Live HN triage: a harsher critic than the community
216 live stories from HN's public API, classified by JEV as they arrived: is this a project, and how much would it reward a developer's attention? Compared against the community's own points.
What the data says
A taste knob, not a ranking
JEV keeps 6.9% of the feed and those items average 52 points vs 29 - but it keeps only 10% of the community's top-30. HN rewards drama and industry news; JEV's described criteria favor substance. Different taste, not broken judgment.
Continuous triage is effectively free
The whole 216-story feed re-classifies in 10 seconds for $0.004. Watching HN's new stream continuously costs ~$0.03/day - two orders below any chat LLM doing the same.
Described levels are the tuning surface
"High = significant release, deep technical work" is what makes the filter strict. Reword the levels, change the feed. Thresholds live in the prompt, not in code.
39.8% of HN's new stream shares a project
The is_project pass alone separates launch chatter from discussion - the same split the Ideas Hub pays an LLM for, here at $0.00002/story.
Bottom line
A calibrated decisions model can sit in front of a live feed all day for pennies, keeping the sixth of stories it judges substantive - 1.8x higher quality per item read. This page's data is the first run of the live-triage app.
How JEV does this
One call per story: title + link as state, two questions - is_project Choice and a worth-attention Score with three described levels ("high: significant release, deep technical work, or important industry shift"). The whole 216-story feed is one 10-second bulk run:
The JEV-flagged feed (top 15 of the run)
- Qwen Image 2.1 (704 pts)
- Grok 4.7 (27 pts)
- Python Workers are now generally available (18 pts)
- Tokenizers v1 (9 pts)
- How Blacksmith runs more than 10M CI jobs a day (5 pts)
- Show HN: All of FreeCAD 1.1.3, every workbench and FEM, running in the browser (3 pts)
- Can We Trust Open Source Models for Production Code? Eval Results (2 pts)
- Running an Optimal Trace (2 pts)
- Agent execution runtime engineered to run millions of sandboxes (2 pts)
- Step 5 Preview: StepFun's frontier model with 1M context and video input (2 pts)
- Mac Studio M5 Max review: The best reason to run AI locally (2 pts)
- Nginx Control API: View In-Memory Configuration and Reload via HTTP Requests (1 pts)
- Automatically Retrofitting JIT Compilers (1 pts)
- SAML: A Fractal of Bad Design (1 pts)
- Beyond the Hype: The Efficiency-Throughput Gap with GitHub Copilot (1 pts)
Explore the full run
| story | pts | verdict | att |
|---|