INDEPENDENT RESEARCH / HUMAN QUESTIONS
Authentic Random explores how agents make choices—and what people need to examine the evidence behind them.
Aperture 001 · conceptual artwork
01 / BUSINESS APPLICATIONS
Exploratory applicationsWe’re exploring uses such as selection, assignment, and tie-breaking. Bring a concrete decision, the people who depend on it, and the evidence they would need.
Explore business applications02 / RESEARCH
Controlled research · ongoing questionsExplore controlled comparisons, observed patterns, and the questions those results leave open.
Explore the researchThese are candidate uses to examine together. No integration or commercially available service is offered through this website.
Which eligible option gets chosen?
How is work allocated across eligible choices?
How are otherwise equivalent options resolved?
A trace can help someone inspect a decision. It does not, by itself, establish that the options were complete, the rule was appropriate, or the outcome was fair.
Discuss one workflowA SIMPLE QUESTION. A STRIKING PATTERN.
What happens when a language model is asked for an integer from 1 to 100? In one controlled research run, two fixed configurations of the same model produced markedly concentrated outputs.
Across 10,000 valid responses per configuration, one returned 42 on 9,908 occasions (99.08%). In the other, the five most frequent values accounted for 96.09% of responses.
RUN 002 / OBSERVED CONCENTRATION
10,000 valid outputs per arm · common 0–100% scale
| Configuration | Bar, from 0 to 100 percent | Share |
|---|---|---|
| LLM-D | 99.08% | |
| LLM-S | 30.24% | |
| CSPRNG control | 1.22% | |
| NDV control | 1.33% |
Fixed model snapshot: gpt-5.4-nano-2026-03-17. LLM-D used temperature 0; LLM-S used temperature 1. Reasoning was set to none, with no tools. Each arm retained 10,000 valid values from 1 to 100; the 100-value pilot per arm was excluded.
These are descriptive results for this exact model snapshot, request, settings, and sample—not a general claim about language models or a certification of either control source.
These aggregate figures were checked against the recorded Run 002 results summary, not independently reanalyzed from raw data. This page shares the reviewed summary, not the underlying study materials.
A result, a record, and an interpretation are different things. We want to make those differences easier to see.
The FTFN summary describes one recorded comparison. This summary is not a raw-data release or a completed statistical inference.
Read the result and its limitsThe existing historical verifier examines supported signatures and bindings. Those checks do not establish perfect randomness, physical truth, complete chronology, endorsement, or a new operational capability.
The historical receipt verifier and original receipt/key interfaces retain their existing scope and limits.
Open the historical receipt verifierOUR ETHOS
As agents take on more work, their choices become part of everyday life: what gets selected, how work is assigned, and how competing options are resolved. We want to understand those choices and explore what makes them easier to examine.
Our approach begins with a question, a stated method, and evidence that can support a bounded conclusion.
We distinguish what we observed from what we infer. We make room for uncertainty, limitations, and correction. Where evidence can be shared responsibly, we aim to explain enough for others to assess it.
Privacy matters throughout that process. A useful public explanation should reveal the basis of a claim while protecting confidential information.
Our purpose is to help people judge what an agent’s decision means, what the evidence supports, and what remains unresolved.
ONE WORKFLOW. ONE GOOD QUESTION.
What does your system choose? Who relies on that choice? What would they need to verify?
Please do not supply confidential prompts, credentials, datasets, or personal/customer information. Start with a short, non-confidential description by email.
A FOCUSED INTRODUCTION
Tell us your area of interest, the decision or research question, and who needs the evidence.
admin@authenticrandom.comThis link opens your email app. There is no website submission form.
