Vision Brief — for the dev

1 page. What we are building, how, and where we start. Full technical detail: technical spec. Method: academy-notes.

In one sentence

PANDORA is not a trading robot. It is a decision engine that industrializes an already profitable method, as a layer on top of UW + Skylit + Trading Volatility — 7% human coverage → 100% machine, without degrading quality. Autonomy is the roadmap, not the MVP.

Phasing (do not mix) — detail: V1-BUILD-SCOPE

  • V1 · Scanner (the brain, deliverable in 8-10 weeks): ~750 tickers (final universe TBD with the data API) → features → rules (method, incl. gamma squeeze + rotation by family) → explained score → top 10 → dashboard → journal. 100% deterministic.
  • V2 · Companion (risk manager): portfolio · calendar (OPEX/earnings/macro) · discipline coach · conviction · LLM that explains (never decides).
  • V3 · Agent (autonomy): memory/learning · IBKR · execution.

Golden rules

  1. The LLM never decides (deterministic code decides; the LLM explains/summarizes — V2).
  2. In V1 we log, we do not learn — the moat = the proprietary decision dataset.
  3. No broker in V1. Decision engine, not trading robot.
  4. Every setup comes out explained (contributions per feature).

Key decisions (revised after CTO feedback) — phase tag

Topic Phase Decision
Data V1 UW + Skylit + Trading Volatility — all 3 as soon as possible (degraded start on 2 while waiting for the Skylit API).
TA V1 Deterministic on OHLC (S/R, trend: Price>SMA50>SMA200 / HH+HL / breakout). LLM-vision review = V2.
Score + Explainability V1 Weighted sum + contributions per feature (the "why"). Verdict GO/WATCH/PASS. §5.
Families & rotation V1 Family tag (Setup.family, families.csv done) + strength/rotation by family (daily aggregation, gets richer with the panel history). §6.2.
Gamma squeeze V1 Multi-data detector → gamma_squeeze score, boosts longs. Core of the method. §6.3.
Market read (morning) V1 directional bias + risk-on/off (trinity SPX/SPY/QQQ · GEX/DIX · VIX/term) → global throttle, at the top of the dashboard. §0.2.
Horizon V1 intraday plan → extendable to swing if the level holds.
IV-expansion earnings V1 separate setup category: earnings 7-12d + IV rank↑ + inverted term → capture the IV ramp.
Journal V1 Every decision → proprietary dataset (the moat). We log, we do not learn yet. §3,§7.
Discipline V1 Every score goes through the gauntlet before deployment. §7.
LLM (explains, never decides) V2 Explains / challenges / summarizes — outside the decision path.
Companion V2 Portfolio risk · calendar (OPEX/earnings/macro) · discipline coach · conviction with uncertainty. §6.4.
Autonomy / IBKR execution V3 Hybrid paper→live. Out of MVP.
Learning (recalibration/memory) V3 Walk-forward weight recalibration + LLM memory.
TradingView annotation (Léo) V2+ annotates charts → dataset labels (feeds the TA-vision).
Team — 1 senior full-stack Python dev + founders, multiplied by AI. 2nd profile only if a bottleneck proves it.

Stack (existing PANDORA base, to be confirmed)

  • Python (reuses the PANDORA arsenal: journal, gauntlet, metrics, data adapters).
  • LLM: Claude (Anthropic) — V2, outside decisions: explanation / summary (+ vision on charts later).
  • Broker: Interactive Brokers (API) — V3 (execution), not in the MVP.
  • Data: UW (API), Trading Volatility (API/tv-mcp, free tier available), Skylit (API coming) — behind an abstraction layer (swappable adapter).
  • Annotation: TradingView (export of the levels/zones annotated by Léo).

Reusable existing code (already built)

  • GEX-regime engine + gauntlet, decision journal (meta-labeling), cross-sectional panel collector (built; free TV demo = ~daily collectable snapshot; constraint = reliable collection [cron]; UW later for richer data), re-scoring prototype, live dashboard.

To settle with Léo (before costing)

  • Deal / timeline: amount, runway, deadline.
  • Definition of “done” for the thin slice: how many tickers, which sources plugged in, which success criterion for v1.