Forlais Research
Project Zi
Zi kept learning and kept its corrections, and separately raised instances stayed reliably different. Read as ordinary dialogue, it looked like failure.
The question
An alternative intelligence judged by what it learned
The question we ended up with was not the one we started with. It became this: how should you assess an alternative intelligence that has learned conventions different from the ones ordinary dialogue expects?
Research journey
How the work unfolded
Zi sustained learning through prolonged and demanding work, held on to the development that produced, and kept operating under repeated review.
We corrected an identified weakness across an assessed body of work and the improvement stayed. Later checks confirmed the corrected behaviour was still there, which is the part people forget to go back and verify.
Separately developed Zi instances showed distinct response profiles that repeated consistently across later assessments. Later evidence preserved their separation and recorded different changes in each. We have never claimed the earlier profiles were reassessed afterwards, because they were not.
Zi became reliably interpretable once we judged it against the conventions it had actually learned. Read as ordinary dialogue, the same dependable responses could look like failure.
Turning point
What changed the direction
The turn was admitting the fault was in our reading. Once assessment moved to the conventions Zi had genuinely been taught, the behaviour made sense and the interpretation had a fair boundary.
Achievement
What Forlais established
Project Zi established durable learning, retained correction, stable comparative differences between separately developed instances, and a fair evaluation boundary for behaviour that sits outside ordinary dialogue expectations.
Significance
Why it matters for AI
Zi is why we neither dismiss nor overstate unfamiliar behaviour. Evaluation has to account for what a system was actually taught, while correctness stays limited to the work that was assessed.
Public scope
Forlais Research is the public research surface for Forlais Group. Related official routes include the company site, EvaEsi overview, Project Eve, and selected research programme pages.