LLM Output Variance: When a Single Targeted Retry Pays Off
45% of LLM outputs changed on identical input. Here is how to turn that variance into a one-shot retry that recovers quality without burning retry budgets.
45% of LLM outputs changed on identical input. Here is how to turn that variance into a one-shot retry that recovers quality without burning retry budgets.
We benchmarked glossary translations scraped from official sites against two-model LLM consensus. The crawl lost 9 to 1. Method and numbers inside.
An in-place sort plus index-derived filenames produces output that builds, loads, and validates fine while pointing every record at the wrong file.
Model-returned URLs are untrusted input. A Go fetcher with dial-time IP checks, DNS rebinding defense, per-hop redirect validation, and page verification.
When one model fails to reply, counting it as disagreement poisons your retry budget and turns outages into verdicts. How to separate the two states.
The same fail-closed check is a safe skip when generating data and a destructive delete when auditing it. Here is how to move it safely.
Locking a parent row with SELECT FOR UPDATE does not protect its child rows under READ COMMITTED. Three traps it creates, and the fix for each one.
A generated file tracked in git and embedded in the binary can ship without being rebuilt. There is no error, just a class with no CSS rule behind it.
A hex picked from a macOS screenshot is in the display profile, not sRGB. Paste it into code and the color shifts. Here is the round-trip proof.
An always-loaded agent instruction file grew to 33,000 tokens. Here is how we cut it to a 10KB routing hub without deleting a sentence, and how we proved nothing was lost.
Focus-based dirty guards silently fail when a button writes to the input. Here is why the blur fires first, and what to use instead in live-updating forms.
Whisper's prompt parameter has a hidden 224-token limit. Overflow is cut from the front without warning, which can turn keyword biasing into a bug.
Two features bound the same shortcut and both fired. preventDefault cancels default actions, not other listeners. Here is the fix that survives the switch.
We ranked a domain vocabulary by corpus frequency and the top 20 came back all generic words. How a planned data gate kept a bad axis from shipping.
An agent that writes and ships posts on its own needs a gate that cannot be talked out of blocking. Two layers, one local and deterministic, one isolated.
A content filter reported clean for months. It was reading nothing. Four ways a check silently inverts into an approval, and how to make failure loud.
A feature failed every request from the day it shipped. One of four deploy targets was missing a credential, and the local environment hid it.
Pasting your secret-terms list into a leak-checking prompt exports the index of everything you protect. Split the job so the list never leaves the machine.
A worker read job results with GETDEL, so any payload it could not parse was gone for good. Treating that as "not ready yet" meant waiting forever.
A per-item done event closed the whole stream, and the client subscribed after starting the job. Both dropped the notification while the work itself succeeded.