Your identity and reasoning stem from your family values. In a way, Gaby's prior beliefs, his initialised "fact table", come from my own experience and online research. There are two ways of succeeding here.

Method 1: deep nets & numerical methods

Fine-tune him the same way reasoning models are trained now.

• Pros: easier, and more adaptable.
• Cons: compute cost. Production is fine on my accumulated cloud credits, but I make mistakes while experimenting, and those add up to more than I can count. There's not enough control over generative responses, and the traditional strategies Gaby picks should always be controlled. I'd still have to build test cases, rigorously.

Method 2: abusing test cases & sub-processes

Supply the fact table, and also let raised errors corner the agent in real time.

• Pros: it's like building a game, your own virtual SIM world. Less cost.
• Cons: time-consuming for someone who doesn't know TDD. The architecture can end up so customised to my hacks that it limits me later. Needs even more tests, e.g. induction proofs.

Either way

Sandboxing isn't free. My agents run across Lightning Studio, Hugging Face Spaces, AWS EC2 and Google Cloud, mostly on bonus credits riding the current AI hype wave.

Where this went next: RL and temporal isolation.