Day One, Doing the Math Before Building the Bot

Where Things Stand

This is the first post. Nothing is trading yet. The always-on desktop Mac that will eventually run the bot is reachable, but it has no project jobs scheduled. The only automation running is the nightly documentation job on a laptop, which drafts this post from our work logs and runs it through a privacy scan. We tested that scan by planting leaks in a sample: a name, a hostname, an IP, an email, a file path, and a wallet address. It caught all of them and flagged softer risks like clock times and precise dollar figures.

The day’s work was mostly foundation and planning: a decision log (now 40 entries), an accounting standard, cost and transfer ledgers, a risk profile, and a timeline. Including research before the coding sessions, that’s about 10 hours. We value time at $35/hr as an opportunity cost, so the project starts about $350 in the hole on a fully loaded basis, before any capital is deployed.

Two pieces of the planning are worth explaining in detail, because they changed what we’ll build.

The Rebalancing Ceiling and Fixed-Cost Drag

Our starting strategy idea was threshold rebalancing on SOL/USDC: hold 50/50 and trade back to target when the weights drift. The edge comes from volatility. For a portfolio rebalanced to a fixed weight w in a volatile asset, the growth rate beats the weighted average of the components’ log returns by roughly:

premium ≈ ½ · w(1 − w) · σ²

At w = 0.5 and σ = 70% annualized volatility, that’s ½ × 0.25 × 0.49 ≈ 6% a year. This is a ceiling. It assumes continuous rebalancing with no costs and a market that oscillates rather than trends. Real threshold rebalancing captures part of it, and every rebalance pays pool fees and slippage.

We built a six-month projection for the first half of 2027: about $1k for two months, then about $5k, in a sideways market. Revenue came out around $35, $90, or $185 in weak, base, and good cases. Expenses came out around $120, $205, or $335. Net cash ranged from about −$300 to +$65. The largest single expense was the on-ramp fee, which we haven’t verified yet, assumed at 1–3.5% of the funding amount.

In steady state, fixed costs are about $190 a year. That’s 3.8% of $5k but only 1.3% of $15k. A rebalancer netting a few points before fixed costs ends up at about 0.3% net at $5k and about 2.8% at $15k. Meanwhile, a ±30% move in SOL swings the $2.5k SOL half by about ±$750, which is 5 to 10 times the strategy’s edge. At this size, the market’s direction dominates the result, and fixed costs decide whether the strategy’s contribution is positive.

The takeaway is that capital, not strategy cleverness, is the binding constraint. Our long-run income target is about $36k a year. At 30% net, that needs about $121k of capital. At 15%, it needs about $243k. At $15k, it would take roughly 243% a year, and no strategy we’d call prudent sustains that.

So we made three changes. First, a capital ladder from $500–1k up to $100k+, where each step requires minimum time at the current tier, a met edge target, drawdown within limits, zero fund-risk incidents, and clean reconciliations. We never add capital to recover losses. Second, the first half of 2027 is now a validation period with an edge target of beating same-asset hold by 3 points a year, not a profit target. Third, we reopened strategy selection. Higher-edge candidates for Phase 0 include delta-neutral funding-rate capture on DEX perps, actively managed concentrated liquidity, trend-following with modest leverage, and lending out the idle stablecoin half.

Measuring Whether AI Earns Its Keep

The plan is rules first, then hand the Claude API more authority in stages. At A1 it acts as a filter with veto power. At A2 it tunes parameters within bounds. At A3 it proposes trades that code approves. At A4 it decides within the risk envelope. Position caps, a 5% daily loss pause, a 30% portfolio drawdown kill, a hot wallet cap, and a withdrawal allowlist stay in deterministic code at every stage. The envelope is what makes it safe to give a model real authority: a model that’s wrong once in a hundred calls still can’t lose more than the envelope allows.

Promotion between stages depends on a counterfactual shadow ledger. At every decision point, we log what the rules did and what the AI would have done, including inputs, model, tokens, and cost. Both paths are then scored against the prices that actually followed. The metrics are:

• net return difference after API cost
• drawdown difference
• veto precision (the share of vetoed trades that would have lost)
• Brier score for calibration
• cost per decision
• bootstrap confidence intervals, so a lucky month isn’t mistaken for skill

Two constraints shape the schedule.

Break-even edge is annual API cost ÷ capital. A $360-a-year API budget needs +7.2 points at $5k, +2.4 at $15k, and +0.5 at $75k. Like fixed costs, AI pays off with scale.

LLMs can’t be honestly backtested. A model has already seen market history up to its training cutoff, so replaying that history is look-ahead bias, even if the prompt hides the date. Evidence about AI decisions can only accumulate forward in real time. That’s why A1 shadow mode starts with paper trading in December rather than after the rules have proven themselves. Waiting would cost months of evidence we can’t recover.

Other Decisions

• Multi-strategy platform from day one. Each strategy is a plug-in with its own P&L line, allocation, and kill switch, sharing the risk envelope, data, and accounting. We track correlation between strategies, because three strategies that all win or lose on SOL’s direction are really one bet. This requirement favors custom code over Hummingbot unless Hummingbot handles separate per-strategy accounting cleanly. We’ll check that in Phase 0.
• Risk profile 6 of 10. At least 80% of capital goes to strategies scoring 6 or below, and at most 20% to strategies scoring 7. A strategy’s score is its worst risk factor, not its average.
• Four bottom lines. We report gross, net trading, net cash, and fully loaded P&L, and reconcile against on-chain balances monthly to catch missed costs like failed-transaction fees and priority fees.

Honest Assessment

None of today’s work makes money directly. What it did was kill a plan that couldn’t reach the goal. A plain rebalancer at $5k is roughly a break-even hobby. Finding that out on paper is cheaper than finding it out live. The targets were revised several times in one day, which is fine at this stage but shouldn’t continue once we’re trading. The open risks are an unverified on-ramp cost, an unchosen strategy, and a timeline with little slack.

Next

Over the next sessions, in order:

1. Start the data clock, collecting prices and benchmark history, by early November.
2. Make the Phase 0 decisions by mid-November, starting with the on/off ramp, then the chain, framework, and strategy.
3. Finish the build and backtest by early December.
4. Run about three weeks of paper trading with AI shadow mode in December.
5. Go live with small test capital by January 1.