DSH-Reforge

DSH-Reforge documentation

Project facts: real-file preview, explicit goals and complete five-arm comparison. Supports32/48 inputs;96/96 agreement is conditional on successful projection. Wrong offers9→0 with extra target/goal inputs, correct offers7→11 while coverage falls33.3%→22.9%; not default DSH or Agent benefit.

Read-only inspector: screenshots, scoped interaction comparison and runnable zero-model demo. Local search, outcome filters and candidate details use the existing bounded state; scripted clicks are not human or Agent benefit.

Applicability prototype: evidence-bound contracts and three-state rejection. On48 synthetic holdout queries, wrong offers13→1 with extra authored task facts; precision45.8%→92.3%, coverage50%→27.1%. Not natural-language extraction or default DSH rollout.

Developer evaluation: frozen retrieval comparison, per-query explanations and index performance. Offline synthetic results are separate from downstream Agent benefit.

Live integration: bounded workbench decisions and three-arm service-chain measurements. Every request rechecks eligibility; the text policy is unchanged.

Multilingual prototype: 120-query BM25/dense/hybrid comparison and complete errors. Local vectors improve synthetic positive recall but retain substantial wrong offers; default DSH is unchanged.

Actual results: automatic learning and all pilot metrics. The real-model automatic path passed; both retrieval arms abstained on held-out tasks, so no memory-treatment effect was measured.

中文 · Version and release status

Start with the video, run the same case, configure read-only access, authorize learning separately, then inspect the results. Source verification, automatic-loop execution and future-task benefit are different kinds of evidence.

Goal Read
See the experience Two-minute video and acceptance notes
Run the same project without model charges Complete customer-form example
Configure and verify each first-use step Quickstart
Install and authorize automatic learning Installation guide (Chinese)
Check exact host support Version matrix
Troubleshoot common problems Bilingual FAQ
Upgrade, retain data and recover safely Data operations
Understand evidence limits Requirements, real-model pilot protocol, earlier evaluation
Participate or report an issue External pilot, issue forms, security reporting
Make a first contribution Contributing, newcomer tasks

Development: Architecture · ADRs · Readiness tracker · Development status · Historical roadmap · Changelog. Historical reports describe their own versions and are not current installation instructions.

Read or edit the source on GitHub ↗0.1.0-alpha.2