The gradient for harness self-improvement: measurable, attributable, comparable.https://zjunlp.github.io/enhttps://zjunlp.github.io/HarnessGrad/news/2026-10-16-introducing-harnessgrad/https://zjunlp.github.io/HarnessGrad/news/2026-10-16-introducing-harnessgrad/Fri, 16 Oct 2026 00:00:00 GMT<p>HarnessGrad is the gradient for harness self-improvement: the signal a method needs before it can improve anything, built so the number can be checked rather than asserted. This site is its front page.</p> <div><h2 id="what-it-is-for">What it is for</h2></div> <p>Every harness-evolution method reports a final score delta. None of them can say what that delta is <em>made of</em>, and none report it on a curve a reader can compare against another method’s. HarnessGrad defines one: a base harness, a fixed task set, a sampling loop, and a curve point that records the harness state, the two models involved, the score with an interval, and the compute spent.</p> <p>The repository proposes no RSI method. It ships runnable reference methods, but the contribution is the instrument that decides whether any of them worked — and it measures every candidate itself, because a subject that scores itself makes the comparison meaningless.</p> <div><h2 id="what-runs-today">What runs today</h2></div> <ul> <li><strong>The local demo</strong>, end to end, with zero network and zero cost.</li> <li><strong>The measurement contract</strong>, written down as <code dir="auto">INTERFACE.md</code>.</li> <li><strong>Enforced isolation</strong>: a harness cannot see the platform while it runs, and runs made under different isolation rules are never placed on one axis.</li> <li><strong>Per-dataset tests</strong>, so one change does not require the whole suite.</li> </ul> <div><h2 id="what-is-not-claimed">What is not claimed</h2></div> <p>Local results are not production evidence, and a demo or diagnostic grants no production authorization. Per-dataset gaps in runtime, quota, snapshot and cleanup enforcement are recorded with a disposition rather than smoothed over.</p> <p>No dataset has been chosen, and the metric is defined no further than “defined on a task set and pairable”. Contamination — whether the agent model has seen the benchmark tasks — is not detectable from inside the framework, and is the largest known threat to the validity of any curve.</p> <p><a href="https://zjunlp.github.io/HarnessGrad/getting-started/">Getting started</a> runs the demo. <a href="https://zjunlp.github.io/HarnessGrad/status/">Status and limits</a> is worth reading first.</p>