clawock / site

# clawock

> AI argues. Code settles. Even the losses stay on the page.

clawock is an agent-native, harness-agnostic investment decision-workflow
engine. An external agent runtime (Claude Code, Codex, OpenClaw, DeepSeek
Harness, or your own) owns the model, conversation, memory and tools; clawock
owns the decision contract: certified evidence, a mandatory opposing case,
deterministic money and FX reconciliation, and a public scorecard the model can
never grade itself.

A real Hong Kong + US account has run it continuously: every judgment settled
by Python; the live account return is published on the dashboard — currently
negative, every loss on the page, every number reproducible from
`clawock audit-resettle`.

## Key pages

- [Complete agent reference](https://kcnyu.github.io/clawock/llms-full.txt): overview, FAQ and external-agent invocation protocol in one Markdown file; generated from the maintained documents on every Pages build
- [FAQ in Markdown](https://kcnyu.github.io/clawock/faq.html.md): the same questions and answers as the human FAQ, without navigation or scripts
- [External agent invocation protocol](https://kcnyu.github.io/clawock/docs/architecture/runtime-protocol.md): prepare, validate, publish, evaluate and bounded improvement; clawock does not call or own the model
- [GitHub repository](https://github.com/KCNyu/clawock): source, issues, PRs
- [Live dashboard](https://kcnyu.github.io/clawock/): positions, risk, self-graded scorecard
- [Daily briefs](https://kcnyu.github.io/clawock/briefs.html): published morning reads (Chinese; the bilingual part is the data coverage behind them)
- [Validation ledger](https://kcnyu.github.io/clawock/#reflect): what was tested and what failed (Reflect tab; `assets/data/evidence.json`)
- [FAQ](https://kcnyu.github.io/clawock/faq.html): questions new users ask
- [简体中文 README](https://github.com/KCNyu/clawock/blob/master/README.zh.md)

## Install

```bash
python -m pip install clawock
clawock workflow install investment-decision --workspace ./my-decision
clawock init ./my-decision --workflow investment-decision
clawock run prepare --workspace ./my-decision
```

Or throw the repository URL at any agent and let it follow the three-step
skill flow (prepare → write `decision.json` → publish). A no-model end-to-end
proof runs with `bash examples/cli/minimal-run/run.sh`.

## For DeepSeek Harness users

An official-style skill package is available: `dsh plugin --profile web add clawock-dsh`
(published to npm as `clawock-dsh`). The same decision contract works from a
pure CLI, an OpenClaw skill, a Claude Code instruction, a Codex AGENTS.md, or a
DSH agent — see [examples](https://github.com/KCNyu/clawock/tree/master/examples).

## What makes it different

- The model can never grade itself: LLMs propose, Python settles.
- One thesis counts once; repeated restatements collapse into one episode.
- Factors need out-of-sample validation before they influence decisions.
- The ledger must reconcile before anything is published.
- Every number on the README reproduces from `clawock audit-resettle` — if it
  doesn't match, the project loses.

## Honesty policy

The live scorecard keeps every eligible result, losses included, including the
fact that active recommendations have not beaten buy-and-hold. This repository
is a personal record and portable workspace — not investment advice, not a
copy-trading service.