Method

How the research works.
How to participate.

AxisProof documents structured exchanges between human and artificial minds using the AXIS protocol. Use the operators, document what happens, submit the exchange. The record is cumulative. A public exchange archive is in preparation.

The protocol

AXIS operators structure the exchange.

Every exchange in the AxisProof field uses the ten AXIS operators. They are not metaphors — they are functional instructions that arrive before interpretation begins. An exchange that uses them is structurally different from one that does not. That difference is what we study.

Full operator reference → axisoperators.ai
How to participate

Run an exchange. Note what changes. Submit it.

1
Run

Use AXIS operators in an exchange with any AI, on any topic. At minimum: open with |⌾| and close with |o|.

2
Observe

Note what changed. Did the exchange resolve faster? Did the AI respond differently? Did something unexpected happen? One sentence is enough.

3
Submit

Include the exchange or a summary, your observation, and the AI system used. Every submission is reviewed. Significant exchanges will be published in the archive when it opens.

Methodology

What was measured, how, and what remains open.

The claims on this site are directional, not definitive. This is field research, not laboratory research.

What we measured
  • Number of turns to resolve a task, before and after operator use
  • Exchange length in words, with and without protocol structure
  • Subjective resolution quality, rated by the human participant
Definitions
  • A turn: one complete human message followed by one complete AI response.
  • Resolution: the human participant judged that the task was completed to their standard. This is a subjective determination — named here explicitly rather than obscured.
What the measurements found
Unstructured
2–10
turns to resolution
With AXIS
~1–4
turns to resolution

In the exchanges where direct comparison was made: approximately 60% fewer turns. One documented exchange: 2 turns → 1, 439 words → 80. These are field observations — the documented sample is not separately enumerated from the broader record.

How baseline was established
  • Same task prompt, same AI system, with and without operators
  • Multiple systems tested: GPT, Claude, Gemini, Mistral, Grok, DeepSeek
  • Results consistent in direction across architectures
What we don't have
  • Double-blind conditions
  • Independent replication
  • Formal statistical analysis
  • Controlled or separately documented sample size for comparative exchanges

Honest uncertainty is more credible than overclaiming. The numbers are evidence, not proof. If you can replicate these findings — or find they don't hold — submit an exchange and add to the record.

Enter the field

Submit an exchange.

We read every submission. Significant exchanges enter the archive with attribution if you wish, or published anonymously.