Woojay Poynter

Dad, blah blah blah, ramen, blah blah blah, Kubernetes, blah blah blah.

[Woojay Poynter @w00jay]

Same model. Same tasks. Same temp=0.

I ran my LLM eval rig twice, a week apart.

qwen3.5-9b on research/finance_hard_01:
Apr 18: 0.90
Apr 24: 0.60

Thirty percentile points of pure judge variance. No model change. No prompt change.

The takeaways ↓

[Niall Harbison @NiallHarbison]

When she was taking her last breaths I promised her legacy would never be forgotten.

Tina’s hospital for sick street dogs is flying up ❤️

[がとりんぐ三等兵: Gatling3 @gatring3]

プラレールガン:Plastic railgun

[Marysia @marysia_cc]

Jean Mallard

[ArtButMakeItSports @ArtButSports]

Lamentation (flipped, detail), by Gerard David, 1515-23, 📸 by Roberto Schmidt

[ArtButMakeItSports @ArtButSports]

Composition VIII, by Wassily Kandinsky, 1923, (rotated, detail), 📸 by @ElsaGarrison

[derek guy @dieworkwear]

notice the outfit has shape and drape

[World Central Kitchen @WCKitchen]

WCK founder Chef José Andrés has arrived in Venezuela to join our teams on the ground. We're serving thousands of meals every day alongside impacted communities and scaling up to reach as many people as possible. ❤️ #ChefsForVenezuela

[Chef José Andrés 🕊️🥘🍳 @chefjoseandres]

Today on Anthony Bourdain’s birthday, let’s honor Tony by sharing a meal…with a friend or with a stranger…and making sure they’re okay. @wckitchen we are honoring him in a way by being next to the people of Venezuela after the earthquakes. Sharing food and hope. #BourdainDay

[Ramin Nasibov @RaminNasibov]

Each circle has a different story

[Woojay Poynter @w00jay]

@HqFhj +1

[INK @0xInk_]

the 2.5D illustrator

[Den Beauvais @DensLairStudio]

Godzilla 2026 painting done! 30"x40" oils on canvas. Available, message me if interested. #godzilla

[Woojay Poynter @w00jay]

Full writeup — GGUF metadata diff, chat-template diff, both run reports — in the article:

https://t.co/sZQUzSMCQp

Next post: running this as a Kubernetes CronJob. From "eval rig on bare metal" to "platform that runs itself at n=40+ without babysitting."

[Woojay Poynter @w00jay]

https://t.co/uppT7669Ud

[Woojay Poynter @w00jay]

Same model. Same tasks. Same temp=0.

I ran my LLM eval rig twice, a week apart.

qwen3.5-9b on research/finance_hard_01:
Apr 18: 0.90
Apr 24: 0.60

Thirty percentile points of pure judge variance. No model change. No prompt change.

The takeaways ↓

[Woojay Poynter @w00jay]

Two more findings two runs unlocked:

→ qwen3.5-27b: 0.62 composite both runs. qwen3.5-9b: 0.62 both runs. But 27b takes 3× the wall-clock time. Not earning its size.

→ coding_artemis_medium_01: 0.42 every model, both runs. coding_mcp_hard_01: 0.25 every model, both runs.

When every model on a task scores the same across runs, the task is broken — not the models.

[Woojay Poynter @w00jay]

https://t.co/uppT7669Ud

[Woojay Poynter @w00jay]

@dklineii This is a great template. Thank you for sharing.

[Woojay Poynter @w00jay]

https://t.co/o3zuIxHAtU