clawock

AI argues. Code settles. The losses stay on the page.

Live dashboard  ·  Daily briefs  ·  Evidence  ·  简体中文

_{“The market doesn't care how confident the model was.”}

127	881	144	43	5	0
days live on a real HK + US account	decisions on the public ledger	episodes settled by code	data modules across 8 layers	agent harnesses, one contract	scores the model wrote for itself

_{Real positions, real P&L — −17.30% since day one, published exactly as it is — graded in the open. Numbers and previews refresh weekly; the live dashboard updates through the trading day.}

Every trading day, clawock turns raw market information into decisions that get graded:

• Collect. 43 fetch and compute modules across 8 layers: quotes, SEC and HKEX filings, capital flow, bilingual news, Reddit and influencer feeds, with multi-source fallback. Python fetches; the model only reads the assembled context.
• Compute factors. Quant factors, cross-sectional ranks, peer residuals and a trend × volatility leverage dial, all computed deterministically in Python.
• Backtest. A factor's clustered bootstrap interval has to clear 50% before it may influence a decision; the cross-sectional layer is pre-registered; the leverage dial is scored out of sample. What fails is published on the evidence page.
• Decide. Four analyst lenses, a bull and a bear, three risk voices and a judge argue over the same context and write plan.json.
• Settle. Python settles every decision against real prices. The model never touches its own score, and every result lands on the public scorecard.

The whole pipeline plugs into the agent you already use: Claude Code, Codex, OpenClaw, DeepSeek Harness, or your own → install and the full loop

What this is

This started as one account, not a package. A multi-agent desk debates the evidence on a real brokerage account with separate Hong Kong and US books and proposes trades; the account owner still places the orders. What comes out of that is the record: real positions, a growing decision history, and a public scorecard the model has no say in — not a get-rich bot, and not a copy-trading service.

clawock is the part of that desk pulled out to be reusable. Your runtime keeps the model call, the conversation, memory, tools, permissions and credentials. clawock adds the decision contract on top: certified evidence, a required opposing case, checked money and FX arithmetic, and outcomes linked back to the decision that caused them. It is files and a CLI, so switching harness leaves the contract unchanged. examples/ runs the same decision from a pure CLI, an OpenClaw skill, a Claude Code instruction, a Codex AGENTS.md, and a DeepSeek Harness agent.

To try it without installing anything locally, open a Codespace and run examples/cli/minimal-run/run.sh: a clean virtualenv, no credentials, no broker, and the same script CI runs against every published wheel.

What makes it different

• A workflow plugin, not another agent or harness. The runtime keeps its model, chat, memory, skills engine, tool loop and permissions; clawock owns only the decision contract, so it moves between runtimes unchanged.
• The loop continues after the answer. Evidence, the opposing case, thesis, decision, execution, and observed outcome share one lineage. Measured results can propose bounded parameter changes, but never silently rewrite strategy.
• Real money, graded in public. One live Hong Kong + US brokerage account, with a public scorecard that keeps every eligible result — the losses included, and the fact that the active calls haven't beaten buy-and-hold. Each published headline names the ledger slice, window and commit it was computed from, and clawock scorecard-provenance --check recomputes it from memory/decisions.jsonl — a re-graded row inside a published window shows up as a mismatch.
• The model can't grade itself. LLMs propose trades; Python settles them and computes the scorecard.
• One thesis, one episode. Repeated opinions on the same thesis count once. Each episode is settled from canonical vendor bars, with declared gap-fill rules when a session is missing.
• The ledger has to reconcile. A money-conservation check runs before every push; if cash, positions, and P&L don't balance, nothing is published.
• Built to keep running. Scheduled Hong Kong and US sessions produce the daily briefs and refresh the live dashboard through the trading day.

For the canonical EN/ZH rendering of every project term — composite, regime, DSR, CSCV, triple barrier, run card and the rest — see the glossary.

How it works

The product boundary is simple: the external agent reads and reasons; clawock owns the portable decision workflow and the deterministic truth around it.

The KCNyu deployment then applies that product boundary to one live portfolio. This second diagram is the deployed KCNyu desk, not the reusable package boundary.

Every trading day the system pulls fresh prices, FX, volatility, earnings and macro context plus news and social sentiment; hands that normalized context to a multi-agent debate; applies deterministic risk, schema, and ledger gates in Python; delivers a brief to WeChat; and updates the public dashboard.

The information layer

Reading the market is most of what the LLM does, so the widest part of the system is data collection. The repository catalogs 43 fetch and compute modules across 8 layers, with bilingual Hong Kong + US coverage — live quotes, SEC + Eastmoney filings, capital flow, earnings calendars, macro (VIX / DXY / 10Y), Reddit and news sentiment, and market-moving social feeds. Each brief consumes the subset relevant to that market and session. Collection stays broad; the decision layer stays constrained.

Coverage is bilingual, but it is not symmetric, and the asymmetry is in research breadth rather than in the basics. Quotes, fundamentals, news and cash-flow reconciliation all have real Hong Kong branches. Two research-breadth capabilities do not: same-industry peers are discovered automatically for US names and read from a curated map for Hong Kong ones (peer_discovery.py — the mechanism is verified, the flag stays off until the peer-residual rules are re-registered against the wider universe), and US trading halts arrive as a structured feed while a Hong Kong suspension arrives as an announcement that the triage rules mark for a human (mover_evidence.py). So: Hong Kong base coverage on par, Hong Kong research breadth behind US.

What each run actually receives

Collection is broad, but no run gets everything. Each scheduled job's preflight assembles only the blocks that job can act on, writes them to a context file, and the model reads that file rather than fetching for itself.

sources
  ──► preflight (Python, deterministic)
  ──► context.json
  ──► LLM prose
  ──► postflight (Python)
  ──► publish

Pre-open gets the most: full position truth, risk, signals, the evidence graph, research state, and it writes the day's plan. The open/midday/close runs travel light — a fresh quote, risk only when signals demand it. Intraday check-ins (every 30 minutes a market is open) sit in between: more signal detail, but no research production and no evidence-graph rebuild, because that is a daily artifact and would be stale by construction.

Influencer radar

The system scans eight sources across US and HK twice every trading day over a rolling 48-hour lookback window: Trump (Truth Social, first-party), Musk (news aggregation), Cathie Wood / ARK Invest (their published daily trades — ticker, direction, share count, ETF weight), Serenity (public Substack posts), and four media-proxied figures with no fetchable first-party feed — 段永平, 洪灏 (HK media), Michael Burry and Pelosi (congressional disclosures). An LLM then filters the noise and links what's left to actual holdings and sectors: stance (endorse / oppose), relevance, and a plain-language summary. Who said what, and whether it touches your book, is already sitting in the pre-open brief — nobody has to go scroll social media for it.

Each source carries its own candidate budget, so no single loud feed (Trump can post dozens of times a day) can crowd the others out of the LLM batch. What each source is and how fresh it is (a first-party post, a news proxy, a disclosed trade from 30–45 days ago) is kept per item and shown in the dashboard card.

A concrete example: in the scan of 2026-08-17 21:54 UTC, five Musk/SpaceX posts all matched real holdings (held_hits=5, the SPCH/SPCX cluster), and the next morning's brief carried it verbatim — 撞持仓 (5 条全中 SPCH/SPCX). The scan the following day's brief quotes (2026-08-18 21:52 UTC) found 1 post and zero holding hits — an empty result is published as an empty result, not skipped. Both entries can be checked against the published briefs of those two days. Misses go in the brief exactly as often as they happen.

How it decides

Analysis resolves into explicit, gated strategy decisions — and one stock can carry several at once.

• Several strategies, graded separately. core_position, risk_rebalance, intraday_t, event_trade, and tactical_entry can coexist on the same name, because a long-term thesis and an intraday trade can legitimately disagree. Each is graded in its own episode.
• Attribution-first. Every decision is tagged by its dominant driver, and that driver's edge is measured dynamically from the record — no hit rate is hard-coded into the logic.

Low-frequency add campaigns

Adding to a position needs two independent evidence families to agree: price-relative (factor rank plus curated-peer residual), point-in-time news (reliable positive surprise or accelerating attention), or a confirmed un-overheated 20-day breakout — not one moving average mistaken for alpha. Any two families authorise a capped exploration slice; validated authority still requires decision-usable evidence on both the price and the information side, so a price pattern can never promote a leveraged name. Negative information or peer-laggard evidence blocks it outright, sizing stays capped and tranche-based (a warming policy earns a small exploration slice, not validated authority), and a small rank wobble can't churn permission off and on.

• Falsify, don't confirm. In a risk-on tape the default is HOLD. A bullish story doesn't trigger a buy until it clears a disconfirming check and an "is this already priced in?" test on the last few days' move.
• Regime over timing. Leverage isn't timed; a 200-day-trend × volatility dial sets the cap. The backtested lesson: the edge was in de-leveraging in the wrong regime, not in calling tops.

The debate

The daily deep brief runs a structured multi-agent debate, adapted from TradingAgents for separate Hong Kong and US books. More agents isn't the point: the protocol demands an opposing case, and the Judge attributes each resolution to a named strategy frame.

• Analyst lenses. Fundamental, technical, sentiment, and sector-rotation agents read the same context and merge into one table. Every claim must cite numeric context.
• Bull vs Bear. Two researchers build opposing cases, each citing concrete analyst data points. The protocol asks them to genuinely disagree on at least one position and to record it, so unanimous agreement reads as a flag rather than evidence.
• Devil's advocate. The Bear researcher is assigned to name and attack the session's strongest consensus view — never the weakest — so a lopsided bull tape cannot go unchallenged by cherry-picking, and the attack it lands is written down.
• Risk voices + a Judge. Aggressive, Conservative, and Neutral each argue their corner; the first-mover voice rotates every four trading days so the framing cannot calcify. A Judge weighs them, names the strategy frame driving each decision, and resolves the argument into plan.json — which enters the next session's grading pipeline.

The public scorecard

Every call is settled mechanically and published — wins, losses, and the cases that can't be graded. Nothing is hand-tuned after the fact.

1. Record — the model submits a versioned decision with its strategy, condition, regime, size, and confidence. The authoritative ledger is memory/decisions.jsonl.
2. Trigger — Python evaluates it against canonical unadjusted daily bars, counted on each market's own calendar. An unfinished session grades nothing, and a gap straight through a trigger fills at the open — never at a price that was never available.
3. Group — repeated calls of the same strategy collapse into one episode, so holding a position for five mornings does not manufacture five samples.
4. Grade & publish — code settles the outcome, scores it against a plain directional baseline, and renders it. Shut sessions, calls that need human evidence, and instruments that didn't trade are published as ungradeable — out of the win-rate denominator, but kept visible in the coverage count instead of silently dropped.

The model submits decisions; it can never write or amend its own evaluation. That isolation stops the desk from grading itself — it does not make the market data or the metric definitions correct. Treat the record as a diagnostic, not as proof of return.

_{Cumulative episode win rate against a 50% directional-hit line — how often the direction was right, not what it earned. The buy-and-hold comparison is the Shadow Portfolio under Holdings; this is a different question. Refreshed weekly by GitHub Actions; live figures are on the Holdings tab.}

What we tested, and what failed

前往 GitHub