Say what the AI did, and what it did not do.
Two distinct uses of the word “AI” coexist in LIVEX: the assistance of generative tools during development, and the behaviour of the entities of the simulated world. This page tells them apart.
A solo project, augmented and responsible.
LIVEX is carried by a single person. It would have been hardly achievable alone without the help of artificial-intelligence tools, which speed up documentation, verification and certain repetitive tasks. This is a fact, and concealing it would be a form of dishonesty.
This page describes where these tools intervene, how their output is controlled, and what they are not in the project.
Guarded assistance, never blind.
Assistance mostly concerns written output and verification. It decides nothing on its own: every phase below remains under the author's approval.
| Phase | Use of AI |
|---|---|
| Verification | Review and consistency checks on passages already written. |
| Rewriting | Help with clarity and concision, subject to the author's approval. |
| Documentation | Correction and formatting of repository documents. |
| Code | A few functions, reviewed and tested like any other contribution. |
| Runs and calibration | Help preparing and analysing simulation runs. |
| Review | Double-checking of proposals before integration. |
The decision remains human.
Design choices, architecture trade-offs and sensitive wordings remain written by the author, then verified. Nothing is published on the sole basis of an automatic generation: any claim about a capability, a figure or a property is confronted with the real state of the project before being released.
What bounds the assistance.
Four structural controls keep assisted output from turning into an unverified claim.
Human review
Any assisted output goes through a human review before being kept.
Tests
Assisted code is submitted to the same test suites as the rest of the project; no preferential treatment.
Traceability
The repository, its commits and its changelog remain the verifiable record of the work.
Register of claims
Sensitive claims on the site carry a dated internal source, regardless of how they were written.
The agents are not language models.
The entities of LIVEX are not language models, and the world is not “learned” by an AI. Their cognition rests on explicit, deterministic and inspectable rules. No generative model intervenes at the heart of the simulation: the question LIVEX asks — what local, replayable laws produce — requires a cognition that is poor but transparent, not an opaque network.
See also the distinction between simulated agent, learning AI and scripted character inEmergence explained.