Monkey Science · UltraScalp

The core is DATA.
Theories. Execute. Adapt.

Dangerous for data
Guarded for live
Intelligence stays local
Demo only · live capital blocked
Engineering spine → data→ RunPack→ decision trace→ MFE / MAE→ construction audit→ stress gate→ promotion
THE CORE DATA protect integrity THEORIES from data only EXECUTE guarded · paper first ADAPT writes new data INGEST normalise · isolate RunPack · traces · KG cBot · paper MFE/MAE · gate · promote DOM · news

UltraScalp is an explainable, evidence-driven system for short-lived opportunities. Data is not an input. Data is the product. Theories are assembled from it. Execution is allowed only after cost, sizing, and invalidation survive a gate. Adaptation does not mutate live logic — it writes new data.

Do not ask: was the signal right?
Ask: was the trade constructible after cost, sizing, timing, and invalidation?
01   AI is never an untracked oracle. Analyst proposes. Gates promote.
02   Deterministic code owns risk and order permission. LLMs cannot override a hard gate.
03   No strategy passes on gross P&L. After-cost or it does not exist.
04   Fail closed. Missing data reduces authority. It never manufactures confidence.
05   Learn offline. Reviewed proposal — never an automatic live mutation.
01 · The core

DATA

Ingest, normalise, isolate. Two lanes stay labelled. Dirty data never enters a primary verdict.

  • Databento DBN/CSV — DOM, MBP-10, orderflow, replay
  • cTrader journals, trade logs, Open API context
  • News / calendar feeds v3
  • Candles, indicators, symbol relationships
  • Capture bus → incoming → RunPack archive
  • Metals / non-normalised trades isolated
02 · From data only

THEORIES

A theory is a RunPack plus an experiment id. Never inferred after the fact. Never promoted by the model that proposed it.

  • RunPack: features · decisions · ai_traces · trades · verdict
  • Deterministic spine + AI selector / veto / chooser
  • Sequence AI Labeler V3.1 · Orderflow V3
  • Edge Probe · Signal Diagnostic · Timebox Scalp
  • Research Workbench · Evidence Board · Knowledge Graph
  • Clear picks ranked by evidence, not vibes
03 · Guarded

EXECUTE

Paper first. Broker adapters only. Intelligence stays local. Live capital is blocked.

  • Cost preflight before any size
  • Risk-normalised sizing · symbol risk caps
  • Armed entry: Market / Pullback / Confirm — or expire
  • Fail-fast: adverse-first · no-follow · giveback
  • cBot · paper executor · broker router
  • AI veto / exec shadow-first · live adjust = false
04 · Writes new data

ADAPT

Measure the path, not the story. Survivors promote. Failures become labels. That is why data stays the core.

  • MFE / MAE · endpoint · SL/TP sim · cost model
  • Construction audit — early? cost-positive? sized? MFE first?
  • Stress gate: OOS · placebo · after-cost · stability
  • Promotion ladder → next logic profile
  • Evidence Board: dominant failure + next experiment
  • Promote / reject overnight into the graph

Lane A · Futures / Databento

DBN / MBP-10 → ingest → normalised features → RunPack → order-flow research → verdict (hit-rate, OOS, placebo, after-cost)

Lane B · CFD / cTrader

cBot logs → trade_log → construction / Achilles audit → verdict (sizing, cost, late entry, no-MFE, exits)