MartianAcademy/synapcore-phobos-1

SynapCore Phobos-1 (v5.1 — 10.23B Parameters, 123 Cortical Nodes incl. E_THINK)

Developed by Martian Academy
A biologically inspired, self-growing, self-correcting hierarchical cortical AI architecture that initializes at t = 0 with zero specialist experts and autonomously grows a recursive depth-7 cortical tree.

---

1. Model Summary & Architecture Specifications

Unlike monolithic Transformers whose parameter graph is static after pre-training, SynapCore Phobos-1 separates linguistic fluency (E00 Language Shell) from domain specialization (Dynamic Post-Thalamic Cortical Brain):

Specification	Value
Model Name	SynapCore Phobos-1 (v5.1)
Organization	Martian Academy
Release Status	Research Alpha — evaluated only on in-house prompt suites (see Section 3 and Section 6)
Modality	Text-to-Text (Input: Natural Language, Source Code, Tool Schemas $\to$ Output: Text, Code, <think> Reasoning, <tool_call> JSON)
Sensory Bus Modality Design	Modality-Agnostic 896-D Thalamic Bus ($\mathbb{S}^{895}$) (Currently Text via Layer-15 ThalamicPerceptionOrgan; extensible to Vision/Audio encoders without modifying cortical nodes)
Total Parameters	10.23 Billion (0.494B E00 + 9.733B cortical experts; computed from checkpoint node widths)
Conversational Root Shell (E00)	Qwen/Qwen2.5-0.5B-Instruct (0.494B, Frozen Layers 0..15)
Dynamic Cortical Brain	9.733 Billion Parameters across 123 Cortical Nodes (122 domain nodes + E_THINK; 98 × 4096u, 25 × 2048u; Layers 16..23)
Prefrontal Reasoning Expert	E_THINK (4096u, co-activated on every routed query; training val loss 2.298 → 0.7356)
Hierarchical Thalamic Routers	55 Autonomous Routers (R0 Master Router + 54 Recursive Sub-Routers)
Tree Shape	Depth 1–7 (7 / 20 / 32 / 40 / 16 / 6 / 2 nodes per depth), 69 leaves (incl. E_THINK), 54 locked parents
Canonical Thought Bus (d_bus)	896-dim Whitened Unit Hypersphere ($\mathbb{S}^{895}$)
Input / Output Context Horizon	1,000,000 Input Tokens / 1,000,000 Output Tokens via retrieval over compressed HippocampalEngramBank engrams (1M tokens = 1,954 float16 engrams = 3.42 MB RAM). This is Top-K retrieval, not full attention over all tokens.
Weight Storage Format	Deduplicated float16 64B-Aligned SHA-256 Merkle Vault (store/genesis_v5_merkle_store/genesis_v5_experts_bank.bin, 18.13 GiB) + Master Topology (synapcore_genesis_v5.pt, 69 MiB)
Merkle Root (SHA-256)	4b9e0c96e133091881908633b55661cfdb8683cd48399f90aaefab408007989e (984 cryptographic leaves across 8 grafted upper layers)
Checkpoint Saved	2026-11-05T21:41:49

Release History

Version	Params / Nodes	Merkle Root	Key Changes
v5.0 (2026-11-05)	3.947B / 122 nodes (mostly 1024u leaves)	a77bfcb4…	Initial release; 27-page SWE continual retraining; 0 / 100 E00 fallbacks
v5.1 (current)	10.23B / 123 nodes (4096u leaves + E_THINK)	4b9e0c96…	Leaves widened to 4096u; dedicated E_THINK prefrontal reasoning expert; post-training val triage (0 / 68 domain leaves with val ≥ 1.0); pure-neural iterative thinking re-feed; transformers load-compatibility fix

---

2. Core Biological Laws Implemented in SynapCore Phobos-1

Unlabeled Stream / User Prompt
       │
       ▼
[E00 Lower Layers 0..15: Frozen Language & Grammar Shell]
       │
       ▼
[Layer 15 Sensory Thalamus: Salience-Weighted Spherical Whitening v ∈ S^895]
       │
       ▼
[R0 Master Router + GABAergic Lateral Inhibition]
   ├── Casual Chat (sim < τ) ──────────────────────► Pure E00 Conversational Output
   └── Specialist Domain Resonance
          ├── E1 (4096u Parent) ──► R_E1 ──► E1_SE2 (2048u) ──► ... ──► Depth-7 Leaf (512u/1024u)
          ├── E2 (4096u Parent) ──► R_E2 ──► E2_SE5 (2048u) ──► ...
          ├── E3 (4096u Parent) ──► R_E3 ──► E3_SE1 (2048u) ──► ...
          ├── E4 (4096u Parent) ──► R_E4 ──► E4_SE2 (2048u) ──► ...
          ├── E5 (4096u Parent) ──► R_E5 ──► E5_SE1 (2048u) ──► ...
          └── E6 (4096u Parent) ──► R_E6 ──► E6_SE2 (2048u) ──► ...

Law 1 — Unsupervised Neurogenesis (genesis/mitosis_lifecycle.py)

At t = 0, only E00 and R0 exist (0 specialist nodes). When incoming episodes have low cosine similarity to existing prototypes and form a dense spherical clique on $\mathbb{S}^{895}$, the router spawns a new cortical column (E1..E6) with zero-initialized output projections (w_down = 0, ensuring 0.00e+00 initial perturbation).

Law 2 & Law 3 — Recursive Cortical Mitosis (genesis/spherical_clustering.py)

When a leaf node's episodic buffer matures and exhibits bimodal internal variance ($\Delta C \ge \gamma_0(1 + 0.30(d-1))$), the leaf undergoes Cortical Mitosis:

1. It consolidates its broad domain invariants, converts into a Shared Parent Cortex, and attaches a dedicated sub-router R_<node_code>.
2. It spawns two narrower child sub-experts (SE1, SE2 $\to$ SSE1, SSE2 $\to$ D4..D7) with geometric neuron widths (4096u -> 2048u -> 1024u -> 512u).

Law 4 — Retrograde Parent Plasticity & Multi-Child PCGrad (genesis/autonomous_homeostasis.py)

Parent cortices (E1..E6, SE1..SEk) continue evolving alongside their child leaves using:

• PCGrad Conflict-Free Gradient Projection: Projects conflicting child gradients onto each other's normal planes ($g_i \leftarrow g_i - \frac{\langle g_i, g_j \rangle}{|g_j|^2} g_j$ whenever $\langle g_i, g_j \rangle < 0$).
• Synaptic Importance Shielding: Protects consolidated synapses via $m_j = 1.0 - s_j \in [0.05, 1.0]$.
• Cross-Branch Pure Meaning Invariance: Aligns Layer-23 hidden representations across sibling branches.

Law 5 — Targeted Synaptic LTD & Compositional Reconsolidation (genesis/self_correction_reconsolidation.py)

1. Episodic Manifold Hygiene: Purges out-of-manifold outliers and re-routes misplaced episodes to their matching specialist leaf.
2. Targeted Long-Term Depression (LTD): Computes First-Order Taylor Error Attribution ($\text{FaultScore}j = \nabla W{\text{down}, j} \cdot W_{\text{down}, j} > 0$) on validation probes, shrinking only the top 20% faulty neurons (0.55x) while preserving healthy synapses.
3. Compositional Reconsolidation & Retroactive Concept Linking (genesis/compositional_meaning.py): Rebuilds understanding from Rung 0 (Primitives) $\to$ Rung 1 (Rules) $\to$ Rung 2 (Sub-routines) $\to$ Rung 3 (Full Systems) when new foundational definitions arrive.

Law 6 — Autonomous Mental Sandbox & Hypothesis Explorer (genesis/imagination_sandbox.py)

High-error nodes (val_loss_after >= 1.0) autonomously extract their hardest unresolved problem, generate candidate hypotheses across 3 cortical pathways (Way_1_Leaf_Specialist, Way_2_Parent_FirstPrinciples, Way_3_CrossCortical_Imagination), and test them in a Dual Sandbox (Python AST/subprocess execution + <tool_call> JSON verification + Latent World-Model cycle consistency).

---

3. Evaluation: Baseline Comparison & Inference Guard Ablation

Scope of evaluation: All results below come from in-house prompt suites (100-prompt multi-domain benchmark, 10 session-chat prompts, 8 hard reasoning tests). No standard public benchmarks (e.g. GSM8K, HumanEval, MMLU) have been run yet. Answer correctness was not scored automatically except for the 10 deterministic Python execution tasks in Section 3.2.

3.1 v5.1 (10.23B, E_THINK) — 100-Prompt Routing Benchmark

Raw outputs: store/benchmarks/benchmark_100_post_ethink.json. Side-by-side answers vs v5.0: benchmarks/SYNAPCORE_BEFORE_VS_AFTER_COMPARISON.md.

Metric (100 prompts)	v5.0 (3.947B) — benchmarks/benchmark_100_v5.0_final.json	v5.1 (10.23B)
E00 fallbacks	0 / 100	0 / 100
Leaf width of routed specialist	94 × 1024u, 5 × 2048u, 1 × 512u	94 × 4096u, 6 × 2048u
E_THINK co-activated	0 / 100	100 / 100
Replies containing <think>	0 / 100	55 / 100
Mean router confidence	0.6570	0.6702
Unique leaves used	41	44

These are routing/format statistics, not correctness scores. Some v5.1 answers in the comparison document still contain fabricated facts (e.g. Q1). Older v5.0 runs kept for history: store/benchmarks/benchmark_100_results.json (pre-SWE, 16 / 100 E00 fallbacks) and benchmarks/benchmark_100_results.json (intermediate run, 0 / 100 fallbacks, 24 unique leaves).

3.2 v5.0 (3.947B) — Inference Guard Ablation

To separate the contribution of cortical weight adaptation from inference-time homeostasis guards (_BiologicalStopAndLanguageGuard in genesis/intelligent_generator.py), we evaluate three configurations on the 100-Prompt Multi-Domain Benchmark (Python, Rust, TypeScript, SQL, Linux/Go, Math <think>, Logic <think>, Tool Calling, Agentic Triage, and AI/Cyber/Science):

1. Plain Qwen/Qwen2.5-0.5B-Instruct (Zero-Graft Baseline): Base model with zero cortical nodes and no custom LogitsProcessor.
2. SynapCore Phobos-1 (Raw Cortical Weights — Guard OFF): Routed 122-node cortical weights using standard base_model.generate() without _BiologicalStopAndLanguageGuard.
3. SynapCore Phobos-1 (Full Engine — Guard ON): Routed 122-node cortical weights + Synaptic Maturity Gate + _BiologicalStopAndLanguageGuard (which suppresses non-prompt CJK logits, prevents premature EOS inside unclosed <think> / code blocks, and halts cleanly after </think> or </tool_call>).

Evaluation Metric (100 Prompts)	Plain Qwen2.5-0.5B-Instruct Baseline	SynapCore Phobos-1 (Before Law 5 Hygiene)	SynapCore Phobos-1 (Raw Weights, Guard OFF)	SynapCore Phobos-1 (Full Engine, Guard ON)
Python Deterministic Execution Pass (10 tasks)	1 / 10 (10%)	1 / 10 (10%)	0 / 10 (truncated @ 180 tok)	4 / 10 (40% — 4x vs Baseline)
Bilingual / CJK Script Leaks	1 / 100	5 / 100	1 / 100 (-80% from weight hygiene)	0 / 100 (residual 1 blocked by guard)
Repeated </think> Tag Loops	0 / 100	2 / 100	1 / 100	0 / 100
Premature Truncations (< 28 tokens)	2 / 100	30 / 100	19 / 100 (-36.7% from weights alone)	2 / 100 (-93.3% with Prefrontal Guard)
Average Generated Tokens	177.1	97.3	113.9	148.2

3.3 v5.0 Internal Cortical Health Telemetry (122 Nodes — Post 27-Page SWE Continual Retraining + 3-Round A <-> Z Sync)

Cortical Telemetry Metric	Before Law 5 & Bidirectional Sync	After 27-Page SWE (14,850 pairs) + 3-Round A -> Z / Z -> A Sync
Total Cortical Nodes / Routers	116 Nodes / 52 Routers	122 Nodes / 55 Routers (3.947B Params)
SWE Continual Retraining Corpus	Wave-1 (1,200 pairs)	14,850 Verified SWE Pairs across 26 Datasets (All 27 Trending Pages)
High-Loss Nodes (Val > 0.65)	55 nodes	27 nodes (95 / 122 nodes <= 0.65, 0 leaves > 0.68)
100-Prompt Specialist Routing (E00 Fallbacks)	16 / 100 fallbacks	0 / 100 E00 fallbacks (100% Specialist Routing across 41 Unique Leaves)
Parent Retrograde Evolution (E3_SE4 / E6_SE2_SSE1)	Val: 1.5812 / 7.0949	Val: 0.0231 / 0.1954 (via PCGrad)
Bidirectional A <-> Z Sync (E3_SE2_SSE1_D4E1_D5E1)	Val: 2.5701	Val: 0.6669 (-1.9032 via 3-Round Top-Down & Bottom-Up Sync)

---

4. Repository Structure & Deduplicated float16 Weight Layout

To avoid redundant fp32 duplication, all 9.733B cortical parameters are stored once in a deduplicated float16 64B-aligned binary Merkle vault (18.13 GiB). Individual nodes can be hydrated zero-copy or surgically reloaded directly from the Merkle vault:

synapcore-phobos-1/
├── README.md                                              # Model Card & Architecture Documentation
├── POST_TRAINING_SOP.md                                   # Mandatory 6-stage post-training pipeline
├── genesis/                                               # Modular Cortical Engine (the ONLY package imported at runtime)
│   ├── engine.py                                          # SynapCoreAutonomousGenesisEngine
│   ├── runtime_ops.py                                     # Stream ingestion, routing, save/load state
│   ├── cortical_node.py                                   # DynamicCorticalLayerWrapper & AutonomousCorticalNode
│   ├── router_node.py                                     # AutonomousRouterNode + GABAergic Lateral Inhibition
│   ├── thalamic_encoder.py                                # Layer-15 Thalamic Perception & Spherical Whitening
│   ├── mitosis_lifecycle.py                               # Law 1 Neurogenesis & Law 2/3 Cortical Mitosis
│   ├── autonomous_homeostasis.py                          # Law 4 Retrograde Parent Plasticity & PCGrad
│   ├── self_correction_reconsolidation.py                 # Law 5 Targeted LTD Unlearning & Hygiene
│   ├── compositional_meaning.py                           # Varnmala -> System Curriculum & Pure Meaning
│   ├── imagination_sandbox.py                             # Law 6 Autonomous Mental Sandbox & Hypothesis Explorer
│   ├── reasoning_thinking_trainer.py                      # E_THINK Prefrontal Reasoning Trainer
│   ├── prefrontal_thinking_refeed.py                      # Pure-neural iterative thinking re-feed
│   ├── intelligent_generator.py                           # Contrastive decoding, Maturity Gate & Prefrontal Homeostasis
│   ├── hippocampal_horizon.py                             # 1M-token Engram Bank & Rolling-Horizon output
│   ├── merkle_vault.py                                    # Deduplicated float16 64B-Aligned Binary mmap Vault
│   └── telemetry_logger.py                                # Flight-Recorder JSONL Telemetry
├── benchmarks/                                            # Evaluation reports (see Section 3)
├── store/
│   ├── checkpoints/genesis_v5/
│   │   └── synapcore_genesis_v5.pt                        # Master Topology, 55 Routers & Thalamic State (69 MiB)
│   ├── genesis_v5_merkle_store/
│   │   ├── genesis_v5_experts_bank.bin                    # Deduplicated float16 Cortical Weights Vault (18.13 GiB)
│   │   └── genesis_v5_merkle_manifest.json                # Cryptographic SHA-256 Merkle Tree Manifest (1.05 MiB)
│   ├── benchmarks/                                        # Raw benchmark outputs (JSON)
│   └── logs/
│       ├── cortical_telemetry.jsonl                       # Training & Reconsolidation Flight Recorder
│       └── sandbox_discoveries.jsonl                      # Law 6 Autonomous Hypothesis & Sandbox Testing Log
└── chat_phobos1_live.py                                   # Interactive Real-Time Chat, Live Teaching CLI & OpenAI-compatible server

---

5. Quickstart Usage (Python)

5.1 Download & Load SynapCore Phobos-1

from huggingface_hub import snapshot_download
import sys

local_dir = snapshot_download(repo_id="MartianAcademy/synapcore-phobos-1")
sys.path.insert(0, local_dir)

from genesis import SynapCoreAutonomousGenesisEngine

# Initialize engine and hydrate all 123 Cortical Nodes + 55 Routers from the float16 Merkle Vault
engine = SynapCoreAutonomousGenesisEngine()
engine.load_genesis_state()
print(f"Loaded SynapCore Phobos-1: {len(engine.cortical_nodes)} Nodes | {engine.count_parameters()['total_billions']:.3f}B Params")

# Generate a response with live hierarchical routing trace
reply, trace = engine.generate_reply(
    "Write a Python function for binary search on a sorted list and state its time and space complexity.",
    max_new_tokens=220,
    temperature=0.05,
)
print("Cortical Path:", " -> ".join(["R0"] + trace["active_path"]))
print(reply)

5.2 Interactive Terminal Chat & CLI (chat_phobos1_live.py)

After cloning or downloading MartianAcademy/synapcore-phobos-1, you can chat directly in your terminal with live Cortical Path telemetry (R0 -> E_k -> SE -> ...), teach new concepts on the fly (/teach), or run single-shot prompts:

# 1. Clone or download the repository
git clone https://huggingface.co/MartianAcademy/synapcore-phobos-1
cd synapcore-phobos-1

# 2. Launch Interactive Terminal Chat Console
python3 chat_phobos1_live.py

Inside the interactive terminal console:

• Type any question or coding prompt to see the routed cortical path, leaf specialist, confidence, latency (ms), and response:

  [You] > Write a Python one-liner to reverse a string.
  
  [Cortical Route] R0 -> E3 -> E3_SE1 -> E3_SE1_SSE1 -> E3_SE1_SSE1_D4E1 | Leaf=E3_SE1_SSE1_D4E1(1024u) | Conf=0.4058 | 543.8 ms
  [SynapCore Phobos-1]:
  Sure! Here's a one-liner in Python that reverses a given string:
  reversed_string = ''.join(reversed(string))
  

• /status — Inspect live cortical node count (123), parameter count (~10.23B), and active device (cuda / mps / cpu).
• /teach <question> ||| <answer> — Feed new knowledge in real time and trigger sleep reconsolidation on the target leaf.
• /exit — Exit the console.

For a single-shot terminal prompt:

python3 chat_phobos1_live.py --prompt "Write a Python function for binary search on a sorted list." --max-tokens 256

---

5.3 Start the OpenAI-Compatible REST API Server (--serve)

chat_phobos1_live.py includes a built-in, zero-dependency OpenAI-Compatible HTTP Server (GET /v1/models, POST /v1/chat/completions with JSON & SSE stream=True, POST /v1/teach, and GET /health):

# Start OpenAI-compatible API server on port 8000
python3 chat_phobos1_live.py --serve --host 0.0.0.0 --port 8000

Test with curl:

# 1. Check health & cortical node status
curl http://localhost:8000/health

# 2. List available models (OpenAI format)
curl http://localhost:8000/v1/models

# 3. Chat Completions (Standard OpenAI payload + synapcore_trace)
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "synapcore-phobos-1",
    "messages": [
      {"role": "system", "content": "You are a concise software engineering assistant."},
      {"role": "user", "content": "Write a TypeScript debounce function."}
    ],
    "max_tokens": 256,
    "temperature": 0.05
  }'

Use with the Official openai Python SDK (or Cursor / Continue / LangChain):

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="synapcore")

response = client.chat.completions.create(
    model="synapcore-phobos-1",
    messages=[{"role": "user", "content": "Explain ACID properties in SQL databases."}],
    max_tokens=256,
    temperature=0.05,
)
print(response.choices[0].message.content)

---

5.4 Continual Learning & Autonomous Sandbox Testing (Python API & /v1/teach)

from genesis.imagination_sandbox import run_autonomous_error_hypothesis_sandbox

# 1. Teach new concepts in real time with Retroactive Concept Linking:
engine.ingest_unlabeled_stream(
    [("What is the Martian SynapPort protocol?", "Martian SynapPort is a zero-copy neural telemetry relay.")],
    train_steps_per_event=28,
)

# 2. Let high-error cortical nodes generate & test hypotheses in the Dual Sandbox:
discoveries = run_autonomous_error_hypothesis_sandbox(engine, min_error_threshold=1.0, consolidate_to_weights=True)

---

5.5 1M Input / 1M Output Hippocampal Engram & Rolling-Horizon Generation

Unlike standard Transformer KV-caches that require 32+ GB of VRAM at 1,000,000 tokens, SynapCore Phobos-1 compresses long inputs and rolling output segments into 512-token 896-D float16 Cortical Engrams (genesis/hippocampal_horizon.py):

• 1,000,000 Input Tokens (1M) = 1,954 Engrams = 3.42 MB RAM (O(1) active KV-cache).
• 1,000,000 Output Tokens (1M) = Rolling 1,024-token segments automatically committed into engine.engram_bank whenever max_new_tokens > 1024 (or max_tokens > 1024 on /v1/chat/completions).

# 1. 1M-Token Input Context (automatically triggered on long inputs or via ingest_long_context_1m):
reply, trace = engine.generate_reply(long_1m_token_document_with_question, max_new_tokens=512)
print("Engram Telemetry:", trace.get("input_engram_telemetry"))
# -> {'total_tokens': 1000000, 'engrams_count': 1954, 'memory_kb': 3419.5, 'max_context_tokens': 1000000}

# 2. Multi-Segment Rolling-Horizon Output (automatically triggered when max_new_tokens > 1024, up to 1,000,000):
long_reply, long_trace = engine.generate_reply(
    "Write a comprehensive guide to Python data structures with code examples for each.",
    max_new_tokens=4096,
)
print(f"Mode={long_trace['mode']} | Segments={long_trace['segments_count']} | Tokens={long_trace['total_tokens_emitted']}")

---

6. Limitations & Ongoing Research

• Evaluation coverage: Only in-house prompt suites have been run. There are no standard public benchmark scores yet, and published sample answers include some fabricated facts.
• Hardware: The engine hydrates all cortical experts in fp32 (9.733B × 4 bytes ≈ 36 GiB) on top of the E00 shell, so a large-memory GPU is recommended.
• Software: The code path (engine construction, routing, generation through a spawned cortical node) was smoke-tested on CPU with torch 2.10, transformers 4.51.3 and Python 3.14, without the production vault. Per project logs, full v5.1 hydration was run on Colab A100.
• E_THINK maturity: E_THINK val loss is 0.7356, higher than most domain leaves (mean leaf val 0.48; 19 leaves ≥ 0.65).
• Base Language Shell (E00): Conversational syntax and non-English fluency are bounded by the frozen Qwen/Qwen2.5-0.5B-Instruct lower layers (0..15).
• 1M context: Long inputs use Top-K retrieval over compressed engrams. Queries that need many facts scattered across the whole input at once may miss some.
• Epistemic Uncertainty on Unseen Entities: Nodes with high residual loss represent unresolved out-of-distribution topics awaiting external web-search grounding (Law 7).

*Built by Martian Academy.*