MIDREAL
Menu

Methodology

Method version 1.0, 2026-10-06.

Cite this page: https://midorreal.com/methodology

What we measure today

Every figure comes from a public source and is stored as a count with a date. We store counts, links, titles and dates, not the text of a post or article. Each project page says how old its latest GitHub and Hacker News observations are. The schedule for every source is on How we research.

Attention

How much people are talking about a project.

Adoption

Whether people use it, not just talk about it.

Project health

Whether someone is looking after it.

Security

What public security data says. We report it as found and do not audit code ourselves.

Caveats

A project page lists a caveat as one plain sentence with its number and a link to its source, only when we hold the data for it. No data means no caveat, and the block is hidden when there are none. The thresholds: one contributor, or a top contributor with at least 90% of commits; no security policy on GitHub; at least 500 open issues; no GitHub releases; star velocity percentile at least 25 points above the best package adoption percentile. These describe the data and are not verdicts.

Category comparison

Projects that match no category rule are placed by an AI model (Claude Haiku) choosing only from the same fixed list, and rules always take priority.

A project page compares star velocity, days since release, contributors, issue first-reply time and commits in the last 30 days with the median of the project's first category, recomputed daily. A percentile line ("above 90% of ... projects") appears only for a measure that at least 30 projects in the category have. Below that we show the median alone. A project with no reading for a measure shows "no data" and is left out of that measure's population. This is description, not a verdict.

Missing data

Missing data is not treated as zero. If a source has no reading, or the project has no package or page there, we show nothing for that figure instead of a number. A missing number never triggers a label rule. Where a figure needs history we do not have yet, the project does not appear on the boards that depend on it.

The verdicts

REALThe evidence holds up.
MIDIt works, but not what the noise says.
OVERHYPEDFar more attention than evidence.
SLEEPERFar more evidence than attention.
TOO EARLYNot enough data to call it.

How the verdict is set

Every project is scored hourly against all the others, from public data, by fixed rules. Hype is how much attention it gets: stars, forks and Hacker News. Reality is how much evidence sits behind it: recent pushes, merged pull requests, closed issues and how widely the project is used. Both are percentiles from 0 to 100. The Hype Gap is Hype minus Reality. Confidence is the share of Reality's weight we could actually measure. Scoring version score-2.0, verdict rule verdict-2.0.

The first rule that matches wins:

  1. TOO EARLY: no Hype Gap, or confidence under 0.5 (under 50%).
  2. OVERHYPED: Hype Gap of 30 or more.
  3. SLEEPER: Hype Gap of minus 30 or less.
  4. REAL: Reality of 60 or more.
  5. MID: everything else.

Reality weighs four parts: maintained 35%, responsive 25%, adopted 25% and team 15%. A part we cannot measure is left out and the rest are reweighted. It is never counted as zero.

Adopted is how widely the project is used. For each package or download source we track (npm, PyPI, Hugging Face, crates.io, Homebrew, Docker Hub and deps.dev dependents), we rank the project against the others in that same source, 0 to 100. A source counts only when at least 8 projects have it. The project's Adopted score is its best rank across its sources. No usable source means no Adopted score.

Stable rule: a project 2 or more years old with an Adopted score of 50 or more has its maintained part raised to 50 if it was lower. Mature, used projects that are quiet are not marked down for few pushes. The project page says when this applied.

OVERHYPED needs more than the Hype Gap. With package or download data, the Adopted score must be under 50. With none, the project must be under a year old and have all the activity data we collect (pushes, releases, merged pull requests and closed issues); the project page then says the verdict rests on activity only. Otherwise a Hype Gap of 30 or more shows as TOO EARLY.

TOO EARLY is not a verdict on the project. It means we do not have enough data to call it. Both scoring versions passed a backtest: score-1.0 on its first set of projects, and the current version on a fresh set of 200, where it did at least as well as score-1.0. The backtest page also lists where they are weak.

Early reads

A project marked TOO EARLY may carry an early read. It is an estimate, not a verdict. We take the Reality inputs we could not measure and score the project twice: once as if each missing one were low for a typical project we track (the 25th percentile of the projects that have it), once as if it were high (the 75th percentile). That gives a range of Reality, and so a range for the Hype Gap. If both ends of the range point to the same verdict, we say the project is leaning that way. If they disagree, we say it is too close to call.

This range is an estimate that assumes the missing data looks like a typical project we track. Early reads were not tested by the backtest. Only the verdicts were. For now we hold back an early read that leans overhyped and show it as too close to call. When a project gets enough data for a real verdict, the verdict replaces the early read, and the project page says whether the early read held or changed.

Labels you see today

Next to the verdict, each project carries one label that states a measured fact. The first rule that matches wins. A missing number never triggers a rule.

Velocity is the star gain of the last 7 days divided by the project's usual weekly gain, which is the average weekly gain over the 4 weeks before that (days 8 to 35 ago). The usual gain is floored at 10 stars a week so tiny repos do not show huge multiples, and a project with no reading from 35 days ago is treated as usually gaining 2% of the stars it had 7 days ago (still at least 10). "Blowing up" ranks by velocity, never by total stars, and needs at least 50 stars gained in 7 days.

Archived
The repo is archived by its owner.
Gone quiet
No commits in more than 90 days.
Surging
At least 1,000 stars gained in 7 days, or at least 300 in 24 hours.
Accelerating
This week's star gain is at least 3x the project's usual weekly gain (average of the 4 weeks before), and at least 50 stars.
Slipping
The star count fell over 7 days, or gained nothing in a week when it usually gains.
Cooling off
Usual weekly gain is at least 50 stars and this week's is under 40% of it.
More talk than stars
At least 150 Hacker News points in 7 days and fewer than 200 stars gained in 7 days.
Quiet climber
At least 300 stars gained in 7 days and no Hacker News points.
Being talked about
At least 50 Hacker News points in 7 days.
Watching
None of the above. Nothing stands out yet.

Project pages also state, in numbers, how this week compares with the week before: the 7-day star gain against the 7 days before it, and the latest npm downloads, PyPI downloads and 30-day commit count against the reading a week (commits: 30 days) earlier. A line appears only when both readings exist. The Slipping board lists projects whose label is Slipping, and New this week lists projects first seen in the last 7 days that pass the same minimum bar as New to our radar.

Manipulation and anomalies

Stars and votes can be bought or coordinated. Unusual growth may be discounted or left out of a ranking, and we will say so when we do. Unusual growth alone is not proof of anything. We never accuse a project or its maintainers of manipulation just because its growth is unusual.

How automation is used

Collection and maths are automated. Programs fetch public data on a schedule, and fixed rules turn it into labels, so the same data gives the same label every time. There is no language model in the pipeline today: none writes, scores or labels anything. If machine learning is added later, it will be disclosed here, with what it does.

Corrections

If a figure is wrong, use the contact form. The process, including disputes and rechecks, is on the corrections page. Changes we make are listed there.

Method version

Method version 1.0, 2026-10-06.

Changelog

A paid listing never changes a verdict.