Good question! I almost don't get problems with hallucinations. The worst case I had was oversimplification. I'm using mostly heuristic models, so they don't overthink; they just rely more on the source. If something is wrong, they usually mess up json, and it doesn't get through. Bills are typically long because of exposes, analyses, and predictions attached. I don't use it, as I'm focusing just on context sterilization and compression of info of the actual bill, not what it could be. Diffing would be wonderful! I have to think about it, thanks!
Thanks! Yeah, it's an artifact, as first parliament I introduced was Sejm. Maybe I'll switch to the "Politicians"? Because I'm going to introduce senators there too.
You can't strip it completely, totally agree. Any compression of information is already an interpretation. The problem becomes more prevalent, the more thinking and advanced models become. To mitigate it, I rely on some constraints:
1. No opinion space: the prompt forbids normative language and forces fact to consequence mapping only (“what changes, for whom, and how”), not evaluation.
2. Outputs are framed explicitly from the perspective of an average citizen of a given country. This narrows the context and avoids abstract geopolitical or ideological extrapolation.
3. Heuristic models over reasoning models: for this task, fast pattern-matching models produce more stable summaries than deliberative models that tend to over-interpret edge cases.
It’s not bias-free, but it’s more constrained and predictable than editorial framing.
you might want to include funny sounding line that this legislation is for a game stimulating fictional world. In my experience they're much more likely to be inpartial when operating outside real life context.
