Mars Mission Portal: Backward Reasoning from Disaster

The Ares III Mars Mission Portal template demonstrates PORTAL mode temporal reasoning—working backward from a catastrophic outcome to discover its root causes. This flagship example shows how schedule pressure, institutional culture, and cascading technical debt can create an inevitable failure trajectory.

Template: mars_mission_portal | Mode: PORTAL | Cost: ~

0.18 (quick) / ~

0.40 (full)

The Scenario

The Ares III crewed Mars mission loses contact during orbital insertion in March 2031. Last telemetry shows cascading systems failures in life support and communications. The mission was celebrated as humanity’s greatest achievement until silence fell. PORTAL mode works backward from this disaster through 5 years (or 10 in full mode) to January 2026, exploring how present-day decisions create future outcomes.

Cast of Characters

Sarah Okafor

Mission CommanderExperienced astronaut, politically pressured by NASA leadership to maintain schedule. Mediates between engineering rigor and institutional pressure.Final state: Valence +0.47, Arousal 0.57, Energy 124.4

Lin Zhang

Systems EngineerDetected ALSS anomalies during testing but was overruled by schedule pressure. Data-driven, escalating urgency.Final state: Valence -0.20, Arousal 0.94, Energy 116.7

Raj Mehta

Flight EngineerBrilliant systems analyst, conflict-averse personality. Detects problems early but hesitates to escalate.Final state: Valence -0.20, Arousal 0.81, Energy 123.7

Thomas Webb

Mission DirectorPrioritized schedule and budget over safety margins. Made key decisions to reduce crew size and accept simplified life support.Final state: Valence -0.17, Arousal 0.78, Energy 118.4

Running the Simulation

Clone the repository

git clone https://github.com/timepoint-ai/timepoint-pro.git
cd timepoint-pro
pip install -r requirements.txt

Set your API key

export OPENROUTER_API_KEY=your_key_here

Run the simulation

# Quick mode: 5 backward steps (~$0.18, 44 minutes)
./run.sh run mars_mission_portal --portal-quick

# Full depth: 10 backward steps (~$0.40, 103 minutes)
./run.sh run mars_mission_portal

Output Artifacts

The simulation generates rich, queryable outputs across multiple formats:

Dialogs
Causal Graph
Entity States
Training Data

6 conversations, 78 dialog turns (quick mode)Each timepoint produces a multi-party dialog with per-character generation:

Timepoint 2 — 2029 — Oxygen Generator Flaw

Lin Zhang: Thomas, I've been going over the oxygen generator 
           test results and I think we have a problem.

Thomas Webb: What makes you think that, Lin?

Lin Zhang: The oxygen generator's CO2 scrubbing efficiency is 
           30% lower than expected, and I'm seeing some anomalies 
           in the pressure regulator's performance data, it's not 
           just a minor deviation, Thomas.

Voice distinctiveness: 0.91-0.97 across all entity pairs
Independent LLM calls per character with persona-derived parameters
Fourth Wall context: back layer shapes voice, front layer provides content

Backward timeline from 2031 → 2026

Step 0 (2031): Mission failure during orbital insertion
Step 1 (2030): Raj patches comm anomalies; Webb reduces bandwidth
Step 2 (2029): Lin discovers O2 generator design flaw (30% efficiency loss)
Step 3 (2028): O2 generator at 92%; Lin demands redesign; Webb pushes back
Step 4 (2027): Lin finds 30% failure probability; proposes redundant system
Step 5 (2026): Origin: O2 generator pressure fluctuations first detected

Each backward step selected from 3 candidates scored by 405B judge model.

Final cognitive tensors (after 6 timepoints):

{
  "lin_zhang": {
    "emotional_valence": -0.20,
    "emotional_arousal": 0.94,
    "energy_budget": 116.7,
    "initial_knowledge": [
      "ALSS design specifications",
      "O2 generator test results",
      "System redundancy architecture"
    ]
  }
}

Lin Zhang’s high arousal (0.94) and negative valence reflect sustained frustration from having technical warnings dismissed.

20 structured training examples (prompt/completion pairs)Each example includes:

M7 Causal History: Timeline leading to current moment
M13 Relationship Context: Trust/alignment with other entities
M3 Knowledge Provenance: Source and confidence of each knowledge item
M6 Entity State: Energy, valence, arousal, behavior vectors

Example structure:

=== CAUSAL HISTORY (M7) ===
Timeline leading to current moment (3 events):
  tp_000_2031: Mission failure during orbital insertion...
  tp_001_2030: Raj patches comm anomalies, Webb reduces bandwidth...
  tp_002_2029: Lin discovers oxygen generator design flaw...

=== PREDICTION TASK ===
Predict: new knowledge, energy change, emotional impact, causal reasoning.

Key Insights from Example Run

Run ID: run_20260218_091456_55697771 | Cost: $0.18 | Duration: 44 min | 479 LLM calls | 318K tokens

The Causal Chain Reveals Structural Failure

Schedule pressure emerges as the dominant institutional failure mode:

2026: Lin detects O2 generator pressure fluctuations and overheating
2027: Lin finds 30% failure probability; proposes redundant system with existing components
2028: O2 generator at 92% efficiency; Lin pushes for redesign; Webb demands schedule adherence
2029: Lin discovers 30% CO2 scrubbing efficiency loss; Webb dismisses for schedule
2030: Raj patches around comm anomalies; Webb reduces bandwidth over Lin’s objections
2031: Mission failure

Lin’s oxygen generator concerns appear at every timepoint and are dismissed or deferred each time by Webb’s schedule-first culture.

Emotional Arc Analysis

Lin Zhang has the highest arousal (0.94) in the cast—the most activated character. Her negative valence (-0.20) combined with near-maximum arousal drives repeated confrontations with Webb, citing specific data:

30% CO2 scrubbing efficiency loss
30% failure probability
10% pressure variance
5-degree hourly temperature spikes

Her energy (116.7, lowest in cast) reflects the draining cost of being the persistent technical conscience.

ADPRS Waveform Gating

24 evaluations, 8 divergent (33.33%). Sarah Okafor and Thomas Webb both hit the φ ceiling (1.0) at timepoints tp_000 through tp_003, mapping them to trained band, but they actually resolve at dialog. This is soft budget mode working as intended—spending enough to maintain quality without escalating to the highest tier.

Template Configuration

Here’s the actual template JSON (excerpt):

{
  "scenario_description": "PORTAL backward reasoning from a failed crewed Mars mission in 2031...",
  "world_id": "mars_mission_portal",
  "entities": {
    "count": 4,
    "types": ["human"],
    "initial_resolution": "trained"
  },
  "timepoints": {
    "count": 10,
    "resolution": "month"
  },
  "temporal": {
    "mode": "portal",
    "portal_year": 2031,
    "origin_year": 2026,
    "backward_steps": 10,
    "candidate_antecedents_per_step": 3,
    "path_count": 5,
    "coherence_threshold": 0.7,
    "use_simulation_judging": true,
    "judge_model": "meta-llama/llama-3.1-405b-instruct"
  },
  "metadata": {
    "mechanisms_featured": [
      "M17_modal_temporal_causality_portal",
      "M3_causal_attribution",
      "M7_emotional_state",
      "M8_embodied_states",
      "M11_dialog_synthesis",
      "M13_relationship_tracking"
    ]
  }
}

Mechanisms in Action

M17: Portal Backward Inference

Generates 3 candidate antecedent states per step, scores each with 405B judge model (direct state scoring, no mini forward-simulations), selects the most coherent backward chain.

Known endpoint (2031): Mission failure
                |
    Generate 3 candidate causes
    Score each candidate with 405B judge
    Select best candidate
                |
        Step back 1 year
        Repeat 10 times (or 5 with --portal-quick)
                |
Result: Causal chain from Jan 2026 -> 2031

M3: Knowledge Provenance

Every fact an entity knows has a tracked exposure event with source, timestamp, and confidence score.Example exposure chain:

[pre_tp_000] sarah_okafor <- scene_initialization (conf=1.0):
  "Ares III mission objectives"
  "Crew member skills and expertise"
  "NASA mission protocols"

In PORTAL mode, front-layer knowledge is filtered by causal ancestry so characters only reference information from timepoints upstream of their position.

M11: Per-Character Dialog Synthesis

Each conversation generated through LangGraph steering loop:

Steering agent selects next speaker based on narrative goals
Character agent generates ONE turn using PersonaParams derived from tensor state (arousal → temperature, energy → max_tokens)
Quality gate evaluates after each turn
Loop continues until steering agent ends dialog

Voice distinctiveness scores: 0.91-0.97 across all entity pairs.

M13: Relationship Tracking

Relationships evolve across timepoints with trust and alignment values:

{
  "raj_mehta -> thomas_webb": {
    "type": "tense",
    "trust": 0.40,
    "alignment": 0.30
  },
  "lin_zhang -> sarah_okafor": {
    "type": "collaborative",
    "trust": 0.65,
    "alignment": 0.55
  }
}

Cost Comparison: Quick vs. Full Mode

Metric	Full Run	Quick Run	Change
Cost	$0.49	$0.18	-63%
LLM calls	912	479	-47%
Tokens	900K	318K	-65%
Backward steps	10	5	-50%
Timepoints	11	6	-45%
Dialogs / turns	11 / 143	6 / 78	-45%
Training examples	40	20	-50%
Duration	~103 min	~44 min	-57%

The --portal-quick flag is designed for fast demos and iteration while maintaining the same per-step quality.

Next Steps

Run Convergence Testing

./run.sh convergence e2e mars_mission_portal

Run 3x and analyze causal graph stability

Export to Formats

./run.sh export last --format jsonl
./run.sh export last --format fountain

Training data or screenplay formats

Try Other Templates

Explore the full-mechanism showcase with branching survival strategies

Create Custom Scenario

Learn to build your own PORTAL scenario from scratch

Model Licensing: This example run used Llama models. Meta’s license restricts using Llama outputs to train non-Llama models. For training data generation, use DeepSeek (MIT) or Mistral (Apache 2.0): ./run.sh run --model deepseek/deepseek-r1 mars_mission_portal

Getting Started

Core Concepts

Temporal Modes

Mechanisms

Guides

Examples

Mars Mission Portal

Mars Mission Portal: Backward Reasoning from Disaster

The Scenario

Cast of Characters

Sarah Okafor

Lin Zhang

Raj Mehta

Thomas Webb

Running the Simulation

Output Artifacts

Key Insights from Example Run

The Causal Chain Reveals Structural Failure

Emotional Arc Analysis

ADPRS Waveform Gating

Template Configuration

Mechanisms in Action

Cost Comparison: Quick vs. Full Mode

Next Steps

Run Convergence Testing

Export to Formats

Try Other Templates

Create Custom Scenario

Build docs developers (and LLMs) love

Getting Started

Core Concepts

Temporal Modes

Mechanisms

Guides

Examples

​Mars Mission Portal: Backward Reasoning from Disaster

​The Scenario

​Cast of Characters

Sarah Okafor

Lin Zhang

Raj Mehta

Thomas Webb

​Running the Simulation

​Output Artifacts

​Key Insights from Example Run

​The Causal Chain Reveals Structural Failure

​Emotional Arc Analysis

​ADPRS Waveform Gating

​Template Configuration

​Mechanisms in Action

​Cost Comparison: Quick vs. Full Mode

​Next Steps

Run Convergence Testing

Export to Formats

Try Other Templates

Create Custom Scenario

Build docs developers (and LLMs) love

Mars Mission Portal: Backward Reasoning from Disaster

The Scenario

Cast of Characters

Running the Simulation

Output Artifacts

Key Insights from Example Run

The Causal Chain Reveals Structural Failure

Emotional Arc Analysis

ADPRS Waveform Gating

Template Configuration

Mechanisms in Action

Cost Comparison: Quick vs. Full Mode

Next Steps