dsh-evidence-arena

When you want to compare two or three AI models on the same coding task, what you really need is not another round of chat, but an isolated, reproducible, and auditable evaluation pipeline. dsh-evidence-arena fills exactly that gap: it lets two or three Builders solve the same task in isolated Git worktrees and independent DSH child runtimes, invisible to one another, so results reflect raw model capability rather than cross-context contamination. Each Builder can use a different Provider, model, credential reference, and system prompt, enabling true cross-configuration comparison.

Once the run finishes, independent Reviewers verify every gate in zero-tool, sealed-evidence mode. You can inspect per-file unified diffs with line numbers, launch a disposable loopback preview for frontend candidates, read runtime logs, and record a pass/fail/inconclusive human verdict. All evidence is packaged into a versioned JSON report that includes metrics, hashes, file metadata, gate results, and Reviewer verdicts while automatically redacting local paths, credential references, and child Session identifiers, making it safe to share across teams or externally.

If you are evaluating model differences, Provider strategies, or prompt variations on coding tasks, or if you need a reproducible and auditable evaluation chain to support technology decisions, this plugin offers a lightweight path that requires no patches to official DSH source and zero extra environment setup. It is well suited for developers and research teams working within the DeepSeek Harness ecosystem, as well as any engineer who wants an objective evidence trail in multi-model coding scenarios.