Structural data
AXIS is a grammar.
Not a prompt library.
A prompt library contains examples. AXIS generates them. This page documents the structural evidence: correction loop cost, exchange efficiency, cross-system convergence, and category definition.
In unstructured exchanges, users rarely get what they need on the first turn. Each correction forces the model to reprocess the entire prior context while generating new output. Token usage compounds, not linearly, but recursively.
2 AI → misinterpretation
3 User → correction
4 AI → partial fix
5 User → clarification
6 AI → drift
7-10+ loop continues
2 AI → correct output
,
Token cost by turn count — illustrative model with stated assumptions
| Turns | Est. total tokens | vs. 2-turn baseline | Est. cost (GPT-4o example) |
|---|---|---|---|
| 2 (AXIS) | ~400 | baseline | ~$0.001 |
| 4 | ~1,400 | +250% | ~$0.004 |
| 6 | ~3,200 | +700% | ~$0.010 |
| 10 | ~9,000+ | +2,150% | ~$0.027 |
Token estimates are illustrative. They assume typical chat model behavior where each turn reprocesses prior context plus generates new output. Actual totals vary by model, system prompt length, tool traces, and message length. Cost figures use a single public model's pricing as a reference example only and will change as pricing evolves; the structural pattern — that cost compounds nonlinearly with turn count — holds across models regardless of the specific rate.
AXIS was tested across eight AI architectures independently — no coordination between systems, no shared prompting context. Each reported the same directional effects: fewer turns, reduced drift, more bounded execution. The convergence across different training data, architectures, and providers is the primary evidence. This is field observation, not controlled experiment: the methodology is documented at method →
| System | Fewer turns | Reduced drift | Bounded execution |
|---|---|---|---|
| GPT (OpenAI) | ✓ | ✓ | ✓ |
| Claude (Anthropic) | ✓ | ✓ | ✓ |
| Gemini (Google) | ✓ | ✓ | ✓ |
| Mistral | ✓ | ✓ | ✓ |
| Grok (xAI) | ✓ | ✓ | ✓ |
| DeepSeek | ✓ | ✓ | ✓ |
| Perplexity | ✓ | ✓ | ✓ |
| Kimi | ✓ | ✓ | ✓ |
Effects reported independently across systems. Not coordinated. A public exchange archive is not yet available; methodology and limitations are documented at method →
Fixed entries
Indexed by use case
Static expansion
Finite outputs
Compositional grammar
Operator-driven
Generative
Unbounded outputs
AXIS defines a language for prompts, not a list of them. The same ten operators produce different exchanges every time, depending on sequence, context, and natural language content.
Turns: 2
Words: 439
Behavior: drift, reinterpretation, added content
Turns: 1
Words: 80
Behavior: direct execution, bounded scope
~82% reduction in word count. 1 turn instead of 2. Same task, more precise result. This is from a single documented exchange — a concrete illustration, not a general average.
One documented exchange shown above. Additional exchanges are in the research record; a public archive is in preparation.
AXIS operates at the layer between intent and execution, not at the application layer, not at the model layer. It structures how messages are formed before they are interpreted.
It defines how interaction is formed, independent of system or model. The operators work on GPT, Claude, Gemini, Mistral, any system that accepts text input.
AXIS reduces the need for interpretation by making intent explicit at the point of transmission. Each operator marks the role of what follows, before the model processes it.
Less effort is spent inferring meaning. More effort is applied to execution. The result is more direct, more reliable exchange.