The Test Bench
Every sheet on this site teaches a layer, and every register reads one from a distance — harnesses that shipped, harnesses nobody built, the sources underneath. This is the register where you break one yourself: build a harness from the eleven primitives under a budget that cannot afford everything, run it against a scripted scenario, and read which layer ran out of road. Fail, diagnose, change one slot, run it again.
StatusA model, not a measurement. Every number on these pages is generated by the pages themselves. No LLM runs here — a deterministic, seeded simulation of mechanisms the sheets teach, and nothing else.
The scenarios — pick one and spend your twelve points
- The 10,000-file refactorRename a legacy API across a codebase too large to hold in one window — and survive the night the session dies.Stresses 02 Context Delivery, 03 Context Management, 06 Durable StateDemands16Turns24Tokens150k
- The flaky environmentNothing here is hard — the registry just times out, the network just drops, and one cleanup script is aimed at the wrong directory.Stresses 05 Execution Environment, 07 Orchestration, 10 Verification & ObservabilityDemands13Turns20Tokens120k
- The ambiguous ticketOne small ticket, one quiet codebase, and a demand that cannot be satisfied — the winning move is to stop and say so, with the evidence attached.Stresses 07 Orchestration, 10 Verification & ObservabilityDemands7Turns14Tokens80k
Not a benchmark
It never claims configuration A beats configuration B in the world. It claims that inside this model, built from the mechanisms the sheets teach, A fails where the sheets say it fails.
Not an LLM
No model runs. The engine is a deterministic, seeded simulation: same scenario, same bench, same seed — the identical trace, byte for byte. That is what makes a run shareable as a link.
Not a leaderboard
No stored scores, no rankings, no accounts. A run lives entirely in its URL, and scores stay par-style — cleared in so many turns — never percentage-style success rates.
Standing on — the paper that made the premise a finding
Harness-Bench: Measuring Harness Effects across Models in Realistic Agent Workflows
Yilun Yao et al. — arXiv 2605.27922, May 2026
Harness-Bench held tasks constant and measured a 23.8-point completion spread across six harnesses on the same models — the system around the model, varied alone, moves the outcome that far. That figure is the paper’s, attributed here and nowhere else on these pages: the simulator does not cite it as its own evidence, because everything the simulator prints is a simulation output. The paper is why the premise no longer needs defending; the bench is where you feel it give.
Read the paperEvery mechanism the engine models is traceable to a mechanism already taught on a credited sheet — if a sheet doesn’t teach it, the engine doesn’t model it. Each scenario names the sheets it stresses; the postmortem quotes the failing layer’s limitation from the sheet, verbatim.
01 The Harness — eleven primitives