mdskills are context injections — AGENTS maps and skill recipes handed to the model. We make the database the single source of truth for that text, compose it into bundles, and pin every bench result to the exact bytes that produced it — all without the harness ever knowing the database exists.
The data model — fragments → bundles → blobs
atomic
Fragments
Small reusable .md/.txt snippets — an AGENTS repo-map, a "fix the failing test" recipe. Versioned + hashed in the DB.
composed
Bundles
Ordered combinations of fragments. Spin up as many as you like from the same library — none, agents, skills, both, or any new mix.
exact bytes
Blobs
The concatenated text a bundle actually injects, stored by sha256. This is the immutable record of "what the model saw".
The bench flow — materialize → vanilla harness → stamp
Why it's built this way
The harness must run in production, where there is no benchmark database — just real .md files on disk. So the DB never goes into the harness. A wrapper generates the files in front of the run and tags the results after; pull the wrapper away and the harness behaves identically off whatever files are present.
The payoff: every score on the board carries the sha of the exact injection text that produced it, so we can ask "which combination of fragments lifted which model on which bench" and get a precise answer — the clean way to measure the v1.7 + mdskills lift without the harness ever depending on the database.
IMR R&D · harness note · v1.7 + mdskills + the orchestration wrapper, June 2026.