T BOT is a self-calibrating, multi-surface trading system where Claude AI holds decision authority — running live on real money, recalibrating itself every night, and reporting numbers it measures, not invents.
The Problem
Algorithmic advantage has always lived on institutional balance sheets. Manual trading loses to speed. Static bots lose to regime change. Neither learns. T BOT is different.
Institutions execute in microseconds; retail reacts in minutes. The edge gap is structural, not skill-based.
Most retail bots run fixed parameters. Markets shift, win rates erode, operators panic-adjust on feeling, not data.
Managed funds charge management fees regardless of performance. You pay whether you win or lose.
The Solution
Fully isolated trading surfaces under a single Claude AI intelligence layer — each with its own capital, env file, state machine, and kill switch. A change to one cannot touch another. Profits compound into a passive long-hold sleeve.
Weather, economics, sports props. NWS observation-aware models + Claude signal generation through a 14-layer guard chain. NO-directional edge.
Setup-classified entries with conditional-Kelly sizing on real per-context edge. BUY-only after the model showed a structural short bias.
Fed funds, CPI, GDP. EV-gated + per-market re-entry caps after a churn fix; an EV-cooldown pre-filter skips known −EV markets before they cost research.
Long-only 200-day trend-follower across 27 large-caps; benchmark is buy-and-hold SPY. Goes live on IBKR funding.
Held to paper after an honest deep dive: the dip-buy edge was smaller than retail taker fees. Kept running to find a fee-clearing signal — capital isn't risked until it does.
Sweeps 25% of every winning trade into a buy-and-hold 70% index / 30% quality portfolio — the trading surfaces are the cash engine, this is the wealth store. Benchmark: SPY.
How it works
Every surface runs the same five-stage pipeline. No manual intervention. Each stage has an escape hatch. Claude holds decision authority within calibration-discipline guardrails.
Identify opportunities by volume, anomaly, and momentum composite
News, observations, filings — Claude synthesizes a per-market signal
Edge + confidence + Kelly sizing through upstream guards
Signed order placement, file-lock dedup, fail-closed safety, real fill
Resolution, P&L, Brier calibration; nightly auto-rebalance
# Auto-calibration engine # strategy dimensions update from data nightly FOREX_DIRECTION_FILTER=*:BUY FOREX_MIN_EDGE=0.15 ← engine-managed ST_DIRECTION_FILTER=NO ST_MIN_CONFIDENCE=0.80 ← data-proven # Shape engines A–H cover every # calibration type: weight / cap / # threshold / set / boolean / # time-window / blocklist / ratio CALIBRATION_APPLY=true SAFE_RESTART_GATE=unit tests pass
The Holy Grail
Strategy dimensions update autonomously from data every night — no human touches a parameter. The model learns its own edge (FX is long-only, ST is NO-only) and stops betting where it bleeds.
Transparency
Most trading products are black boxes. T BOT measures itself in public: every figure in our documents and dashboards is regenerated from live trade logs — not typed by hand, not invented.
Allocations, the revenue cone, and the realized ledger auto-inject daily from the live snapshot. The doc can't drift from reality.
Each calibration change and trade carries its reasoning to an audit ledger — queryable, reversible, never silent.
We label every number: measured (from logs) or modeled (explicit, stated assumptions). Nothing is presented as real that isn't.
Business Model
One model, one sentence: subscribers pay a commission only on the profit they actually withdraw. If the bot doesn't make money, neither do we.
We earn nothing until you're above your previous withdrawal peak. Losses carry forward.
Commission crystallizes at withdrawal, not mark-to-market. Paper gains never trigger a fee.
Early subscribers pay nothing for the first 90 days.
No subscriptions, management fees, or seat counts. Pure profit-share.
Projections
We separate what we measure from what we model, and we never blur the two. The engine's edge is measured live; the business is modeled openly, with every assumption on the table.
| Subscribers | @ 8% return | @ 15% return | @ 25% return |
|---|---|---|---|
| 10 | $360 | $675 | $1,125 |
| 50 | $1,800 | $3,375 | $5,625 |
| 250 | $9,000 | $16,875 | $28,125 |
| 1,000 | $36,000 | $67,500 | $112,500 |
MODELED · annual commission = subscribers × $10,000 avg capital × net return × 30% withdrawal × 15% fee. The only input meant to be grounded in measurement is net return — and today it's early (FX shows a measured profit factor of 2.95 BUY-only over 79 trades; ST is in a measured drawdown the latest fixes target). So the return columns are planning scenarios, not validated baselines: the model shows the structure — revenue is linear in every input, with zero owed when the bot loses. Returns are the gate, and they're earned, not assumed. Offering to subscribers is RIA/custody-regulated; live deployment carries the compliance wrapper.
Security
Two external security audits completed; every critical and high finding closed. Defense-in-depth across independent layers.
Authenticated Origin Pulls — only Cloudflare reaches the origin; direct IP hits fail the handshake.
Secrets in an isolated vault, rotated vault→CLI→env. Never appear in shell, chat, or logs.
HMAC-SHA256 signed at generation, verified at execution. File-perm 600. Injection vector closed.
Ambiguous mode flags default to DRY_RUN. The safe state is the default state.
Per-surface NAV/balance breakers with hysteresis, plus a cross-surface risk governor (halt + throttle).
GitHub Actions → droplet via SSH; pre-restart check aborts if any env file is missing.
Audit record: two external security audits completed. All critical and high findings closed; remaining items are deliberate, scheduled hardening — not structural risk.
Traction
Two surfaces have run autonomously on real capital since May 2026. Every metric below comes from production trade logs, not backtests — at honest, early-stage scale.
Roadmap
The system is built and live. The path to scale is funded data, a proven live return, and multi-subscriber infrastructure — not new code.
Isolated surfaces, auto-deploy, PM2 cluster, Cloudflare security stack, Bitwarden vault, native mobile app, circuit breakers.
Autonomous nightly calibration, cross-surface risk governor, self-maintaining docs, audit ledger, two security audits closed.
Turn the measured edge net-positive across the live book. IBKR equity account funding in progress; weekly audited reviews. The return is the gate.
Scale the funded surfaces, launch the 90-day beta cohort under the 15% HWM model, ship per-subscriber state isolation.
Full subscriber dashboard, compliance wrapper (RIA/custody), and a signal API for third-party consumers.
Founder
"The infrastructure, the AI stack, the security model, the business model — every line is mine. I'm not building a demo. This is live money, live data, and a system that recalibrates itself every night and reports only what it can measure."
The Ask
Capital funds the paper surfaces into live production, premium data, and multi-subscriber infrastructure. The system is already running — we're raising to feed it and prove the live return.
Fund the IBKR equities surface and scale the live FX + ST bankrolls to a return-relevant size.
Premium market-data feeds, lower-latency execution, redundant failover.
Multi-user dashboard, per-subscriber capital isolation, the compliance wrapper, and the HWM commission engine.