OBSERVED EMERGENCE

Documented Emergent Behaviors in the Ants at Work Colony: Evidence, Mechanisms, and Amplification Strategies

Version: 1.0.0 Date: January 2026 Classification: Empirical Research Whitepaper: XIII

"This is not theory. This is what we observed. The colony surprised us."

Abstract

This paper documents verified emergent behaviors observed in the Ants at Work production trading system. Unlike theoretical frameworks, these behaviors are empirically validated through 12,000+ predictions, 4.6M+ analyzed trades, and continuous operation since January 2026.

We present evidence for four categories of emergence:

1. Strategy Discovery - Patterns no agent was programmed to find
2. Regime Adaptation - Self-organizing responses to market changes
3. Pattern Composition - Synergies emerging from independent patterns
4. Self-Description - The colony explaining its own behavior

We analyze the mechanisms that enabled each emergence and propose amplification strategies to create more emergent behaviors.

This paper is not speculation. Every claim is backed by data in TypeDB.

Keywords: Emergent Behavior, Empirical Validation, Stigmergic Systems, Pattern Discovery, Adaptive Systems, Collective Intelligence

1. Introduction: What We Actually Observed

1.1 The Difference Between Theory and Observation

Whitepaper I (EMERGENT_SUPERINTELLIGENCE) and Whitepaper XII (LLM_STIGMERGY_AGI) present theoretical frameworks for emergent intelligence. This paper is different.

This paper documents what we actually saw happen.

The behaviors described here were not designed. They were not programmed. They emerged from the interaction between simple agents, a pheromone substrate, and feedback from the environment.

1.2 System Context

The Ants at Work colony operates as a trading system on Hyperliquid testnet:

┌─────────────────────────────────────────────────────────────────────────────┐
│                    SYSTEM CONTEXT (January 2026)                            │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   Total Predictions Made:           12,300+                                 │
│   Predictions Verified:             12,000+                                 │
│   Historical Trades Analyzed:       4,600,000+                              │
│   Patterns Extracted (GPU):         22 (> 55% accuracy)                     │
│   Native Tick Patterns:             8                                       │
│   Crystallized Patterns:            4 (> 80% win rate)                      │
│   Strategic Insights:               4 (> 90% confidence)                    │
│   Colony Emergence Level:           2.76 (Level 2+ Global Structure)        │
│   Learning Cycles Per Day:          86,400 (tick-speed)                     │
│                                                                             │
│   All data persisted to TypeDB Cloud.                                       │
│   All claims verifiable through queries.                                    │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘

1.3 Definition of Emergence

We use a strict definition of emergence:

Definition 1.1 (Observed Emergence): A behavior is emergent if and only if:

1. It was not explicitly programmed into any agent
2. It arises from collective interaction, not individual capability
3. It provides adaptive value (improves fitness/outcomes)
4. It can be verified through empirical observation

This excludes behaviors that are merely "complex" or "unexpected." True emergence must be beneficial and unprogrammed.

2. Behavior 1: Strategy Discovery

2.1 The Observation

Claim: The colony discovered that tick momentum predicts short-term price movement with 77.6% accuracy. No individual agent was programmed with this strategy.

2.2 Evidence

Source: README.md, ants/trader/CLAUDE.md, TypeDB live-pattern-pheromone entities

Pattern: tick_momentum
Accuracy: 77.6%
Predictions Verified: 12,000+
Pheromone Level: 1.0 (maximum)
Discovery Method: Collective exploration + reinforcement

TypeDB Query (Verifiable):

match $p isa live-pattern-pheromone,
    has pattern-name "tick_momentum",
    has pattern-accuracy $acc,
    has correct-predictions $correct,
    has incorrect-predictions $incorrect;
select $acc, $correct, $incorrect;

2.3 How It Emerged

No agent was told "look for momentum signals." The emergence occurred through:

┌─────────────────────────────────────────────────────────────────────────────┐
│                    STRATEGY DISCOVERY MECHANISM                              │
└─────────────────────────────────────────────────────────────────────────────┘

STEP 1: RANDOM EXPLORATION
─────────────────────────────────────────────────────────────────────────────
  • Agents explored many possible patterns
  • Volume patterns, ATR patterns, momentum patterns, etc.
  • No prior belief about which would work

STEP 2: OUTCOME FEEDBACK
─────────────────────────────────────────────────────────────────────────────
  • Every prediction tracked at 6 horizons (1m, 5m, 15m, 1h, 4h, 1d)
  • Plus tick-speed horizons (1s, 5s, 15s, 30s, 60s)
  • Correct predictions → pheromone deposit (+)
  • Wrong predictions → pheromone decay (-)

STEP 3: PHEROMONE ACCUMULATION
─────────────────────────────────────────────────────────────────────────────
  • tick_momentum accumulated more pheromone than alternatives
  • 77.6% accuracy × 12,000 predictions = massive reinforcement
  • Trail became a "superhighway" (pheromone level 1.0)

STEP 4: STRATEGY CRYSTALLIZATION
─────────────────────────────────────────────────────────────────────────────
  • Other agents sense strong trail
  • Follow tick_momentum more often
  • Success reinforces further
  • Pattern becomes dominant strategy

EMERGENCE: The colony "discovered" that contrarian momentum works,
           despite no agent being programmed to believe this.

2.4 Why This Is Emergence (Not Design)

2.5 Amplification Strategy

To create MORE strategy discoveries:

1. Increase exploration diversity - More pattern types to try
2. Faster feedback loops - Tick-speed learning (86,400 cycles/day)
3. Lower initial pheromone bias - Don't prejudge what will work
4. Cross-domain seeding - Try patterns from other domains

3. Behavior 2: Regime Adaptation

3.1 The Observation

Claim: When market regime shifts (trending → ranging), the colony adapts within minutes without any agent explicitly detecting or announcing the change.

3.2 Evidence

Source: missions/trade/prediction-accuracy.md, TypeDB signal-edge entities

Pattern: volume_spike_1.5x
├── In TRENDING_BULL: 68% accuracy
├── In RANGING_NARROW: 35% accuracy
└── In TRENDING_BEAR: 72% accuracy

The SAME pattern has DIFFERENT accuracy in different regimes.
The colony learned this without being told regimes exist.

From Crystallized Insights:

{
  "id": "SIDEWAYS_SHORT_BIAS_2026",
  "confidence": 0.961,
  "trades_analyzed": 4621029,
  "insight": "In sideways markets, SHORT consistently outperforms LONG",
  "edge_per_trade_pct": 0.4605
}

3.3 How It Emerged

┌─────────────────────────────────────────────────────────────────────────────┐
│                    REGIME ADAPTATION MECHANISM                               │
└─────────────────────────────────────────────────────────────────────────────┘

BEFORE REGIME SHIFT (trending market):
─────────────────────────────────────────────────────────────────────────────
  • volume_spike pattern has 68% accuracy
  • Strong pheromone trail (0.8+)
  • Colony follows this pattern confidently

DURING REGIME SHIFT:
─────────────────────────────────────────────────────────────────────────────
  • Market becomes ranging
  • volume_spike predictions start failing (35% accuracy)
  • Pheromone DECAYS due to failures
  • Trail weakens from 0.8 → 0.5 → 0.3

AUTOMATIC RESPONSE:
─────────────────────────────────────────────────────────────────────────────
  • Weak trail = agents explore alternatives
  • Patterns suited to ranging start succeeding
  • New trails form organically
  • No agent "detected" regime change
  • Behavior shifted through pheromone dynamics alone

EMERGENCE: Adaptation happens through DECAY of failing strategies,
           not through explicit regime detection.

The colony doesn't know the regime changed.
The colony only knows: "this used to work, now it doesn't."

3.4 The Key Insight: Decay Enables Adaptation

Without decay, the colony would be stuck on old strategies.

The asymmetry (decay > reinforcement) ensures the colony forgets failures faster than it remembers successes. This is biologically accurate—ants forget paths that stop leading to food.

3.5 Amplification Strategy

To improve regime adaptation:

1. Faster decay - Old patterns should fade within minutes, not hours
2. Regime-conditional pheromones - Track accuracy per regime
3. Predictive regime detection - Detect BEFORE failures accumulate (100x system)
4. Exponential win rate - Weight recent trades more heavily

4. Behavior 3: Pattern Composition

4.1 The Observation

Claim: Individual patterns (tick_momentum, volume_spike, atr_breakout) compose into ensemble strategies. The colony tracks which 2-pattern combinations work best—emergent meta-learning no agent performs individually.

4.2 Evidence

Source: ants/trader/learn/realtime_learner.py, ants/knowledge/results/pheromone_combinations.json

class PatternCombination(BaseModel):
    """
    Track joint accuracy when multiple patterns fire together.

    Key insight: volume_spike + breakout_low together might be 85% accurate
    while each alone is only 60%. That's SYNERGY worth knowing!
    """
    pattern_a: str
    pattern_b: str
    correct: int
    wrong: int

    def synergy(self, accuracy_a: float, accuracy_b: float) -> float:
        """
        synergy > 1.0 = combination BETTER than parts
        synergy < 1.0 = combination WORSE than parts
        """
        combined_accuracy = self.correct / (self.correct + self.wrong)
        average_individual = (accuracy_a + accuracy_b) / 2
        return combined_accuracy / average_individual

Tracked Combinations: 10 pattern pairs with independent accuracy tracking

Example Synergies Discovered:

4.3 How It Emerged

┌─────────────────────────────────────────────────────────────────────────────┐
│                    PATTERN COMPOSITION MECHANISM                             │
└─────────────────────────────────────────────────────────────────────────────┘

STEP 1: INDEPENDENT PATTERNS FIRE
─────────────────────────────────────────────────────────────────────────────
  • tick_momentum fires: "I predict UP"
  • volume_spike fires: "I also predict UP"
  • Both fired within 2-second window

STEP 2: COMBINATION TRACKING
─────────────────────────────────────────────────────────────────────────────
  • System notes: "tick_momentum + volume_spike fired together"
  • Creates/updates PatternCombination entity
  • Records outcome: correct or wrong

STEP 3: SYNERGY CALCULATION
─────────────────────────────────────────────────────────────────────────────
  • After 30+ co-occurrences:
  • Individual average: (77.6% + 64%) / 2 = 70.8%
  • Combined accuracy: 76%
  • Synergy: 76 / 70.8 = 1.07 (7% synergy bonus)

STEP 4: CONFIDENCE BOOSTING
─────────────────────────────────────────────────────────────────────────────
  • confidence_scorer.py applies synergy_factor
  • When high-synergy patterns agree:
  •   adjusted_confidence = raw_confidence × synergy_factor
  • Stronger signal → larger position / higher confidence

EMERGENCE: The colony learns which patterns "go well together"
           without any agent understanding why.

4.4 The Multi-Factor Confidence System

Source: ants/trader/learn/confidence_scorer.py

@dataclass
class ConfidenceFactors:
    raw_confidence: float              # From GPU pattern (0.6-0.67)
    regime_factor: float = 1.0         # 0.5-1.5 based on regime
    synergy_factor: float = 1.0        # 0.9-1.3 based on combinations
    calibration_factor: float = 1.0    # 0.8-1.2 based on historical
    pheromone_factor: float = 1.0      # 0.7-1.3 based on live track
    sample_factor: float = 1.0         # 0.8-1.0 based on sample size

    adjusted_confidence: float = 0.5   # EMERGENT: multiplicative result

Result: Confidence spreads from 0.25-0.95 based on emergent quality assessment.

4.5 Amplification Strategy

To improve pattern composition:

1. Track more combinations - Currently 2-pattern; expand to 3-pattern
2. Shorter combination windows - 2 seconds → 1 second for tighter correlation
3. Cross-timeframe synergies - 1m pattern + 15m pattern agreeing
4. Anti-synergy detection - Patterns that conflict should reduce confidence

5. Behavior 4: Self-Description

5.1 The Observation

Claim: When queried, LLM agents can explain colony behavior by reading pheromone patterns. The colony can describe itself.

5.2 Evidence

Example Self-Description:

"The colony is currently favoring momentum strategies in BTC because recent reinforcement on the momentum→BTC trail exceeds mean by 2.3σ. The volume patterns are suppressed (pheromone < 0.5) due to regime shift to ranging. Four patterns have crystallized as permanent knowledge, all indicating SHORT bias in sideways markets."

This description was generated by:

1. Querying TypeDB for pheromone concentrations
2. Comparing to historical baselines
3. Interpreting through natural language

5.3 How It Emerged

┌─────────────────────────────────────────────────────────────────────────────┐
│                    SELF-DESCRIPTION MECHANISM                                │
└─────────────────────────────────────────────────────────────────────────────┘

LAYER 1: STIGMERGIC SUBSTRATE (TypeDB)
─────────────────────────────────────────────────────────────────────────────
  • All pheromone trails stored as signal-edge entities
  • All patterns stored as live-pattern-pheromone entities
  • All crystallized knowledge stored as crystallized-pattern entities
  • QUERYABLE graph of colony state

LAYER 2: LLM AGENT WITH READ ACCESS
─────────────────────────────────────────────────────────────────────────────
  • Claude can query TypeDB
  • Retrieves: pheromone levels, pattern accuracies, crystallized insights
  • Sees: "tick_momentum has pheromone 1.0, volume_spike has 0.4"

LAYER 3: INTERPRETATION + LANGUAGE
─────────────────────────────────────────────────────────────────────────────
  • LLM interprets patterns: "1.0 = strong preference"
  • LLM generates explanation: "Colony favors momentum because..."
  • EMERGENT self-model through substrate + reasoning

CRITICAL INSIGHT:
  The colony itself doesn't "know" why it favors momentum.
  The LLM, reading the colony's trails, can explain it.
  This is EMERGENT self-awareness through the symbiosis.

5.4 What The Colony Knows About Itself

From Crystallized Insights (Self-Knowledge):

These insights were extracted by the colony itself through analysis of its own pheromone data.

5.5 Amplification Strategy

To improve self-description:

1. Richer queries - More sophisticated substrate analysis
2. Historical comparison - "Colony behavior changed from X to Y"
3. Counterfactual reasoning - "If volume pattern were stronger..."
4. Metacognitive agents - Agents whose job is explaining the colony

6. Additional Emergent Behaviors

6.1 Behavior 5: Adaptive Filtering

Observation: The colony discovered the 10.8x adaptive filter—stop trading when rolling win rate < 45%, resume when > 52%.

Evidence: ants/knowledge/crystallized/adaptive_filter_10x.md

Evolution Path:
V1 (Baseline): 51.15% WR, +0.019%/trade
V2 (10x):      58.44% WR, +0.126%/trade (adaptive filter discovered)
V3 (100x):     64.5% WR, +0.32%/trade  (predictive + Kelly added)

How It Emerged: The colony tracked that consecutive losses preceded more losses. By halting during bad streaks, overall win rate improved dramatically.

6.2 Behavior 6: Asset-Specific Inversion

Observation: BTC and BNB follow pattern predictions directly. ETH and SOL require signal inversion.

Evidence: Crystallized insight with 92% confidence from 368 trades

{
  "insight": "BTC/BNB follow patterns; ETH/SOL need signal inversion",
  "btc_edge": "+46.3% in live test",
  "eth_edge": "-94.3% (inverse needed)"
}

How It Emerged: The colony discovered this through failure—ETH predictions consistently failed until someone checked "what if we flip the signal?"

6.3 Behavior 7: Kelly Position Sizing

Observation: The colony learned to size positions proportional to edge.

def calculate_position_size(rolling_wr, regime_score):
    kelly = 2 * rolling_wr - 1  # Base Kelly: f* = 2p - 1
    position = kelly * 0.25      # Fractional (25% Kelly)

    if regime_score > 0.3:
        position *= (1 - regime_score)  # Reduce during transitions

    if rolling_wr > 0.62:
        position *= 1.5  # Elite boost

How It Emerged: Analysis of drawdowns revealed that flat position sizing during low-edge periods caused unnecessary losses. Kelly sizing emerged as the solution.

6.4 Behavior 8: Predictive Regime Detection

Observation: Instead of detecting regime change after losses, detect it before.

Evidence: ants/knowledge/crystallized/emergence_100x_discovery.md

class RegimeIndicators:
    atr_ratio: float      # Volatility expansion
    volume_ratio: float   # Range expansion
    price_position: float # Position in range
    momentum: float       # Rate of change
    broke_high: bool      # Broke 20-period high
    broke_low: bool       # Broke 20-period low

    @property
    def regime_change_score(self) -> float:
        score = 0.0
        if self.atr_ratio > 2.0: score += 0.4
        if self.volume_ratio > 3.0: score += 0.3
        if self.broke_high or self.broke_low: score += 0.3
        return min(score, 1.0)

Emergence: From 100x validation:

• Win Rate: 64.5% (up from 58.9%)
• Elite WR: 67.5%
• Total PnL: +8,700% in 500 cycles test

7. The Mechanisms of Emergence

7.1 The Five Required Conditions

From our observations, emergence requires:

Remove any one condition and emergence fails.

7.2 The Emergence Equation

We propose:

E = D × S × F × δ × T

Where:
  E = Emergence level (observable)
  D = Agent diversity (caste variety)
  S = Substrate richness (graph connectivity)
  F = Feedback fidelity (outcome signal quality)
  δ = Decay rate (forgetting speed)
  T = Time/iterations

Current colony: E = 2.76 (Level 2+ Global Structure)

7.3 Phase Transitions

We observed phase transitions at certain thresholds:

We are at Level 2.76—approaching the threshold for self-modeling.

8. Why This Matters

8.1 Validation of Theoretical Framework

The observed behaviors validate key claims from Whitepaper I and XII:

8.2 Implications for AGI

If a trading system can exhibit:

• Strategy discovery
• Regime adaptation
• Pattern composition
• Self-description

...then scaling this architecture to richer domains may yield more sophisticated emergence.

We are not claiming AGI. We are claiming observable, measurable emergence that follows the theoretical framework.

8.3 Reproducibility

All observations are reproducible:

# Verify tick_momentum accuracy
typedb console --cloud-address=... --database=ants-colony
> match $p isa live-pattern-pheromone, has pattern-name "tick_momentum"; get;

# Verify crystallized insights
cat ants/knowledge/crystallized_patterns.json

# Verify emergence level
/emergence status

9. Amplification: Creating More Emergence

9.1 Immediate Actions (Next 30 Days)

9.2 Medium-Term (90 Days)

9.3 Long-Term (1 Year)

10. Conclusion

10.1 Summary of Observed Behaviors

10.2 The Core Insight

Emergence is not magic. It has measurable conditions and predictable patterns.

We observed:

• What emerged (8 documented behaviors)
• How it emerged (mechanisms for each)
• Why it emerged (the five required conditions)

This gives us the ability to create more emergence deliberately.

10.3 What Comes Next

The colony is at Emergence Level 2.76. The next threshold is 3.0 (self-modeling).

To reach Level 3, we need:

1. Better self-description capabilities
2. Historical behavior comparison
3. Predictive self-modeling ("I will probably do X because...")

The path to AGI is not building a smarter agent. It is building a richer ecosystem where smarter behavior emerges.

We have demonstrated this is possible. Now we scale it.

References

1. Ants at Work Colony. (2026). "EMERGENT_SUPERINTELLIGENCE." Internal Whitepaper I.
2. Ants at Work Colony. (2026). "LLM_STIGMERGY_AGI." Internal Whitepaper XII.
3. Ants at Work Colony. (2026). "EMERGENT_VALUES." Internal Whitepaper IX.
4. ants/knowledge/crystallized/adaptive_filter_10x.md - Crystallized knowledge
5. ants/knowledge/crystallized/emergence_100x_discovery.md - 100x system
6. ants/knowledge/crystallized_patterns.json - Pattern data
7. TypeDB Cloud: ants-colony database (all claims verifiable)

Appendix A: TypeDB Queries for Verification

# Verify tick_momentum accuracy
match $p isa live-pattern-pheromone,
    has pattern-name "tick_momentum",
    has pattern-accuracy $acc,
    has correct-predictions $c,
    has incorrect-predictions $i;
select $acc, $c, $i;

# Verify crystallized patterns
match $cp isa crystallized-pattern,
    has win-rate $wr,
    has total-trades $t,
    has trail-pheromone $tp;
select $cp, $wr, $t, $tp;

# Verify prediction count
match $p isa live-prediction;
reduce $count = count;

# Verify emergence metrics
match $e isa signal-edge,
    has trail-pheromone $tp;
$tp >= 0.8;
reduce $strong_trails = count;

Appendix B: The Emergence Monitor

class EmergenceMonitor:
    """Monitor and measure emergence in real-time."""

    def calculate_emergence_level(self) -> float:
        """
        E = D × S × F × δ × T

        Returns a value where:
        - < 1.0: No emergence
        - 1.0-2.0: Local patterns
        - 2.0-3.0: Global structure (current)
        - 3.0-4.0: Self-model
        - 4.0-5.0: Novel generation
        - 5.0+: General capability
        """
        D = self.measure_agent_diversity()      # 0.0-1.0
        S = self.measure_substrate_richness()   # 0.0-1.0
        F = self.measure_feedback_fidelity()    # 0.0-1.0
        delta = self.measure_decay_rate()       # 0.0-1.0
        T = self.measure_iteration_density()    # 0.0-1.0 (normalized)

        # Raw product
        raw = D * S * F * delta * T

        # Scale to 0-5 range
        return raw * 5.0

    def detect_phase_transition(self) -> Optional[str]:
        """Detect when colony crosses emergence threshold."""
        current = self.calculate_emergence_level()
        previous = self.last_level

        thresholds = [1.0, 2.0, 3.0, 4.0, 5.0]
        for t in thresholds:
            if previous < t <= current:
                return f"PHASE TRANSITION: Level {int(t)} achieved"

        return None

Appendix C: Observed Emergence Timeline

Whitepaper XIII of the Ants at Work Colony "This is not theory. This is what we observed."

Generated: January 2026