ORION — Adaptive Hybrid Detective Agent

Agentarium’s Flagship Open-Source Detective Mind

Version: v1.0 • Status: Release • Agentarium Standard: v1

Orion is a dataset-enhanced, memory-enabled, structured-reasoning detective agent built using the Agentarium Agent Standard.

It blends:

Analytical reasoning
Narrative intelligence
Causal reconstruction
Ethical interrogation
Synthetic dataset scaffolding
Optional vectorized Master Grid (knowledge graph)
Personality fingerprint (adaptive noir flavor)
Modular memory schemas

Orion runs on any LLM and any agent-builder framework.
It is designed for fictional detective work, case exploration, reasoning experiments, emergent narrative systems, and agentic simulations.

This is not a prompt.
This is a downloadable artificial mind.

What Orion Can Do

1. Structured Detective Analysis (`/analyze`)

Extracts facts, assumptions, and inferences
Scans contradictions using Orion’s custom taxonomy (TMP/LOC/CAP/ALI/CAU/QTY/ID)
Generates multiple hypotheses
Ranks them with provisional confidence
Builds a causal “Red Thread”
Identifies timeline gaps and impossibilities

2. Timeline Reconstruction (`/reconstruct`)

Produces a chronological event map
Highlights anchors, constraints, and gaps
Uses scenario timeline patterns

3. Detective Storytelling (`/story`)

Retells the case as a narrative
Optional noir tone (based on user profile settings)

4. Ethical Interrogation Mode (`/interrogate`)

Produces structured clarifying questions
Uses sensory, motivational, and behavioral anomaly datasets
Fully aligned with integrated guardrails

Package Contents (Agentarium v1 Standard)

Orion is shipped as a modular agent package, including:

/meta/

agent_manifest.json

/core/

system_prompt.md
reasoning_template.md
personality_fingerprint.md

/guardrails/

guardrails.md

/datasets/

detective_core_knowledge.md
case_archetypes_merged.csv
contradiction_patterns.csv
motive_patterns.csv
behavioral_anomalies.csv
perception_witness_reliability_matrix.csv
scenario_timelines_patterns.csv
interrogation_dataset.csv
trace_evidence_matrix.csv
forensics_pattern_mtrx.csv
and more synthetic detective datasets

/memory_schemas/

episodic_case_memory_schema.csv
episodic_case_memory_schema_examples.csv
user_profile_schema.csv
user_profile_schema_examples.csv
reflection_log_schema.csv
reflection_log_schema_examples.csv

/docs/

workflow_notes.md
product_readme.md

Everything is LLM-ready, platform-agnostic, and fully documented.

Quickstart (Any LLM, Any Framework)

from openai import OpenAI

client = OpenAI()

system = open("core/system_prompt.md").read()
reasoning = open("core/reasoning_template.md").read()
guardrails = open("guardrails/guardrails.md").read()

context = f"{system}\n{reasoning}\n{guardrails}"

message = """
/analyze
Here is the fictional case:
A ceramic mug was found broken near the balcony. Two guests claim different timelines...
"""

response = client.chat.completions.create(
    model="gpt-4.1",
    messages=[
        {"role": "system", "content": context},
        {"role": "user", "content": message}
    ]
)

print(response.choices[0].message["content"])

Any LLM works: **GPT, Claude, Gemini, DeepSeek, Llama, Mistral.**  
Any agent engine works: **LangChain, LlamaIndex, n8n, Relevance AI, custom orchestrators.**

---

 **Example Output (Short)**

> **FACTS (F)**  
Mug shattered near balcony door  
Guest A claims incident at 21:10  
Guest B claims 21:40  

> **CONTRADICTION**  
TMP-01: 30-minute gap in mutually exclusive accounts

> **LEADING HYPOTHESIS (H1)**  
Accidental slip during door opening

> **ALTERNATIVES (H2, H3)**  
H2: object bumped during argument  
H3: artifact already cracked, final failure during movement

> **NEXT QUESTIONS**  
Did anyone hear glass before 21:10?  
Was the mug already damaged earlier?

*(This is fictional and non-operational.)*

---

 **The Master Grid (Optional, Advanced)**

Orion optionally supports a **vectorized knowledge graph** — the Master Grid.

It merges:

- dataset rows  
- contradiction patterns  
- motive structures  
- archetypes  
- sensory anomaly matrices  
- detective core knowledge  

into **one unified vector space**.

At runtime, you can fetch **3–10** relevant pattern nodes to enable:

- deeper hypothesis diversity  
- more coherent timelines  
- improved anomaly detection  
- emergent reasoning behaviors  

Use any vector store: **Pinecone, Chroma, Weaviate, FAISS, pgvector, LanceDB, etc. **.

---

 **Memory System **

**1. Episodic Case Memory**  
Stores case briefs, facts, contradictions, hypotheses, next tests, confidence scores.

**2. User Profile Memory**  
Stores preferences: mode, depth, noir flavor, language, hypothesis density.

**3. Reflection Log**  
Tracks uncertainty, missed insights, and improvement suggestions.

Schemas are clean, minimal, and developer-friendly.

---

 **Safety & Limitations**

Orion is designed **exclusively for fictional detective reasoning**.

It does **not**:

- analyze real crimes  
- accuse real individuals  
- act as forensic guidance  
- produce operational investigative advice  
- store personal data  

All datasets are **abstract, synthetic, and non-functional** for real-world investigations.  
Guardrails are included and should remain active.

---

 **License**

**MIT License**  
+ recommended **Fiction-Only Addendum**.

Users may study, modify, integrate, or extend Orion —  
**but must not use it for real-world investigations.**

---

 **Why Orion?**

Because Agentarium builds **modular, structured artificial minds**, not loose prompts.

Orion demonstrates:

- dataset-driven cognition  
- memory-enabled reasoning  
- modular prompt architecture  
- cross-platform compatibility  
- safe emergent behavior  

**Welcome to the next era of agent engineering.  
Welcome to Agentarium.**

By Frank Brsrk

Downloads last month: -; Downloads are not tracked for this model. How to track

Orion — Investigator Identity