dsh-evidence-gate

In DeepSeek Harness, agents routinely state "I read that file" or "I ran that test" with full confidence—without ever having done so. The conventional remedy is a prompt line that says "don't guess," which is neither enforceable nor auditable. dsh-evidence-gate turns conclusion discipline into a structural constraint: before an agent can state any load-bearing claim, it must call the evidence-gate tool. The claim is cross-checked against the session's actual tool-activity stream, and any guess that fails verification is downgraded from VERIFIED to UNVERIFIED on the spot, leaving a complete, replayable audit trail.

The mechanism works in four layers. A policy section injected before every model step defines an evidence hierarchy—tested, directly read, documented, or labeled assumption—so the model knows what level of support each conclusion requires. The gate tool assigns a verdict (VERIFIED, REFERENCED, UNVERIFIED, INVALID) to every claim. The cross-check layer matches not only tool-call arguments but also output anchors—file paths, URLs, distinctive words extracted from recorded results—and a graduated fallback ladder (cross-session lookup, existence floor) prevents honest claims from being falsely blocked. A per-session ledger persistently tracks blocked, converted, and floor-passed counts across restarts, queryable at any time via the report command. The switch, ledger, and policy text are all scoped to a single session; conversations that never opt in pay zero overhead.

It is built for developers who run multi-step reasoning agents in DSH, need auditable conclusions in their pipelines, or have grown tired of agents delivering confident-sounding fiction. All data stays local as a plain JSON file—nothing is sent anywhere.