Skylit — Agentic trading — Strategy & moat (working doc, 2026-06-16)
Working doc to frame the Skylit project and the investor discussion. Not a finished pitch — the raw material (wedge, moat, architecture, roadmap, ask, risks). Canonical today: the deck (
/skylit), the Vision Brief and the technical spec. If they differ, these 3 prevail.Method (the edge layer): the dealer/Heatseeker read (GEX/VEX) is documented in The Greeks Bible — Skylit Edition. It is the theoretical base behind the re-scoring: the scanner's "maps support / flow / dealer" checks = GEX/VEX nodes, dealer hedge flow, and the vol regime filter (golden rule: trade where Greeks + vol agree).
IV layer (2nd lane): IV Expansion into Earnings — a detector of IV-expansion plays (earnings in 7-12d, IV rank ↑, inverted term structure) next to the long directional checklist. We capture the IV ramp, not the crush.
0. The trigger
TradingVolatility (Jay Wolberg) has just released tv-mcp (PyPI) + a v2 API
"designed for Claude & LLMs": an MCP server that gives, per ticker and cross-ticker,
the options structure (GEX, gamma flip, skew, flow, max pain, expected move) AND a
setup ranking (/top-setups) with direction, entry, stop, sizing, options
structure. Tested in demo mode: excellent quality, LLM-native.
Strategic consequence: the data + scanning + trade-setup layer — which we thought we had to build — already exists and plugs in. So we no longer fight on data. We fight on edge, validation and execution.
1. The wedge
A trader cannot watch 500 tickers and pull out the best trades of a session. The data exists (TV, Unusual Whales…). What is missing: sorting with a method that works, proving it works, and executing — without drowning the human. That is our place.
2. What we INTEGRATE vs what we BUILD (the key reframe)
| Layer | Decision |
|---|---|
| Multi-ticker options data (GEX/flip/skew/flow), raw ranking, raw trade-setup | INTEGRATE — tv-mcp / TradingVolatility v2 API (+ abstraction to swap UW/others) |
| Re-scoring with the methodology (the alpha) | BUILD |
| Anti-overfit validation engine (the gauntlet) | BUILD (we already have it, proven on PANDORA) |
| Market regime gate (DIX/GEX dark-pool, v_joint) | BUILD (PANDORA) |
| Human-in-the-loop + learning loop | BUILD |
| Execution + risk (paper→live) | BUILD (Alpaca/IBKR already wired) |
| Cockpit / UI | BUILD |
Anyone can
pip install tv-mcpand wire Claude (the tweet that put us on the trail did it in one evening). So data is NOT the moat. Our moat is elsewhere.
3. The moat (4 layers, in order of defensibility)
- The methodology edge, validated. TV's
opportunity_scoreis generic. We re-rank it with Léo's discretionary method (zones/maps, Fib retracements in log, flow/volume confluence, multi-TF) and we gate it by the market regime (PANDORA: v_joint, DIX/GEX). TV setups → filtered by a brain with a proven edge. - The validation discipline (the trust moat). We only show/trade what survives out of sample (Deflated Sharpe, walk-forward, costs — the gauntlet). TV gives signals; we prove which ones win. PANDORA is the proof that we know how to kill overfitting (75 strategies → almost all noise, 1 validated champion).
- The decision dataset (the compounding moat). Every human accept/reject becomes labeled data (meta-labeling). The system learns which setups, filtered by our method, are actually taken and work. TV does not have it; a competitor wiring tv-mcp does not have it. The longer we run, the more defensible it is.
- Execution + risk (the product). TV stops at "here is a setup". We close the loop: validated sizing (fractional Kelly), portfolio risk, paper→live. A trade executed and managed, not a suggestion.
Defensibility pitch in one sentence: "We are not a data wrapper. We are the brain (methodology edge), the discipline (anti-overfit validation) and the execution that turn commoditized options data into validated, executed trades — with a decision dataset that gets stronger with every trade."
4. Architecture
TradingVolatility tv-mcp ─┐
PANDORA (DIX/GEX, v_joint)─┤→ FUSION → Léo methodology re-scoring → regime GATE →
Macro / flow (extensible) ─┘ │
▼
Validation gauntlet (shows only real OOS edge)
▼
Cockpit: ranked scanner + card per ticker (readable read)
▼
Human-in-the-loop (accept/reject → learning dataset)
▼
Execution paper→live (Alpaca/IBKR) + risk/sizing
Abstract data layer (interface) → we can swap/add UW, etc. without rewriting.
5. De-risked roadmap (data being solved, it goes fast)
- P0 — PoC (days, not months). tv-mcp → Claude →
/top-setupsre-scored by a v1 of the method + PANDORA regime gate → cockpit where Léo reviews. Goal: prove that our re-scoring beats TV's raw ranking. - P1. Full v1 methodology overlay + gauntlet on the re-scored setups + cockpit + paper logging. Metric: do our top picks beat TV's / Léo's manual selection?
- P2. Learning loop (meta-label of Léo's decisions) + semi-auto execution (paper, 1-click).
- P3. Live + multi-user / scale.
6. The ask (partnership)
Capital to fund 1 senior full-stack dev + the infra (Léo + Gab = product/method/validation, AI multiplies the dev — no team of 3). Order of magnitude: ~55 k€ over 6 months (P1→P4), released in 2 tranches with a validation gate. The PoC is doable in days (data is solved) — so we can show something running before committing to the big build.
7. Risks & mitigations (honest)
- Dependency on TV (third-party data, not our IP, demo = 8 tickers, full universe = paid key). → Abstract the data layer (swap UW/others); eventually, proprietary signals.
- Commoditization (anyone can wire tv-mcp). → The moat = validated edge + decision dataset + execution, not data.
- Regulatory → semi-auto, human-in-the-loop, never personalized advice.
- ML over-promise → we sell a proven method + validation, not "magic ML" (see PANDORA: anti-overfit rigor IS the argument).
Appendix — proof that we can execute (to show)
PANDORA: validated SPX system (v_joint, DSR sweep, walk-forward), live dashboard, automated daily brief, Alpaca paper executor, anti-overfit gauntlet. The scanner = this know-how, extended to multi-ticker with TV's data.