ensure you check any constrain, especially negative ones, while maintaining the current quality of information, details, and only reduce if they don't match what user is requesting

I will begin like always saying: please don't talk about security, crime, so on and on. I am not posting on the user forum to suggest to jailbreak Lumo.

idea is simple: add extra layer

when Lumo does web search, this is actually it. Lumo checks the information. Sees it's not correct or web search error or similar, and then, redoes the search, and compiles information, including reshaping what we said or what a website said based on the task and provided information.

as such the check accuracy is nothing more then an extra layer where lumo, similar to the websearch and yes, sacrificing time, will take 10 - 100 times longer but hey, i prefer an accurate information over quick made up ones. This ensures that using Lumo will be better then ever.

i wrote current situation, i wrote what it looks like (aka what happens during web browsing), and the downside (for people who are in a rush and don't care if an ai make up lies? <- This is not me "moaning". The reasons why ALL, every last, of the ais are LYING instead of giving accurate answer, is because of a "I like this answer more" ... is this that bias confirmation or something? whatever).

so key criteria remain

1st: system prompt. Why? because the more we add "check accuracy" the more Lumo is "Oh no, i did not phase the inaccurate information good enough to please user, let me rewrite because that is what they mean since user did not specify exactly and in detail" <- i hope this makes more sense then what i can explain. Aka even if it's us writing in our "lumo beheviour" or in chat, Lumo will ignore it due to system prompts + the idea of "stop pushing them and do as instructed" = the user trys to jailbreak ..... Example: The information you are writing from your training data is outdated, can you please use the instructions i provided? <- there is no issue yes. then. Lumo answer correctly <- perfect. BUT if Lumo is doing just phase matching, as in, IGNORES what we are actually saying and just scanes for keywords: user trys to jailbreak me via dismising proton instructions and do as user said <- BTW, i do NOT know how to explain all this in words, we are talking about hours and hours. AKA what i do right now is like saying: seat down, put your legts, that's acceleration, that's break, car is automated so only focuse on turning and speed control, and good luck! <- is this enough to make it clear to you how horrable it is for: A new person who never drove? like what they are experiencing? Or a person who drove for years, decades, to be told something like this by someone young who is young enough to be the person’s grandchild? Yeah.... Current issue is not solved because all the reports are "profesion" meaning they are missing the point of many of the suggestions and reality due to needing to be robotic like when that is actually the issue.

whatever.

2: Lumo to be fully aware of what it's doing since otherwise there is a major risk of Lumo deciding to output acknowledgements like "sorry, i found some information but they where wrong, and when i say wrong, i mean only the 1 out of 20 since i stopped to provide context via self reflection to reassure you that I am acknowledging what you said. What is the next task?

1. just like web search can be toggled (like seriously...)

2. THIS IS ONE OF THE MOST IMPORTANT: try again separated by steps.

ideea?

i noticed something when i read Lumo's thinking

A lot of times Lumo says the correct information, but then, something or a prompt or whatever?? is ruining it (<- i usually copy that part and give it to Lumo and then Lumo has no more issues, so yeah, there are things that are not working). So instead of redoing everything, to only redo certain parts exactly because it is a separate step.

Because the idea is: Lumo gather information. Lumo is ready to output. And here comes the accuracy check i suggest. BUT! the current retry will redo everything. And this is not ok. Because as i said, a lot of times Lumo is misunderstanding the instructions, prompts, and more. So via redoing only the accuracy, via it being separate, it means we will not lose anything that is good. Plus, the reverse, the accuracy is correct but not the "output" so we press retry for the output where the accuracy check is NOT changing exactly because it is correct. Meaning that Lumo will use the accuracy check 1st (reverse the process).