Claim-ledger comparisons

Comparisons

This page surfaces only claims with an explicit comparison field. It does not pool cohorts, rank configurations, or synthesize a winner.

Published comparisons

Each page lists the published results for one comparison, newest first, with links to their reproduction packs.

kogen-loop vs CodexHead-to-head tests of kogen-loop, Kogen's small stand-alone coding agent, against OpenAI's Codex command-line agent, using the same AI model at the same effort setting.Kogen vs the same agent used alone, on hard tasksTests of the full Kogen product, which drives another coding agent, against that agent used on its own, on hard multi-step repository tasks.kogen-loop speed, patches and connection routesSmall tests of changes to kogen-loop itself, and of how fast its connection to the model streams, with no second agent being compared.Test setup, rules and sandboxingFindings about how the tests themselves are run: written rules, changes to the test harness, and sandbox tools that confine the agent.Kogen shaping and delivery-build speedMeasurements of Kogen's own shaping step (turning a request into an agreed specification) and of where time goes in its delivery builds. Each result tests Kogen against itself or reads recorded logs; no other agent is compared.Kogen shaping componentResults about the step that turns a request into an agreed specification with executable acceptance checks, each measured against Kogen's own earlier behaviour.Kogen build componentResults about the step that implements an agreed specification and drives its acceptance checks to green, measured against Kogen's own earlier behaviour.Kogen review componentResults about the independent review of a finished change: which real defects it catches, which it misses, and whether it blocks correct work.Kogen verification componentResults about how Kogen runs checks and reports failures: timeouts, leftover processes and how precisely a failure is located.Kogen Context componentResults about the evidence packets Kogen Context gives a model before it works: whether the packet points at the right files, and whether a task finishes correctly, faster or with fewer tokens with the packet than without it.
No automatic poolingRound, task, venue, model, effort, and analysis differences remain attached to each claim. Follow every included and excluded record before drawing a comparison.

No claim-ledger row has a structured comparison field. Comparison pages remain status-only until such records are published.