A Atlantis Core

Core Architecture

Scoring To Permission

Pattern evidence becomes trade hypotheses, signal intents, permission decisions, and reason-coded admission or rejection.

# AtlantisV2 core pipeline
ReferenceTime
  -> MarketSnapshot
  -> FeatureVector
  -> PatternHit[]
  -> TradeHypothesis
  -> SignalIntent
  -> PermissionDecision
  -> PositionIntent
  -> OutcomeRecord

The middle of Atlantis Core separates three decisions that legacy systems often merge:

  1. Is there evidence for a thesis?
  2. What instrument-neutral intent should be proposed?
  3. Is that intent allowed right now?

Keeping these separate is what lets the same thesis be expressed as spot, perp, future, option, paper, shadow, or no trade.

TradeHypothesis

ScoringEngine consumes PatternHit[] and emits TradeHypothesis.

The current scoring slice includes:

  • evidence stack
  • directional bias
  • conviction score
  • confluence across same-direction hits
  • conflict handling
  • quality penalties
  • feature schema version surfacing
  • mixed-underlying rejection
  • mixed feature schema rejection

EvidenceStack

Evidence stack retains the pattern hits that contributed to the hypothesis. This prevents scoring from becoming an opaque number.

The stack carries:

  • pattern ids
  • hit ids
  • direction
  • confidence
  • feature schema versions

ConvictionBreakdown

The framework has a breakdown object so future scoring dimensions can be added without changing the top-level hypothesis contract.

Current implemented dimensions include pattern strength and confluence. Future dimensions may include:

  • regime alignment
  • liquidity quality
  • execution cost
  • historical edge
  • freshness confidence
  • portfolio concentration
  • operator overrides

SignalIntent

SignalIntent is the most important interface in Core. It is the instrument-neutral thesis handoff.

It includes:

  • signal intent id
  • thesis id
  • source hypothesis id
  • underlying id
  • direction
  • preferred exposure
  • horizon
  • invalidation context
  • target context
  • constraints
  • evidence summary
  • feature schema version
  • intent version
  • decision state
  • reason codes

Preferred Exposure

ExposurePreference prevents hidden instrument assumptions:

  • spot
  • perp
  • future
  • option single leg
  • option structure
  • any delta one
  • no preference

A BTC long thesis can prefer ANY_DELTA_ONE without saying whether it must be spot or perp.

Invalidation And Target Context

Current typed shells include:

  • invalidation price level
  • time limit
  • feature condition
  • pattern failure condition
  • regime flip condition
  • max adverse excursion
  • price target
  • move target percent
  • time target
  • expected move
  • minimum viable reward
  • scaleout shape

These contexts are intentionally portable. They should not mention one broker order type.

PermissionDecision

PermissionEngine consumes SignalIntent plus RiskContext and emits PermissionDecision.

Current permission logic includes:

  • required capability defer
  • quality flag rejection
  • paper-only and shadow-only modes
  • kill switch
  • duplicate exposure gate
  • status precedence
  • constraints applied
  • portfolio and operator context snapshots
  • cycle-scoped decision ids
  • enum-only reason codes

Permission Statuses

Implemented statuses include:

  • allow
  • allow with limits
  • allow paper only
  • allow shadow only
  • reject
  • defer

Permission is where risk and operator control belongs. Signal generation remains gate-free so research can see what would have fired even when permission rejects it.