# cs-paper-harness

> Evidence-bound, agent-first harness for CS graduate paper workflows in Codex
> and Claude Code.

## Canonical Pages

- Project knowledge: docs/project-knowledge.md
- Full AI-readable digest: llms-full.txt
- Student install guide: docs/student/install.md
- First paper from zero: docs/student/first-paper.md
- Continue existing paper: docs/student/continue-existing-paper.md
- Experiment artifacts: docs/student/experiment-artifacts.md
- Direct paper edit recovery: docs/student/direct-paper-edit-recovery.md
- Privacy and local state: docs/student/privacy.md
- Codex guide: docs/student/codex.md
- Claude Code guide: docs/student/claude-code.md
- Troubleshooting: docs/student/troubleshooting.md
- Protocol overview: docs/protocol-overview.md
- Release contract: docs/v1-release-contract.md
- Release operations: docs/maintainer/release-operations.md
- Release evidence ledger: docs/maintainer/release-candidate-evidence.md
- Manual student acceptance playbook: docs/maintainer/manual-acceptance.md

## Install Surface

- Runtime package: `pipx install cs-paper-harness==1.0.0`
- Project setup: `cph-install --scope project --agent codex --project /path/to/paper-workspace`
- Global agent setup: `cph-install --scope global --agent all`
- Combined setup: `cph-install --scope both --agent all --project /path/to/paper-workspace`
- Agent package layouts: `agent-packages/codex/` and `agent-packages/claude-code/`

## Runtime Boundary

The `cph` runtime is limited to deterministic workspace, manifest, assembly,
paper-sync, and handoff operations under:

- `cph init`
- `cph doctor`
- `cph manifest`
- `cph assemble`
- `cph handoff`

Runtime must not judge novelty, paper quality, evidence support, final audit
results, or keep/discard/defer decisions for direct `paper.md` edits.

## Privacy Boundary

Real paper workspaces should keep `.cs-paper/`, `AGENTS.md`, and `CLAUDE.md`
private and gitignored. The committed `.cs-paper/` state under
`examples/self-referential-demo/` is synthetic demo-only state.

## Claim Boundary

Do not claim real-world effectiveness beyond evidence. CPH provides workflow
guardrails and deterministic state checks; it does not guarantee paper quality,
guarantee novelty, or provide autonomous research capability.
