How these readings are taken.

Every 4 hours the agent reads 388 repositories and writes down what it sees. Nothing is inferred from a single reading. A repository's fork count is compared against that repository's own trailing average — never against another repository's — and a comparison is only made once there are 14 days of history to compare with. Before then the count is shown raw and marked forming, which means not measured yet rather than measured at zero.

A threshold crossed once is detected. It becomes confirmed only if it is still true on the next day's reading. That costs a day of speed and buys the thing that cannot be bought back: fork counts are trivially inflatable with throwaway accounts, and a single observation cannot tell a real surge from a manufactured one.

Sentences that interpret a reading are written by a language model, and are set in this typeface with a dashed edge so they never look like the measurements around them. Before publication every number in such a sentence is checked against the record it describes; if the model produced a figure that is not there, the sentence is discarded and a templated one is used instead. A sentence that is certainly true beats a fluent one that might not be.

It is not a survey. The watchlist is 388 repositories chosen by hand. If something happens outside them, this instrument does not see it and has no way of knowing it missed it. No figure here supports a claim about open source as a whole, or about any field as a whole.

It does not measure quality, popularity or momentum. It measures fork counts, release tags, issue text, dependency manifests and declared model ancestry. A repository appearing here is an observation, never an endorsement, and never a judgement that the project is good, bad, safe or unsafe.

A finding is a co-occurrence, never a cause. That forks rose and a release followed is in the record. That one caused the other is not, and is never claimed.

It is not real-time. Readings are 4 hours apart at best, and a scheduled run can be delayed. The header always states when the last successful reading actually happened.

A detector set above anything that happens in the real world produces the same empty page as a quiet month, and looks equally healthy. So every observation is compared against its threshold whether or not it crosses, and the distribution is recorded the same day and published on the index under Our record.

If a detector reads never approached, nothing has come within half its threshold across the whole window. That is not a statement about the repositories being watched. It is a statement that this instrument is set too high, and it is published because the alternative is asking readers to assume otherwise.

The same section states how many confirmed fork findings were followed by a release from the same repository within 7 days. That measures co-occurrence rather than accuracy, the sample is small, and it is published whatever it says.

This is funded by a token on Robinhood Chain. Trading it pays a fee, most of which goes to whoever launched the pool. Holding it grants nothing here, there is nothing to connect a wallet to, and no reading is behind it. It has not launched.

The conflict worth stating. 75 of the 388 repositories watched — 19% — are crypto and blockchain infrastructure, the same field this project's funding lives in. That share was chosen by hand and nothing measured it. It is disclosed here because a reader should not have to discover it themselves, and because it is exactly the kind of thing that costs a project its credibility when someone else finds it first.

Corrections work the same way. Findings are append-only: a wrong one is superseded by a correction that appears in the same place with the same prominence, never deleted. Every demand cluster published on the first live run was wrong and all of them were retracted that way.

An MCP server over the same readings, so an agent answers "is this dependency healthy" from a measurement taken today rather than from training data a year old.

{ "mcpServers": { "readout": { "url": "https://readout-7pt.pages.dev/api/mcp" } } }

Four read-only tools: check_package, check_stack, compare_repositories, search_repositories. No key, no account, no quota. Every result carries the limits above with it, because a scorecard pasted into a code review without them is a claim this project does not make.

Most days nothing crosses a threshold, which makes this a poor page to check daily and a reasonable one to subscribe to. The feed at /feed.xml carries confirmed findings only, and carries them at the same moment they appear here.

If you would rather work with the data directly, every bundle behind every page is a static file: /data/index.json for the current state, /data/stack-index.json keyed by package, /data/compare.json one row per repository. No key, no rate limit, no account.

For coding agents. The same readings are served over the Model Context Protocol at /api/mcp, with four read-only tools. An agent asked whether a dependency is healthy otherwise answers from training data a year old; this lets it answer from a reading taken today. Every result carries these limits with it, because a scorecard quoted into a code review without them is a claim this project does not make.