# Goal Spec: VeriHarness V0.1 Board Summary

Status: complete and frozen on 2026-07-13. The published page, case1 trajectory integration,
bilingual prompt view, and compressed progress snapshot now preserve the V0.1 research boundary.

## Goal

Publish a dedicated `VeriHarness V0.1` board page that makes the harness research design,
four experimental conditions, independent four-dimension evaluator, current case1 audit,
versioned system prompt, and design philosophy understandable to a first-time reader.

## User Value

Researchers can inspect the current VeriHarness design from the main project board without
reconstructing it from implementation notes, while retaining the caveats needed to avoid
misreading the frozen case1 run as a valid verification-effect result.

## Context

- The main board already lists VeriHarness under `内部迭代`, but currently links to this README.
- V2 and V3 use dedicated root-level `sp-*.html` pages and generated board/navigation entries.
- Four prompt modules are assembled by `src/veriharness/prompts.py`.
- The current case1 audit scores are historical pipeline evidence; frozen P1 agent model calls
  did not receive image bytes.

## Scope

### In Scope

- A dedicated Chinese board page with the English prompt source preserved verbatim.
- The H0P0/H0P1/H1P0/H1P1 2x2 design and concrete condition names.
- The independent Canvas/Text Geometry/Fresh Visual Bank/Structure Audit evaluator.
- The frozen case1 score table with prominent validity and matcher caveats.
- Detailed design principles, board generation integration, and public-safe deployment sync.

### Out of Scope

- New experiment runs, score changes, verifier calibration, or RL training.
- A full watermark/OCR event replay or raw trajectory viewer.
- Publishing raw transcripts, run workspaces, proxy configuration, credentials, or private
  model responses.

## Constraints

- Reuse the existing board shell, navigation, CSS language, and native `details` interaction.
- Keep the two four-part structures separate: experimental treatment versus outcome measure.
- Preserve prompt text exactly from the four versioned source modules.
- Present the case1 scores as a pipeline audit, never as general causal evidence.
- Register the page in the board generator so a future sync cannot restore the README link.

## Assumptions

- The page and complete prompt modules are approved for the public board.
- Page prose is Chinese; prompt source remains English.
- The frozen case1 result stays immutable and later reruns will be reported separately.

## Acceptance Criteria

| ID | Criterion | Evidence Required |
| --- | --- | --- |
| AC1 | VeriHarness has a dedicated active page under `内部迭代`. | Generated index, attempts page, and sidebar all link to `sp-veriharness.html`. |
| AC2 | The four conditions and four evaluator dimensions are distinct and complete. | Manual content review against the manifest and offline evaluator weights. |
| AC3 | All four prompt modules are complete, collapsible, and traceable. | Snapshot comparison against the prompt source files and prompt assembly order. |
| AC4 | Case1 scores appear with the missing-image, transport, and text-binding caveats. | Manual review against the frozen post-run audit. |
| AC5 | Public sync includes only the page, summary, and prompt sources. | Inspection of `deploy-board/public` after sync. |
| AC6 | Desktop and mobile layouts are readable and error-free. | Playwright screenshots at 1440x900 and 390x844, zero page errors, and no document overflow. |

## Definition of Done

- The dedicated page, Markdown summary, generator override, deployment allowlist, and scoped
  styling are implemented.
- Board sync is idempotent and browser QA evidence is recorded.

## Stop and Ask If

- A prompt source, evaluator score, or public/private boundary differs from the frozen sources
  cited in this spec.

## Open Questions

- None blocking.