What we have measured.
Explore the published studies behind Kin's claims. Each result identifies the tested build, the comparison, and the conditions it covers.
Does the repository work as described?
Can code and recorded changes be read back after moving a project into Kin?
Importing a project and keeping its code intact.
In this test, Kin kept the code intact, recorded a rename and retained the change after a restart. The result covers this particular edit.
The same function read back byte-identically through two separate paths, a three-occurrence rename shown to preserve behavior because the optimized compiler output matched before and after, and a native commit a daemon restart could not disturb.
Checking a published download.
The older download used in the saved-output case passed installation checks on the operating systems named in its release record. Those checks belong to that release.
The v0.7.6 archive that case ran on was built, published and checked by its own release run: public install proof on five runners across Linux, Windows and macOS, provenance checks on both published npm packages, and an anonymous install smoke for each.
Store size and determinism are kept with the code and checked on every change rather than quoted here. The store-size figures
Does it answer a useful question?
Can the answer help you find related code and check it against the source?
A saved answer about connected code.
Kin returned connections to a piece of code and showed where it was used. The case page carries the saved output, exactly as the run printed it. It does not establish that every connection was found.
One function, 25 entities reaching it, and a two-hop route from a caller with the call sites named. The case record contains the same run with every stream committed and hashed.
Website controls that share a piece of code.
React helps developers build website interfaces. Its text fields, dropdowns and text areas share a small piece of code. This written example shows the saved connections and source you can inspect before changing it.
AI agents meeting that archive for the first time.
AI agents tried the older download in separate environments and wrote reports about what they found. These are AI-agent trials; they do not establish how a person would experience Kin.
Three independent runs, each an AI agent (Claude Code, running Opus) with no prior knowledge of the archive, each inside its own 6 GiB container, and all three came back with their own written report of what they found.
Does it improve a workflow?
We have not yet shown that Kin improves a complete team workflow on the current release. The design partner program is intended to test a recurring task with a team: how much setup it takes, whether the answer helps and whether people keep using it.
An older comparison test is available, including a result where ordinary text search did better.
One pre-registered comparison benchmark exists today. It's measured on a Kin commit well behind the current release, and it publishes the axis where a grep-driven baseline came out ahead rather than leaving it out.
Explore the answers yourself. The online code explorer opens on a saved answer from a public project, without an account, and lists the requests it makes so you can check an answer yourself.
Every case above is published in full: the question it asked, what came back, the method it followed and the files behind it.
How a case is built, and what we have not measured The archive Go identities and saved history on v0.7.20 Dated released-byte case The saved-output case