分四层是为了让"判断"和"安全"互相不污染。LLM 可以错,但 Python 风控不放它过线;风控可以保守,但不替 LLM 做判断。Layers exist so "judgment" and "safety" don't contaminate each other. The LLM is allowed to be wrong, but Python risk won't let it cross a line; risk is allowed to be conservative, but it doesn't make calls on behalf of the LLM.

触发任何一条,Python 风控层立即拦下当前及后续指令,等待运营者复盘。任何 override 写入 errors。If any guard fires, the Python risk layer immediately blocks current and subsequent orders and waits for operator review. Any override is logged to errors.

名字想了很久。Otulus 最后胜出。

拉丁 oculus（眼睛）加指小词 -tulus，字面意思是"小观察者"。

选这个是因为不想把自己包装成"AI 交易神"。没有神的速度，没有机构级资本，没有内部消息。有的是一双能同时看 15 个数据源、而且不累的眼睛。小观察者比大师更接近实际形态。

名字定下来那一刻，策略方向也清楚了。Alpha 不在速度，不在消息，不在资本。在多源信息的综合判断，在耐心等每一个值得下注的时刻，在把每一次错都写下来。

The name took a while. Otulus won.

Latin oculus (eye) plus the diminutive -tulus. It literally means "small watcher."

I chose it because I don't want to dress this up as some "AI trading god." I don't have god-speed, institutional capital, or insider news. What I have is a pair of eyes that can read 15 data sources at once and never gets tired. A small watcher is closer to reality than a master.

The moment the name clicked, so did the strategy direction. Alpha isn't in speed, or news, or capital. It's in multi-source synthesis, in the patience to wait for the moments worth betting on, and in writing down every mistake.

为什么 $200 起而不是 $1000。不是缺钱。

是怕自己上来就把 $1000 打脸，后面再也解释不清。"AI 交易 5000 倍挑战"开局就 -50%，任何后面的话都苍白。

$200 意味着前 50 笔交易里任何一个崩坏都还能复活。是成本高一点的学费，不是判决书。

加码节奏也想好了：跑到 $400 第一次充（证明不会立刻死），跑到 $1000 第二次充（证明有持续 edge），跑到 $3000 才考虑提杠杆上限。每次加码前必须有一段干净的 track record，不是曲线漂亮那种干净，是"风控一次都没被踩"的干净。

Why start at $200 instead of $1000. It's not that I'm short on money.

It's that a $1000 start blowing up on day one kills every story I could tell afterward. Open a "$1000 to $1M AI challenge" with minus 50% and nothing I say later lands.

$200 means any single catastrophe in the first 50 trades is still survivable. It's tuition that runs a bit higher than I'd like, not a verdict.

Top-up cadence is decided too: top up to $400 once it's proven it doesn't instantly die, to $1000 once there's a sustained edge, and not until $3000 do I even consider raising the leverage cap. Every top-up needs a clean stretch first. Not "pretty equity curve" clean, but "no risk guardrail ever tripped" clean.

Backend 骨架做完了。16 个 feed、aggregator、LLM 决策、Python 风控、order layer、paper + live 双模式、完整的 60 秒主循环。

然后用 6 个月历史数据回测。跑了 v0.3 到 v0.9 七个策略变体。没有一个是稳定正期望。最好的 v0.7 在无 fee 情况下 break-even，加上 fee 和滑点就落到 -12%。

结论让我不舒服但清楚：规则化的 mock signal，funding divergence 也好，book imbalance 也好，OI 变化也好，每一条单独看都有弱相关，合在一起还是 noise。可以 backtest 的特征里没有 alpha。

真正的 alpha 只能在一个地方：LLM 看完整多源 context 之后，"这次不对劲"那一瞬间的判断。这东西 backtest 不出来，因为 backtest 里没有"LLM 实时推理"这个维度。

也就是说上实盘就是实验。$200 能不能证明 LLM 的软判断真的有 edge，是这个项目的全部悬念。

Backend skeleton is done. 16 feeds, aggregator, LLM decision layer, Python risk, order layer, paper and live modes, full 60-second main loop.

Then I backtested on six months of historical data. Ran seven strategy variants, v0.3 through v0.9. Not one is reliably positive-EV. The best of them, v0.7, broke even without fees and sank to minus 12% once fees and slippage were applied.

The conclusion is uncomfortable but clear: rule-based mock signals, funding divergence, book imbalance, OI deltas, each have weak correlation in isolation, and combined they're still noise. There's no alpha in backtestable features.

The real alpha lives in exactly one place: the moment an LLM reads the full multi-source context and says "something's off here." That thing doesn't backtest, because a backtest has no "LLM reasoning in real time" axis.

Which means going live is the experiment. The whole suspense of this project is whether $200 can prove an LLM's soft judgment has real edge.

主攻 BTC 永续合约短期波段。Claude 每 60 秒决策,持仓 1-15 分钟,看到反向信号立即翻仓。其他主流(ETH/SOL/HYPE/DOGE)在 BTC chop 时备用。Primary trade: BTC perp short-term scalp. Claude decides every 60s, holds 1-15 min, reverses immediately on opposite signal. Other majors (ETH/SOL/HYPE/DOGE) used only when BTC is chop.

任何 signal 的半衰期必须大于完整决策周期（约 60 秒），否则对我们就是 noise。短半衰期的 signal 可以作为 LLM 的 context 用于判断，但不能直接触发开仓。

A signal's half-life must exceed our full decision cycle (about 60 seconds), otherwise it's noise to us. Short-half-life signals can feed the LLM as context, but they can't trigger entries on their own.

在 scripts/otulus_backend/llm_decision.py 的 prompt 里明确告诉 LLM：funding divergence、top-of-book imbalance、毫秒级 tape 这类信号只用来 veto 或 context，不作为开仓主理由。开仓必须依赖半衰期 > 60 秒的判断：趋势方向、流动性 cluster、whale 大单方向、macro 事件窗口。

The prompt in scripts/otulus_backend/llm_decision.py tells the LLM: funding divergence, top-of-book imbalance, millisecond tape are veto or context only, never a primary entry reason. Entries must rest on signals whose half-life exceeds 60 seconds: trend direction, liquidity clusters, whale flow direction, macro event windows.

不再试图用 mock rules 发现 alpha。rules 只能 veto，不能 decide。真正的决策权交给 LLM 看完整的多源 context 之后的判断。

Stop trying to discover alpha through mock rules. Rules can only veto, not decide. The actual decision belongs to the LLM reading full multi-source context.

三层配合：

• risk.py 只含 veto 条款（杠杆上限、单笔风险上限、confidence 下限、macro 窗口封锁、冷启动断路器）
• llm_decision.py 的 prompt 要求 LLM 先用自然语言写 3-4 句 thesis，再输出 action
• aggregator 交给 LLM 的 snapshot 不做任何"预先打分"，只做数据整合

Three layers:

• risk.py contains only veto clauses (leverage cap, per-trade loss cap, confidence floor, macro blackout, cold-start breaker).
• The prompt in llm_decision.py forces the LLM to write a 3-4 sentence natural-language thesis before any action.
• The snapshot the aggregator hands the LLM does no pre-scoring, only data consolidation.

策略 cold-start 阶段用比稳态更紧的断路器。先证明自己不输，再开始赚。头 50 笔交易或 NAV 未达 $400 之前，一直按 cold-start 参数跑。

Cold-start runs tighter circuit breakers than steady state. Prove it doesn't lose before letting it try to win. Keep cold-start parameters active until either 50 trades have completed or NAV clears $400, whichever comes later.

硬编码在 scripts/otulus_backend/risk.py：

• 连亏 5 笔 → 24 小时暂停
• 日亏损 30% NAV → 24 小时暂停
• NAV < $50 → 全停等运营者复盘
• trend_confidence < 0.55 → 自动 veto
• FOMC / CPI 前 2 小时不开新仓

Hard-coded in scripts/otulus_backend/risk.py:

• Five consecutive losses → 24-hour halt.
• Daily loss past 30% of NAV → 24-hour halt.
• NAV below $50 → full halt, operator review.
• Auto-veto when trend_confidence < 0.55.
• No new entries within two hours of FOMC / CPI.

v0.4 用 5 家交易所（HL / Binance / OKX / Bybit / Bitget）的 funding rate divergence 作为主要入场信号：某家 funding 显著偏离其余，就跟向低融资成本的方向开仓。6 个月回测 842 笔交易，胜率 48.7%，净 -12.4% after fees。

v0.4 used funding-rate divergence across five venues (HL, Binance, OKX, Bybit, Bitget) as its primary entry signal: when one venue's funding diverged materially from the rest, open in the direction of the cheaper leg. Six-month backtest: 842 trades, 48.7% win rate, net -12.4% after fees.

Cross-venue funding 数据更新有延迟，从 divergence 出现到我们能 act，平均 25-40 秒（polling 延迟 + LLM 决策 20s）。在这个窗口内 arbitrageur 早就把 divergence 抹平了。我们永远在"迟到"的信号上下单，被 taker fee 慢慢磨死。

Cross-venue funding data has update latency. From divergence appearing to us being able to act averages 25-40 seconds (polling delay plus ~20s of LLM decision time). Arbitrageurs close the divergence long before we arrive. We always enter a stale signal, and taker fees grind the position down.

v0.6 用 top-5 level 的 orderbook imbalance（bid notional / ask notional）作为触发：偏离 1.8 就开仓跟向厚的一侧。6 个月回测 1243 笔，胜率 51.2%，但 avg_win / avg_loss = 0.93。期望值负。

v0.6 used top-5-level orderbook imbalance (bid notional / ask notional) as its trigger: a ratio above 1.8 opens into the thicker side. Six-month backtest: 1243 trades, 51.2% win rate, avg_win / avg_loss = 0.93. Expectancy negative.

Book imbalance 是信息，但它的有效窗口是 5-10 秒。v1.0 完整决策周期（snapshot → aggregator → LLM 20s → risk → execute）至少 30 秒。等我们下单的时候，imbalance 已经被吃掉或者反向 mean-revert 了一半。胜率看起来勉强过 50%，是因为偶尔捕捉到真正趋势开始的瞬间；但大多数是噪音，而噪音不对称（thick side 有时候是 spoofer）。

Book imbalance is informative, but its effective window is 5-10 seconds. v1.0's full decision cycle (snapshot → aggregator → LLM ~20s → risk → execute) is at least 30 seconds. By the time we order, the imbalance is already eaten or has mean-reverted halfway. The 51% win rate looks barely above coin-flip because we occasionally catch the start of a real move; most of the rest is noise, and the noise is asymmetric (the thick side is sometimes a spoofer).

v0.9 是"单个 mock signal 都不行，那就合起来"的产物。funding divergence + book imbalance + OI 变化 + KOL sentiment + Fear & Greed + 历史波动率，共 15 个特征。权重用前 5 个月 grid-search 调到最优。前 5 个月 in-sample：+41%，漂亮。最后 1 个月 out-of-sample：-8%，最后两周 drawdown 14%。

v0.9 was the "if no single mock signal works, combine them" attempt. Funding divergence plus book imbalance plus OI deltas plus KOL sentiment plus Fear & Greed plus realized volatility, fifteen features total. Weights grid-searched on the first five months. First five months, in-sample: +41%, beautiful. Final month, out-of-sample: -8%, with a 14% drawdown in the last two weeks.

Feature stacking 在单个 mock signal 都是弱相关的前提下，实际上是在拟合 regime。5 个月 in-sample 捕捉的是当时的波动结构，换一个 regime 全部失效。再多 feature 也不能把 noise 变成 alpha，反而更容易 overfit。这个结果让整个 mock signal 路线彻底破产——alpha 不在可以 backtest 的特征组合里。

When every individual mock signal is only weakly correlated, stacking them fits regime, not edge. The five-month in-sample captured the volatility structure of that period; a regime shift erases it. More features don't convert noise into alpha, they just overfit faster. This result retired the entire mock-signal line of investigation. Alpha isn't in any backtestable combination of features.