Skip to main content

​
The Core Insight

Time isn’t one thing. The framework supports multiple temporal ontologies, each with different causal semantics and validation rules. Key principle: Each temporal mode changes what “consistency” means and how validation works.

​
M17: Modal Temporal Causality

Five temporal modes, each defining different causal regimes.

​
Temporal Modes

class TemporalMode(str, Enum):
    FORWARD = "forward"           # Standard causality—causes precede effects
    DIRECTORIAL = "directorial"  # Narrative time with flashbacks, ellipsis
    BRANCHING = "branching"      # Counterfactual timelines diverge from decision points
    CYCLICAL = "cyclical"        # Prophetic/mythic time—future constrains past
    PORTAL = "portal"            # Backward inference from known endpoints
From schemas.py:20-27

​
Mode-Specific Complexity Multipliers

complexity_map = {
    FORWARD: 1.0,        # Baseline
    DIRECTORIAL: 1.2,    # Narrative structure overhead
    CYCLICAL: 1.3,       # Prophecy validation
    BRANCHING: 1.4,      # Multiple timelines
    PORTAL: 1.5,         # Backward inference (most expensive)
}
Token budgets scale with modal complexity.

​
PORTAL Mode: Backward Temporal Reasoning

Given a known endpoint (“startup reaches $1B valuation in 2030”) and origin (“founders meet in 2024”), PORTAL discovers plausible paths connecting them.

​
Configuration

config = TemporalConfig(
    mode=TemporalMode.PORTAL,
    portal_description="John Doe elected President in 2040",
    portal_year=2040,
    origin_year=2025,
    backward_steps=15,
    path_count=3
)

​
Architecture

Generation → Scoring → Validation → Pivot Detection
  1. Generate candidate antecedent states working backward from endpoint
  2. Score via hybrid evaluation:
    • LLM plausibility
    • Historical precedent
    • Causal necessity
    • Entity capability
  3. Validate forward coherence: Backward-generated paths must make sense when simulated forward
  4. Detect pivot points: Key decision moments where paths diverge

​
Exploration Strategies

# Reverse chronological (simple)
2040 → 2039 → 2038 → ... → 2025

# Oscillating (better for complex scenarios)
2040 → 2025 → 2039 → 2026 → 2038 → 2027 → ...

# Adaptive (system chooses based on complexity)

​
Entity Inference Fix (January 2026)

Original problem: Portal mode blindly copied entities_present from consequent, producing empty lists. Solution: LLM-based entity inference per generated timepoint. 
# In TemporalAgent._infer_entities_for_timepoint()
inferred_entities = self._infer_entities_for_timepoint(
    event_description="Heated debate at city council meeting",
    available_entities=["jane_chen", "campaign_manager", "opponent"],
    context=context
)
# Returns: ["jane_chen", "opponent"] based on event relevance
From workflows/temporal_agent.py implementation The system:
  1. Extracts available entities from context or causal graph
  2. Prompts LLM to identify which entities should be present
  3. Falls back to regex-based name extraction if LLM unavailable

​
Forward Coherence Validation

Backward-generated paths must make sense when simulated forward. Paths below coherence threshold are:
  • Pruned
  • Backtracked
  • Marked with warnings

​
Pivot Point Detection (January 2026 Fix)

Original problem: _detect_pivot_points() checked children_states which was never populated during backward simulation → zero pivot points detected. Solution: Multi-strategy detection:
  1. Divergence-based: Uses key_divergence_points from path analysis (steps where >30% of paths have unique narratives)
  2. Keyword-based: Detects pivot language (“decision”, “pivoted”, “funding”, “launched”, “founded”, “acquired”)
  3. Event-based: Checks key_events and entity_changes for significant shifts
  4. Score-variance: Flags states with unusual plausibility scores relative to neighbors
Before fix: 0 pivot points detected
After fix: 84 pivot points detected (14 per path × 6 paths)

​
Path Divergence Analysis

After path generation, the system analyzes where paths diverge: 
Step 6: Computing path divergence...
  Key divergence at steps: [23, 24]
  Clusters: 1 (high_coherence)

Step 7: Detecting pivot points...
  Path 1: 14 pivot points detected
  Path 2: 14 pivot points detected
  ...

PORTAL SIMULATION COMPLETE
Total pivot points: 84 across 6 paths
Key divergence points: [23, 24]

​
Quick Mode

./run.sh run <template> --portal-quick
Reduces backward_steps from default (often 24) to 5 for fast demos (~15 min vs 1+ hour).

​
Fidelity Template Scaling

Templates now scale proportionally with backward_steps instead of using hardcoded sizes.
TemplateDistribution
minimalist70% TENSOR, 21% SCENE, 7% DIALOG
balanced33% TENSOR, 33% SCENE, 20% GRAPH, 13% DIALOG
portal_pivots2 TRAINED endpoints + scaled middle
max_quality66% DIALOG, 33% TRAINED

​
DIRECTORIAL Mode: Narrative-Driven Reasoning

In DIRECTORIAL mode, causality serves narrative rather than the reverse. Events happen because the story needs them.

​
Configuration

config = TemporalConfig(
    mode=TemporalMode.DIRECTORIAL,
    narrative_arc="rising_action",
    dramatic_tension=0.8,
)

​
Five-Act Structure

ActTension RangeTemporal DensityFidelity Allocation
SETUP0.2-0.4Sparse (establishing beats)SCENE/TENSOR_ONLY
RISING0.4-0.7Increasing (complications)DIALOG
CLIMAX0.8-1.0Dense (every moment matters)TRAINED
FALLING0.5-0.3Decreasing (consequences)DIALOG
RESOLUTION0.1-0.2Sparse (denouement)SCENE
The arc engine shapes generation:
  • Rising action: Conflict keywords in prompts
  • Climax: Higher temperature (0.7 + tension×0.2) for dramatic unpredictability
  • Resolution: Closure and consequence emphasis

​
Camera System

The “invisible director” controls: POV Rotation:
  • Main character for climax scenes
  • Ensemble for setup
  • Antagonist perspective for dramatic irony
Framing Vocabulary:
  • WIDE: Establish scope, show relationships in space
  • CLOSE: Internal conflict, decision moments, emotional beats
  • OVERHEAD: God’s-eye omniscience, fate bearing down
  • SUBJECTIVE: Character’s limited perception, unreliable narration
  • ENSEMBLE: Multiple simultaneous perspectives, group dynamics

​
Dramatic Irony Detection

The system identifies moments where the audience knows something characters don’t. When detected, generation emphasizes the gap—characters make choices that are tragic or comic because of what they don’t know.

​
Fidelity as Dramatic Investment

Resources concentrate where drama concentrates. A 20-timepoint tragedy might allocate 40% of its token budget to 3 climax timepoints. Background setup exists at TENSOR_ONLY because the story hasn’t asked us to look closely yet.

​
Implementation

./run.sh run hound_shadow_directorial  # Full directorial mode
From workflows/directorial_strategy.py

​
CYCLICAL Mode: Prophecy and Time Loops

CYCLICAL mode supports:
  • Time loop narratives
  • Generational sagas
  • Economic/ecological cycles
  • Mythic/religious narratives

​
Configuration

config = TemporalConfig(
    mode=TemporalMode.CYCLICAL,
    cycle_length=4,
    prophecy_accuracy=0.7,
)

​
Cycle Semantics Interpretation

Key innovation: Cycle type is discovered, not prescribed. | Cycle Type | Variation Mode | What Repeats | What Changes | Example | |------------|---------------|--------------|--------------|---------|| | repeating | Mutation | Events, structure | Details, awareness | Groundhog Day—same day, growing knowledge | | spiral | Amplification | Structural beats | Stakes, intensity | Dynasty saga—same conflict, higher stakes | | causal_loop | Retroactive | Causal structure | Nothing (bootstrap) | Predestination—event causes its own preconditions | | oscillating | Inversion | Poles | Magnitude, timing | Boom/bust—expansion then contraction | | composite | Mixed | LLM-directed | LLM-directed | Complex patterns combining multiple types | The system generates a CycleSemantics object:
  • cycle_type
  • variation_mode
  • escalation_rule
  • prophecy_mechanism
  • key_recurring_elements
  • variation_seeds
All subsequent generation respects these semantics.

​
Prophecy System

Prophecies aren’t decorative—they’re structural. A prophecy creates narrative obligation: either fulfill it (destiny) or subvert it (tragedy/comedy). Tracking:
  • prophecy_accuracy (0.0-1.0): How often prophecies come true
  • fulfillment_confidence: LLM-rated confidence that a prophecy was fulfilled
  • prophecy_source_cycle: Which cycle generated the prophecy (enables “ancient prophecies”)
Mechanisms vary by scenario:
  • Witches’ riddles (Macbeth)
  • Deja vu sensations (time loop)
  • Analyst forecasts (economics)
  • Ancestral curses (dynasty)

​
Causal Loop System

For bootstrap paradoxes where events cause themselves:
  1. Detect opportunities for loop closure (“this event could be what caused X in cycle 1”)
  2. Enforce closure by rewriting states to create explicit causal links
  3. Validate that all opened loops eventually close
This enables legitimate bootstrap paradoxes—but only within CYCLICAL mode’s relaxed causality constraints.

​
Fidelity as Cycle-Awareness

PositionFidelityWhy
Cycle boundariesDIALOGThe “seam” where patterns become visible
Prophecy momentsTRAINEDHigh-stakes narrative pivots
Mid-cycle eventsSCENEThe repeating “texture”
Variation pointsDIALOGWhere this cycle diverges from archetype

​
Implementation

./run.sh run agent4_elk_migration  # Verified Feb 2026
From workflows/cyclical_strategy.py Results: 5 paths, 5 cycles, prophecy resolution, coherence 0.727

​
M7: Causal Temporal Chains

Timepoints form explicit causal chains.

​
Structure

class Timepoint(SQLModel, table=True):
    timepoint_id: str = Field(unique=True, index=True)
    timeline_id: str | None = Field(default=None, index=True)
    timestamp: datetime
    event_description: str
    entities_present: list[str] = Field(default_factory=list)
    causal_parent: str | None = Field(default=None, index=True)
    resolution_level: ResolutionLevel = Field(default=ResolutionLevel.SCENE)
    run_id: str | None = Field(default=None, index=True)
From schemas.py:307-316

​
Validation Constraint

Entity at timepoint T can only reference information from timepoints T’ where a causal path exists from T’ to T. This prevents anachronisms structurally, not heuristically.

​
M8: Vertical Timepoint Expansion

Timepoints can be expanded vertically—adding detail within a moment rather than progressing to the next moment. A “board meeting” timepoint might expand into:
  • Arrival
  • Opening remarks
  • Key debate
  • Decision
  • Aftermath
All causally linked but temporally simultaneous.

​
M12: Counterfactual Branching

Create alternate timelines from intervention points.

​
Implementation

def create_counterfactual_branch(parent_timeline, intervention_point, intervention):
    branch = Timeline(parent_id=parent_timeline.id, branch_point=intervention_point)
    
    # Copy timepoints before intervention
    for tp in parent_timeline.get_timepoints_before(intervention_point):
        branch.add_timepoint(tp.deep_copy())
    
    # Apply intervention at branch point
    branch_tp = parent_timeline.get_timepoint(intervention_point).deep_copy()
    apply_intervention(branch_tp, intervention)
    branch.add_timepoint(branch_tp)
    
    # Propagate causal effects forward
    propagate_causal_effects(branch, intervention_point)
    
    return branch
Enables “what if Madison hadn’t shared his notes?” queries with proper causal propagation.

​
Castaway Colony Example

At Day 7, Commander Tanaka’s decision spawns three counterfactual branches: Branch A (Fortify & Wait):
  • Copies all state before Day 7
  • Propagates conservative resource consumption forward
Branch B (Explore & Adapt):
  • Applies exploration interventions
  • Sending teams out
  • Accumulating injuries and discoveries
Branch C (Repair & Signal):
  • Commits all resources to beacon repair
Each branch is internally consistent:
  • Branch B can’t use cave shelter discovered in Branch A’s timeline
  • Branch C can’t benefit from food sources found during Branch B’s exploration

​
Scenario Anchoring

BRANCHING mode forward steps are anchored to the scenario premise via:
  1. Scenario Anchor Block: First 800 chars of scenario_description with explicit “do NOT drift” instruction
  2. Entity Roster: Rich role descriptions from template’s entity_roster
  3. Accumulated World State: Persistent store tracks key_events_history, resource_states, monitoring
This prevents the LLM from abandoning the scenario (e.g., turning a Kepler-442b colony into a VR narrative).

​
M14: Circadian Activity Patterns

Entities have activity probabilities that vary with time of day.

​
Implementation

def get_activity_probability(hour: int, activity: str) -> float:
    probability_map = {
        "sleep": lambda h: 0.95 if 0 <= h < 6 else 0.05,
        "work": lambda h: 0.7 if 9 <= h < 17 else 0.1,
        "social": lambda h: 0.6 if 18 <= h < 23 else 0.2,
    }
    return probability_map.get(activity, lambda h: 0.0)(hour)
This constrains entity behavior plausibly without requiring explicit scheduling logic.

​
Energy Budget Integration

From validation.py:98-137: 
@Validator.register("energy_budget", "WARNING")
def validate_energy_budget(entity: Entity, context: dict) -> dict:
    # Base cost per knowledge item
    base_expenditure = new_knowledge_count * 5
    
    # Apply circadian adjustments if timepoint available
    if timepoint and circadian_config:
        activity_type = context.get("activity_type", "work")
        expenditure = compute_energy_cost_with_circadian(
            activity_type, 
            timepoint.timestamp.hour, 
            base_expenditure, 
            circadian_config
        )
    
    if expenditure > budget * 1.2:  # Allow 20% temporary excess
        return {"valid": False, "message": f"Energy expenditure exceeds budget"}
    return {"valid": True, "message": "Energy budget satisfied"}

​
Next Steps

Knowledge Provenance

Who knows what, from whom, when

Entity Simulation

Dialog synthesis and entity behavior

Build docs developers (and LLMs) love

Get started for freeTalk to us