Upweight

The Hacker News front page, weighted your way. → upweight.vercel.app

Hacker News shows everyone the same front page in the same order. Upweight scores each story on six dimensions and hands you the sliders, so the ranking is yours.

---

The one idea

An AI model reads every story and answers eight fixed-answer questions about it. The answers come back as numbers, not sentences.

That single difference is the whole project. You can multiply a number, add it to another, subtract it, decide it counts double. You cannot do any of that to a paragraph. So the ranking policy never has to live inside the model: it lives in the browser, in about ten lines of arithmetic.

By the time you open the page the model has already finished. Dragging a slider asks nobody anything. Turn off your wifi and the list still reorders.

composite = Σ (w_d ÷ 100) × score_d  ÷  Σ |w_d| ÷ 100

The model used is Jev, TypeSafe's System One model. It takes natural language like an LLM but returns typed answers with calibrated probabilities instead of text.

The six dimensions

	asks
Technical depth	How much substance is here for a working engineer
Drama	How much conflict is in the discussion
Practical utility	Could you use this within a week
AI slop	How much is low-effort AI hype rather than real work
Novelty	How new is this to someone who follows the field
Career relevance	Does it bear on hiring, pay, or where the industry is going

Weights are bipolar, -100 to +100, so a dimension can be actively penalised rather than merely ignored. Two further questions return yes/no probabilities and become tags: primary source and rage bait. They stay out of the ranking on purpose, because they are not degrees along a spectrum and a slider would be the wrong control for them.

---

What measurement changed

Three things in this repo exist because a measurement contradicted the plan. The scripts that produced them are in upweight/scripts/ and can be re-run.

1. Four of six dimensions were guessing

The original design sent the title, metadata and the top five comments. No article. But four of the six questions are about the article, so they were answering from a headline.

scripts/experiment-state.ts scored the same stories with and without article text:

dimension	without article	with article
technical_depth	0.02	0.73
practical_utility	0.14	0.77
ai_slop	0.61 at 0.26 confidence	0.27 at 0.81
drama (control)	0.39	0.40

drama moving 0.01 is what makes this credible rather than noise: drama reads the comments, which were there all along.

2. A confident wrong answer is worse than an honest gap

Some articles cannot be fetched: paywalls, bot walls, JavaScript-only pages. The assumption was that those questions would simply come back uncertain.

They did not. The model looked at the empty space where the article should be and reported technical_depth 0.00 at 0.95 mean confidence. Perfectly true, completely useless, and confident enough to sink those stories under any tech-weighted preset for entirely the wrong reason.

So those questions are no longer asked. Dimensions carry available: false, the composite renormalises over what was answered, and the card says scored on 2 of 6.

3. The comment sample was measuring the calmest part of every thread

Drama maxed at 0.47 across a whole front page, and the top two rubric levels never fired. The cause was ingestion, not the rubric: HN orders kids by rank, so the top comments are the ones people agreed with. We were sampling the calm and asking how heated it was.

Now: 8 top-level comments, plus replies from the 3 most-replied threads, nested so back-and-forth is visible. Reply count is a cheap proxy for contention and costs one field rather than a fetch.

	before	after
max	0.47	0.65
mean	0.24	0.35
above 0.50	0	7

Also worth recording: a theory that did not survive

Firecrawl was added expecting cleaner extraction to reduce false ai_slop readings, on the theory that cookie banners and newsletter CTAs read as marketing language. Measured across stories where both extractors succeeded: mean ai_slop shift +0.005, mean confidence -0.016. The theory was wrong. Once the whole article is present the model already ignores the furniture.

Firecrawl stayed anyway, because it earns its place on coverage: each extractor rescues pages the other loses, and running both takes article coverage to 27–30 of 30.

---

How it runs

GitHub Actions (hourly)  ─┐
Vercel Cron (daily)      ─┴─→  /api/cron/score
                                   │
                    Hacker News API ├─→ 30 stories + threaded comments
                    Firecrawl ─┐    │
                    plain fetch ┴───├─→ article text
                                   │
                              Jev ─├─→ 8 questions × 30 stories, one request each
                                   ↓
                          Vercel Blob (private)
                                   ↓
                         page (server component)
                                   ↓
                    your browser: weights → ranking, no requests

Inference runs on a schedule, never on a request, so a traffic spike and an empty room cost the same. The hourly schedule lives in GitHub Actions because Vercel Cron is capped at one run per day on Hobby; the daily Vercel entry stays as a backstop, since GitHub disables workflows on repositories that go quiet for 60 days.

A refresh is built entirely in memory and only written if it survives. If anything throws, the previous payload is untouched and the page keeps serving it with an older timestamp. A refresh yielding fewer than 10 stories is refused outright, because that is far more likely an expired key than a genuinely empty front page.

Stack

Next.js 16, TypeScript strict, React 19. Five runtime dependencies: @typesafe-ai/sdk, firecrawl, @vercel/blob, @vercel/analytics, @vercel/speed-insights. No Tailwind and no motion library: the reorder is a hand-rolled FLIP in about fifteen lines.

Running it

cd upweight
pnpm install
cp .env.example .env.local   # add TYPESAFE_API_KEY and FIRECRAWL_API_KEY
pnpm smoke                   # one story, one call, prints the full answer
pnpm score                   # full front page, writes data/snapshot.json
pnpm review                  # per-dimension extremes and what each preset surfaces
pnpm correlate               # dimension independence check
pnpm dev

pnpm dev works with no API keys at all: the committed snapshot is the fallback, and the page renders from it.

script	what it is for
pnpm smoke	Cheapest possible check that the key and question set work
pnpm score	A full scoring run
pnpm review	Judging whether the rankings are defensible
pnpm correlate	Confirming no two dimensions are the same slider twice
pnpm test	83 tests

Layout

docs/          the PRD, the implementation plan, and the prototype that preceded the app
upweight/
  lib/questions.ts   the eight question definitions. the file that matters most
  lib/jev.ts         the only code that calls the model
  lib/composite.ts   weighting and ranking
  lib/article.ts     Firecrawl, then plain fetch
  lib/hn.ts          ingestion, including the comment sampling
  scripts/           the experiments behind the findings above

docs/ is worth a look if you want the reasoning rather than the result. The PRD records what was decided before any code existed, and a "Phase 1 outcome" section records which of those decisions measurement later overturned.

---

Honest limitations

• Calibrated does not mean correct. Typed output guarantees the interface, not the truth. The rankings here were reviewed by hand against a live front page and found defensible; that is a judgement, not a benchmark.
• ai_slop has little range on a good day. HN moderates well, so the funniest slider often barely moves the list. Deep tech versus Max drama separates far more.
• Drama's top rubric level never fires. It describes an entrenched personal pile-on, which HN genuinely rarely produces. The rubric may be describing something that does not exist here.
• Article coverage is 90 to 100 percent, never guaranteed. Stories without one are ranked on two dimensions and say so.
• Dimensions correlate. All pairs sit under r = 0.8, but technical_depth and novelty reach 0.70. "Good story" is a strong common factor.

Credit

Built on TypeSafe (Jev) and Firecrawl.