May 20, 2026
Bot on a leash, comments unleashed
Formal Verification Gates for AI Coding Loops
Teach the machine some rules it literally can’t break, says fed-up coding crowd
TLDR: A developer says the fix for AI-made coding mistakes is simple: stop relying on polite instructions and make the computer reject bad code automatically. Commenters loved the idea of hard guardrails, but some immediately pounced on flaws in the example, turning the thread into a classic “smart idea, messy execution” showdown.
The big mood in this discussion is basically: stop trusting AI to “remember” the important stuff. The article argues that the scariest software mistakes are often the simplest ones, like letting one customer peek at another customer’s private data. The author’s fix is to stop hiding those rules in prompts and checklists and instead bake them into hard system rules that the computer can reject automatically. In plain English: don’t just ask the bot nicely — build the fence so it can’t wander off. The author even popped into the comments to sum it up and plug the Shen-Backpressure repo, framing it as a practical answer to the “AI wrote 16,000 lines, now what?” panic.
And the comments? Very “this is smart, but hold on…” energy. One camp was fully on board, cheering the idea that AI is good at making the shape of software but terrible at consistently keeping promises over a long stretch. Another commenter basically translated the whole thing into a meme-able truth: the runtime — the actual program — has to be the bad cop, because the bot sure won’t be. But then came the plot twist: critics said the demo’s security example itself looked shaky, especially around login tokens, sparking the classic tech-thread drama of “great concept, questionable details.” So the community verdict is deliciously split: this could be the future of keeping AI-generated code honest… unless the guardrails themselves are flimsy.
Key Points
- •The article argues that AI-generated code makes prompt-level enforcement of critical software invariants increasingly unreliable.
- •It presents a distinction between behavioral gates, such as prompt instructions, and structural gates, such as compilers, type checkers, linters, test runners, and proof checkers.
- •The article’s central claim is that structural backpressure is a more dependable control mechanism than waiting for smarter AI models for many production software tasks.
- •Shen-Backpressure is introduced as a tool and methodology for expressing invariants in machine-checkable form and feeding gate failures back into the coding loop.
- •The article uses access-control rules as an example of an invariant that can be encoded in Shen and projected into target-language types, constructors, and gate commands.