Forlais Research
Eros
We built an agent foundation we could trust and ran a learning line beside it, then found that a good final reading can hide real damage done along the way.
The question
Dependable operation and learning judged separately
The idea was to keep two things apart that are usually reported as one. Could an agent foundation hold continuity, authority and accountable evidence, while a separate learning line improved on real data without quietly concealing what it damaged?
Research journey
How the work unfolded
We built the foundation first and gated it independently: exact continuity across a project, firm permission controls, refusal of actions it had no authority to take, and research missions that carried their sources. Later checks confirmed those safeguards still held.
In a separate line, Eros produced repeatable improvement on unseen real data. An independent review recovered the same recorded progression and supported a narrow learning result, which is the only size of result that evidence supported.
That work recovered a useful contribution earlier work had thrown away, while preserving the reference outcome. When we tried to use it more widely it caused harm, so we bounded the benefit and rejected the unsafe extension.
Then came the finding that changed how we assess everything: a favourable final reading could conceal damaging change during the run that produced it. We strengthened assessment until deterioration could no longer hide behind a good ending.
Turning point
What changed the direction
The turn was catching a positive final reading covering for real damage mid run. From that point the whole course of a run became evidence, not just where it finished.
Achievement
What Forlais established
Eros gave us a dependable agent and execution foundation alongside a distinct learning line that delivered repeatable improvement, recovered useful discarded information, preserved prior supported behaviour, and exposed deterioration that had been concealed.
Significance
Why it matters for AI
Eros is why we judge dependable operation and learning separately, and why we hold both accountable throughout a run rather than only at the end of one.
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.