfaced by a single research run; held as an unverified candidate that requires independent validation before earning Verified standing. |
| Not settled | Credible primary studies point in different directions; surfaced transparently as Not settled rather than forced into a single answer. |
Recheck schedule
Static storage is the wrong place for prices or active recalls that can shift overnight. Live prices and active recalls require real-time lookups, while stored axioms re-verify on three scheduled cadences.
| Category | Definition |
|---|---|
| Stable | Rechecked within 12 months. Physics, physiology, chemistry, and durable material mechanisms. |
| Seasonal | Rechecked within 6 months. Product formulations, annual hardware lineups, and updated clinical or technical guidance. |
| Volatile | Rechecked within 90 days. Regulatory thresholds, return policies, and dated compliance rules. |
Internal pipeline evaluations
Whenever we revise ADP, we run internal evaluations before shipping pipeline updates. In our October 2026 skincare audit across 1,111 forged axioms, Step 2 source rewriting (604 axioms rewritten directly from primary sources) raised statement citation accuracy from 41/100 to 81/100, and across seven controlled Truth Pack evaluations (~9,900 graded AI answers), grounding cut wrong claims per answer by 85% on small models (0.167 to 0.025). Our full external benchmark and grounding study releases with open datasets are detailed on our Research page.
Total API and search spend per finished axiom reconciled against provider invoices in our internal harness so protocol changes stay economically viable.
See studies in progressWall-clock time to research, source-check, compare, challenge, and verify an axiom across labs, including retries and rate limits.
See studies in progressShare of specific figures, mechanisms, and thresholds verified directly on opened primary source pages (lifted from 41/100 before Step 2 source rewriting to 81/100 across our 1,111-axiom October 2026 audit, with 604 axioms rewritten from primary sources prior to cross-lab judge gating).
See studies in progressCommon questions
Under the Axioms.ai Axiomatic Distillation Protocol (version 4.4), no model approves its own work as categories are distilled and Truth Graph shelves are re-checked. First, 3-5 AI systems investigate primary sources apart; second, figures and conditions are checked on opened source pages; third, runs are compared for convergence; fourth, a rival lab’s AI challenges the draft; and fifth, 2 cross-lab judges confirm primary-source support.
When underlying primary studies split or a candidate statement fails either independent cross-lab judge under the Axiomatic Distillation Protocol (version 4.4), Axioms.ai never forces a single consensus answer. Instead, the record is marked not settled and preserves the competing primary evidence side by side so downstream AI assistants report genuine uncertainty accurately without guessing.
Axioms.ai assigns every stored axiom an explicit re-verification clock matched to how quickly the underlying facts change. Durable physical, chemical, or physiological mechanisms recheck within 12 months (Stable); product formulations and annual hardware lineups recheck within 6 months (Seasonal); and regulatory limits or return policies recheck within 90 days (Volatile), while daily prices require live lookup.
Yes. Every axiom record published by Axioms.ai and surfaced on Product.ai records its opened primary URLs, check date, recheck cadence, and a specific What would change it falsification condition. Readers, domain researchers, and brands can use the inline Ask us to recheck this control to submit new primary studies, and confirmed updates publish openly with revision history.