Skip to main content
Back to /proof

Earlier tests and results

These examples describe earlier versions of Kin. They remain available so you can inspect what was tested and compare the results with their original claims. Choose a case below; its saved files and limitations stay with it.

Dated cases

Kin ships v0.7.19 today. Every case below is pinned to the build it was captured on, which is what makes it dated: it says what happened on those bytes on that day, and it is not restated as a claim about the current release.

  • Dated released-byte caseCaptured 2026-09-08

    hiredis, migrated and edited on the exact public v0.7.4 release

    Measured against the bytes of one published release archive on one dated day, and pinned to that build rather than to whatever Kin ships now.

    Open the full case record

  • Recorded session on the shipped releaseCaptured 2026-09-09

    kin-demo, asked three questions on the exact public v0.7.6 release archive

    Captured on the bytes of the current published release archive, against a public repository at a named commit, and the clips on /proof are an animation of these exact streams.

    Open the full case record

The older benchmark

One older run, kept where it is.

Before this release we ran one pre-registered retrieval benchmark on a much older Kin commit. It is not a result about the build we ship, so it does not lead this page, and none of it is withdrawn or summarised into something friendlier. The full set is published in one place with its conditions, its artifacts and the axis where the grep-driven agent baseline came out ahead of us. That run has no public verifier: its harness and raw run directory are private, so what you are reading there is our own measurement. Separately we publish a full specification and a standalone checker for our merge-trust review benchmark, which is a different benchmark and neither specifies nor scores that run.

View the archived benchmark details
Why this is not a current-release claim
Measured on a Kin commit well behind the release line this site ships today; the current release's own case is on /proof, not here.
Scope and unfavorable axis
The full pre-registered set, including the axis where the grep-driven agent baseline came out ahead, is published in the package linked below, not summarized away here.
Independent check
Not recorded. That run's harness and raw run directory are private, so it has no public checker; the numbers in the package are our own measurement.
Where the numbers live
In the package and the two records below, with their conditions. This page restates none of them.
Open the public proof package Read the merge-trust specification

Inspect the archived package at the link above. It is the URL this site has always used for that package, kept stable on purpose so a citation made against it does not break. The specification link is for a different, comparison-benchmark result; it does not specify or score this one.

Preserved artifact paths

The files the archived package cites, listed by their own URLs. These paths are not republished or rewritten here; the list exists so each one keeps a place on this site. crawl-archive-artifacts.mjs reads this same list, and a lane runs it by hand before landing a change to this page. It is not part of the build, so nothing here claims these URLs were checked at the moment you loaded the page.

How a case is built names the rules a case follows before it reaches this index.