Já na programação de games faço o contrário. Lá o resultado é o que importa e o método não faz diferença no mundo real, então eu realmente não leio muita coisa!
É isso, payments é como o meu caso, não temos margem pra erro. Na engenharia tá em jogo a segurança de quem ocupa a edificação e a credibilidade do produto. Áreas com impacto severo no mundo real precisam de muita cautela
Lá no final tem um engenheiro que assina o resultado obtido no meu software. Mesmo com a auditoria batendo 100% com a literatura, eu leio e confiro. Resultado certo não garante método certo
A PR inteira não. O pensamento crítico que vc questiona no post tá em saber onde ler. Apesar dos gates e do meu método de auditoria, o núcleo de cálculo eu leio inteiro: funções, coeficientes, unidades etc, tudo tem que bater 100% com a norma/método de cálculo
Building a foundation sizing tool where the AI part doesn't get to pick anything
It organizes the request, asks questions, looks up methods. The engineer OKs every assumption and the math runs in deterministic, tested code
Missing data stays pending, never passes by default
ALT Code of the gate that refuses to run the calc engine until every input is filled and the engineer confirms the assumptions, with a terminal run of three cases
pra crud talvez os gates bastem. em cálculo de engenharia eu ainda leio PR e documentação com calma, porque um detalhe errado ali não vira bug, vira problema na obra. gate ajuda, mas não substitui ler
What the API for my foundation design tool looks like so far. Column, soil and materials in. Footing size, rebar and the checks out
Still in development, but it runs the same engine as the app. Engineer review comes back as not reviewed until a person does it
ALT Request and response code of the footing API, with a real run showing footing size, rebar, the checks, pending items and engineer review not done
ran out of github actions minutes, so my old ryzen 3900x became the ci runner. works perfectly lol. rest of the day: progress on the action rpg i'm building, ui fixes and audited calcs on the engineering side
Every change in the foundation tool that moves a number the user sees, or flips pass, fail or pending, ships with a decision file in the same commit
Context, sources, decision, alternatives, effect, how it was checked. Agent-made calls need independent review before merge
ALT README describing the decision file required for every change that moves a user-visible number
An optional input in the foundation tool showed 0 on screen while the engine got undefined. The check that used it just disappeared and the footing could pass
A half-filled pair stays pending now. In the repro, actually typing that 0 made the same footing fail
ALT Code where a half-filled optional input pair stays pending, plus terminal output
original games and whole genres are still allowed. so claude is fine, the only thing dead is the reskin-someone-elses-game business lol
burned through all 2,000 free github actions minutes this weekend lol. lots of shipping, a bit of engineering, a bit of the action rpg i'm building. any open source ci you'd self-host on a vps?
ALT GitHub Actions usage alert: 2,000 of 2,000 included minutes used this billing cycle, with a full progress bar.
A pile result has several checks and the screen shows one as the reason. It was picking a pending check even when a firm failure was in the same list
A sweep of 1296 inputs found 392 like that. A firm failure goes first now
ALT Code that picks the governing check, putting a firm failure ahead of a pending one
the skills as Agent() part is the interesting bit. can they call each other, like ImplementSpec() running CodeReview() before it returns?
There's also a file listing the places where the foundation tool is more conservative than it needs to be, on purpose
It exists so nobody "fixes" that extra margin without a sign-off
ALT File listing intentional conservative choices in the foundation tool
Angular distortion in the foundation tool was only checked for the combination with the most settlement. A lighter one can be worse if the neighbor settles more
One case showed 1/1084 and passed. Real worst was 1/277, limit 1/300. Every service combination is checked now
ALT Chart of angular distortion per service combination against the 1/300 limit
pra site e tela concordo. no meu caso é cálculo de fundação, e aí "chegar lá com iterações" sem ler o código e auditar funções extensivamente não rola lol
same here. i build engineering software and a wrong number there ends up in a foundation lol. so the AI only reads the inputs, the engineer checks them and the math is plain tested code