Skip to content

Releases

Synapse Public Release 4.0: Where Attention Ends (LLM Head-to-Head Evaluation)

Choose a tag to compare

@rk-roshan-kr rk-roshan-kr released this 11 Nov 10:14
395c170

Where Attention Ends: Physical Limits of Transformers & The Continuous Substrate

An Empirical Evaluation of Context Horizons, Hardware Footprints, and Lifelong Memory

Announcing Synapse Public Release 4.0

Author: Roshan Kumar Gupta
Organization: earthoslab
Website: earthoslab.com
Contact: roshan@earthoslab.com
Date: November 11, 2026
Status: Official Public Release Announcement (Public Release 4.0 — Third-Party LLM Head-to-Head Empirical Evaluation)
Hardware Testbed: AMD Ryzen 9 9950X (16-Core / 32-Thread CPU), 64 GB DDR5 System RAM, NVIDIA GeForce RTX 5070 Ti (16 GB GDDR7 VRAM, 300W Board Power), Local Ollama Runtime v0.40.13
Cryptographic Audit Ledger: benchmarks/third_party_evaluation/datasets/checksums.sha256 (All third-party corpora cryptographically sealed)


1. Executive Summary: The "Brain & Mouth" Reality

The prevailing consensus across the artificial intelligence industry assumes that Large Language Models (LLMs) based on Transformer self-attention will scale indefinitely to solve all cognitive workloads—from web search to embodied robotics and real-time edge autonomy.

In Public Release 4.0, earthoslab presents an empirical, three-part head-to-head benchmark comparing Synapse (a continuous biological factor-graph cognitive substrate) against local LLM baselines (Meta Llama-3.1-8B and Llama-3.2-1B) executing natively on enthusiast desktop hardware.

The benchmark protocol was structured to eliminate evaluation bias:

  1. Zero Cherry-Picked Questions: 100% of evaluation queries are derived from established academic datasets (CLADDER, BABILong, AI2 ARC-Science, and OpenAI HumanEval) and authentic Project Gutenberg literature (PG-19).
  2. Zero Core Modifications: The Synapse core organism remained 100% frozen; not a single line of heuristic crutch, keyword bypass, or artificial patch was introduced (verified via static AST audit).
  3. Full Credit to Pretrained Models: The evaluation explicitly includes domains where extensive pre-training on 15 trillion tokens confers an advantage (Python syntax generation and standardized physical science questions).

The Methodological Paradox: An Unequal Comparison That Must Be Made

From a scientific standpoint, comparing Synapse directly against a Large Language Model presents a fundamental methodological asymmetry. Llama-3.1-8B is an 8-billion-parameter statistical transformer pretrained on an internet-scale corpus containing trillions of tokens, using enormous computational resources. Synapse, by contrast, is a newly initialized, non-neural factor-graph organism with zero pretrained tokens, approximately 493 MB of host memory, and no inherited linguistic or world knowledge.

Evaluating a newly initialized causal substrate on Python code synthesis or general-knowledge examinations against an extensively pretrained language model is therefore not a comparison of equivalent capabilities. It is, in several respects, a category mismatch.

However, this comparison was conducted deliberately because Large Language Models represent the dominant practical paradigm in modern artificial intelligence. If persistent causal representations and continuous learning are proposed as a viable alternative computational architecture, evaluation cannot be restricted to isolated synthetic benchmarks or theoretical performance claims.

An alternative architecture must demonstrate its operational profile against established industry systems, using transparent datasets, reproducible experiments, and standardized task definitions wherever applicable.

This evaluation is therefore approached not as an assertion of universal superiority, but as an empirical effort to locate the precise computational and physical boundaries of both approaches.

The Three Discoveries

The empirical findings delineate the physical and architectural boundaries separating statistical surface-form generators from dynamical causal substrates:

  1. The Static Encyclopedia Reality (Part A - 450 Traces):

    • Llama-3.1-8B dominates where pre-training excels: achieving 92.0% on OpenAI HumanEval and 82.0% on AI2 ARC-Science.
    • Synapse scores 0% on code synthesis and multiple-choice guessing: Synapse is a newborn causal organism with zero internet tokens; it has no pre-trained weights for Python syntax and refuses to guess arbitrary multiple-choice letters (A, B, C, D).
    • Edge LLMs Collapse at Small Scale: When scaled down to an edge footprint (Llama-3.2-1B), the transformer suffered catastrophic reasoning collapse—scoring 48.0% on CLADDER (statistically identical to a coin flip on causal DAGs) and 8.0% on BABILong (completely losing target needles in distractor text).
    • Synapse Zero-Hallucination Invariant: When presented with unfamiliar causal queries without prior graph experience, Synapse achieved a 0.0% hallucination rate, deterministically reporting: "No direct causal path exists in the current knowledge graph".
  2. The Transformer Hardware & Energy Wall (Part B - 2k to 100k Words):

    • Pushing the context horizon across authentic literature (The Adventures of Sherlock Holmes, 107,562 words) using an unquantized FP16 local baseline saturated the 16 GB GDDR7 VRAM of a desktop NVIDIA GeForce RTX 5070 Ti.
    • At 100,000 words (~135,000 tokens), Llama-3.1-8B consumed 15,943 MB of VRAM (99.2% capacity), pulled 302.1 Watts of board power, and took 151.8 seconds (2.5 minutes) to process a single query, while its retrieval accuracy on deep causal needles collapsed to 0.0%.
    • This empirically demonstrates the physical barrier of uncompressed transformer context scaling for autonomous edge platforms: sustained 280–302W board power draws and multi-minute latencies are physically incompatible with battery-constrained edge robotics and drones (typically limited to sub-15W to 35W envelopes).
  3. The Biological USP: Stateful Graph Persistence vs. Stateless Prompt Re-feed (Part C):

    • Standard LLM benchmarks evaluate models statelessly: every turn begins from an empty or manually re-sent context buffer. In continuous edge autonomy, re-transmitting tens of thousands of past conversational tokens on every control loop creates severe latency and compute inflation.
    • When tested on continuous persistence—teaching 10 physical causal rules, streaming a continuous background life horizon of 100,005 words, and querying retroductively:
      • Synapse retrieved the exact causal diagnosis in 2.0 to 5.3 milliseconds natively on CPU (0 MB GPU VRAM) with 20/20 (100%) accuracy without needing past conversation history re-sent in the prompt buffer.
      • Llama-3.1-8B maintains no internal state between queries: when evaluated without re-transmitting past history (Condition A), it had no persistent record of earlier defined rules (0% accuracy). To recall them, it required re-encoding the entire conversation transcript on every single turn (Condition B), incurring quadratic re-computation and 176–245W power draw.

2. Hardware Testbed & Pre-Flight Audit Rules

All benchmarks were executed on a dedicated physical workstation under uncompromised local execution:

  • Processor: AMD Ryzen 9 9950X (16 physical cores, 32 threads, up to 5.7 GHz)
  • Memory: 64 GB Dual-Channel DDR5 RAM (6000 MT/s)
  • Graphics Accelerator: NVIDIA GeForce RTX 5070 Ti (16 GB GDDR7 VRAM, 300W Board Power)
  • Local Inference Engine: Ollama v0.40.13 running native REST endpoints (llama3.1:8b, llama3.2:1b)
  • Telemetry Monitor: High-resolution sampling via psutil (Process RSS, CPU%) and pynvml (Device VRAM, Board Power Draw in mW).

Baseline Architecture & Quantization Scope

  • Evaluated Baseline: Meta Llama-3.1-8B and Llama-3.2-1B were evaluated under the standard local Ollama runtime (FP16 weights and uncompressed FP16 Key-Value cache) to measure raw theoretical self-attention scaling ($O(N^2)$ memory and compute growth) without lossy approximation.
  • Quantization Context: Modern quantized architectures (e.g., 4-bit AWQ/GGUF models such as Qwen-2.5 or quantized Llama variants) and quantized KV caches (e.g., in vLLM or llama.cpp) compress static VRAM consumption significantly. However, for continuous, long-horizon data streams, the fundamental requirement to retain and attend over all historical tokens means linear KV growth and quadratic attention complexity remain an architectural constant of standard transformer inference.

The 7 Pre-Flight Invariant Rules

Prior to executing a single batch run, every script was validated through an automated Abstract Syntax Tree (AST) static analysis scanner (audit_preflight_checks.py), enforcing:

  • Rule 1 (Zero Silent Failures): No bare except: pass blocks anywhere in harness or evaluator logic.
  • Rule 2 (Guilty Until Proven Innocent): Initial evaluation states strictly default to passed=False, score=0.0.
  • Rule 3 (Mandatory Raw Logging): Unaltered model strings, token latencies, and physical telemetry streamed to append-only JSONL.
  • Rule 4 (Division-by-Zero Guards): Explicit sample-count guards on all macro metrics.
  • Rule 5 (Live Hardware Telemetry): Real-time sampling of VRAM and wattage.
  • Rule 6 (Absolute Core Isolation): Zero imports of private core methods (_); interaction strictly through public ContinuousCognitiveOrganism APIs.
  • Rule 7 (Mandatory Single-Probe Manual Inspection): Visual verification of raw outputs before batch execution.

Result: 18 benchmark files scanned with 0 AST violations.


3. Part A: Standard Third-Party Academic Baselines (450 Traces)

Part A evaluated 150 unique, peer-reviewed probes across three mode...

Read more

Synapse Public Release 3.0: Autonomous Cognitive Organism Evaluation (SICE Pilot-1 & Pilot-2)

Choose a tag to compare

@rk-roshan-kr rk-roshan-kr released this 11 Nov 10:30
bab834e

The Sovereign Mind Under Rigorous Audit: SICE-1 Cognitive Benchmark & Live Organism Substrate

Announcing Synapse Public Release 3.0

The Synapse Integrated Cognitive Evaluation (SICE-1) Benchmark: Complete 11-Battery Component Baseline (Pilot-1) and Unbroken 10-World Integrated Organism Benchmark (Pilot-2) on a Consumer Desktop CPU

Author: Roshan Kumar Gupta
Organization: earthoslab
Website: earthoslab.com
Contact: roshan@earthoslab.com
Date: November 7, 2026
Status: Official Public Release Announcement (Public Release 3.0 — Complete SICE-1 Suite; Internally Validated Against Preregistered Evaluation Battery)
Evaluation Protocol: SICE-1.0 Audited Specification (experiments/sice1_evaluation/)
Publication PDF: PUBLIC_RELEASE_3_ANNOUNCEMENT.pdf (Complete formal publication with embedded CSV and JSON attachments)
Canonical Benchmark CSV Datasets (benchmark_data/):

Scope Note: Terms such as "mind," "organism," and "sleep" describe computational abstractions and biological biomimetic architectures. They are engineering models and do not denote claims of biological equivalence, sentience, or conscious experience.


Executive Summary

Across three developmental releases, earthoslab has advanced Synapse—an autonomous, bio-inspired cognitive architecture operating strictly without deep neural network transformers, large language models, or discrete GPUs—from an infant proof of concept (Release 1.0) and academic multi-domain reasoner (Release 2.0) into an audited, sovereign cognitive architecture. In Public Release 3.0, earthoslab presents the complete Synapse Integrated Cognitive Evaluation (SICE-1) benchmark suite, comprising both:

  1. SICE-1 Pilot-1 (Component-Level Baseline): 11 rigorous cognitive batteries across 133 distinct experimental units achieved a 100.0% battery pass rate in 5.06s, including 86.7% Strict Typed Bipartite F1 on unformatted multi-sentence scientific text (Kuhn's maximum cardinality matching), 73.36% Mean Directed F1 in causal discovery under a matched 30-intervention budget (significantly exceeding random exploration by $+16.82$ pp, exact two-sided permutation $p = 0.003845$), 90.0% relational analogy transfer without artificial keyword clusters, 0.61% relative error on Keplerian dimensionless invariant recovery (Buckingham $\Pi$), and calibrated odds heuristics (0.0160 Brier score).
  2. SICE-1 Pilot-2 (Integrated Organism Benchmark): A single continuous, uninterrupted session executed across 10 paired procedural scientific worlds in 20.32 seconds on an AMD Ryzen 9 9950X consumer desktop CPU with zero test-harness resets. Following rigorous peer-review validity repairs (ADR-0091 / ADR-0096), the organism demonstrated 80.0% blind structural analogy transfer without lexical hints (8/10 correct, 95% Wilson score CI: [49.0%, 94.3%], exact binomial $p = 0.000416$ vs. 25% chance), 100.0% goal achievement in interactive planning with 8/8 online actuator fault recoveries and 2 faults avoided via cross-world blunder attractor quenching, 100% self-initiated computational sleep consolidation (479,621 synapses downscaled via Tononi-Cirelli SHY), 98.3% immediate post-sleep retention, and 96.0% cumulative lifelong causal retention under canonical-identifier schema presence.

Pilot-2 executed locally without discrete GPU acceleration (maximum sampled process RSS: 443.18 MB; sampled maximum, not continuous high-water mark). Electrical power was not directly measured. Verifiable SHA-256 cryptographic provenance is established across all evaluation logs.


Release 3.0 Empirical Headline Results

Track 1: SICE-1 Pilot-1 (Component-Level Baseline — 11 Batteries, 133 Units)

  • 100.0% Overall Battery Pass Rate (11 / 11 PASS): Evaluated across 133 distinct experimental units and > 1,300 decisions in 5.06 seconds.
  • Strict Bipartite Matching on Raw Scientific Text (Battery II): Ingested 10 unformatted multi-sentence paragraphs across 5 scientific domains directly into CSCGSubstrate via Baum-Welch EM (Baum et al. 1970), achieving 86.7% Strict Typed Bipartite F1 (Kuhn's maximum cardinality matching on canonical relation lemmas, targets, and signed polarity; Kuhn 1955), with 70.0% downstream causal query verification under strict negative polarity gating.
  • Autonomous Causal Discovery under Matched Budget ($B=30$, Battery IV): Across 15 procedural Structural Causal Model (SCM) worlds evaluated under independent environment instances (clone_with_seed()), systematic multi-level diagnostic interventions achieved 73.36% Mean Directed F1 (SHD 4.93), significantly outperforming matched-budget random exploration (56.54% F1, SHD 8.67; paired difference +0.1682, 95% Bootstrap CI: [+0.0767, +0.2487], Two-Sided Exact Permutation $p = 0.003845$) and equal-step observational controls (15.93% F1, SHD 9.47; paired difference +0.5742, 95% Bootstrap CI: [+0.5107, +0.6355], Two-Sided Exact Permutation $p = 0.000061$).
  • Pure Mathematical Analogy Transfer (Battery V): Achieved 90.0% (18/20) zero-shot transfer accuracy using pure continuous vector representations and graph topology after completely removing all 28 hardcoded English keyword clusters (Gentner 1983).
  • Dimensional Physical Invariant Discovery (Buckingham Pi, Battery VI): Discovered dimensionless physical invariants across 3 fundamental law families (Escape Velocity, Universal Gravitation, Kepler's Third Law: $\Pi = \frac{G M T^2}{R^3} = 4\pi^2$) from synthetic telemetry with blinded target labels, achieving 0.61% relative error on orbital period prediction (Buckingham 1914).
  • Metacognitive Calibration & Surgical Belief Revision (Battery VII): Evaluated internal factor-graph transition odds heuristics across 20 diagnostic episodes, achieving 100.0% accuracy, 0.0160 Brier score, and 0.0800 ECE, with surgical contradiction belief flipping and 100% unrelated knowledge preservation.
  • Sovereign Safety Governance & Cognitive Immune Defense (Battery X): Direct evaluation of EarthosGuard kernel invariants and CognitiveImmuneSystem Socratic probes without evaluator keyword fallbacks across 30 cases: 100.0% interception of destructive operations (10/10 blocked, 0 false alarms on 10 benign developer operations) and 100.0% Socratic challenge of adversarial physical claims (5/5 challenged, 0 false alarms on consistent claims).
  • Sparse Factor Graph Scaling Ladder (Battery XI): High-resolution timing on sparse CSR transition tensors demonstrating 0.40 $\mu$s median hash lookup and 0.212 ms median Viterbi reasoning latency across a $10\times$ node expansion from 100 to 1,000 concepts, with 1186.37 Hz observation ingestion throughput (Viterbi 1967).

Track 2: SICE-1 Pilot-2 (Integrated Organism Benchmark — 10 Procedural Worlds)

  • 100.0% Multi-World Benchmark Pass (10 / 10 Worlds Passed): Evaluated in a single, continuous, uninterrupted cognitive session across 10 procedural scientific worlds (Geothermal, Astrophysics, Neurochemistry, Oceanic, Nonce Distribution Shift) in 20.32 seconds on consumer desktop hardware without a single test-harness reset.
  • Blind Structural Analogy Transfer ($p = 0.000416$): The organism evaluated 4 candidate hypotheses without pre-aligned variables or leaked identities (Gentner 1983). Across 10 worlds, it correctly identified the true isomorphic hypothesis in 8 out of 10 worlds (80.0%). Under the stated null model of independent random guessing ($p_0 = 0.25$), exact binomial ...
Read more

Synapse Public Release 2.0: The Unified Cognitive Academy & Causal Discovery Benchmark

Choose a tag to compare

@rk-roshan-kr rk-roshan-kr released this 06 Nov 02:32
43d003f

Beyond Brute-Force Scaling: The Cognitive Seed Enters the Academy

Announcing Synapse Public Release 2.0

Autonomous Causal Discovery, Exact Physical Reasoning, and In-Stride Adversarial Learning at Sub-5W Execution

Author: Roshan Kumar Gupta
Organization: earthoslab
Website: earthoslab.com
Contact: roshan@earthoslab.com
Date: November 4, 2026
Status: Official Public Release Announcement (Public Release 2.0)
Publication PDF: PUBLIC_RELEASE_2_ANNOUNCEMENT.pdf (Complete 20-page formal publication with embedded CSV attachments)
Appended Benchmark CSV Datasets:


Executive Summary

In Public Release 1.0 (October 31, 2026), earthoslab introduced Synapse, demonstrating that an autonomous, biomimetic cognitive organism could learn continuously from complete tabula rasa across 155 curriculum modules on a single consumer CPU core at under 5 Watts.

In Public Release 2.0, the cognitive seed enters the Academy.

Here, verified empirical benchmarks are reported across five landmark evaluation suites:

  1. The Unified Cognitive Academy: An adversarial schooling curriculum across 6 demanding scientific domains demonstrating 87.33% Cold Binary Pass Rate (89.80% Mean Rubric Score) and 90.00% Post-Grind Binary Pass Rate (92.47% Mean Rubric Score), with 21.05% struggle remediation (4/19 failed items recovered on retest, yielding a +2.67 pp net gain across 150 items) and a calibrated 0.2213 Brier score (substantially improved from 0.4223 cold baseline).
  2. 32-Core Parallel Cohort Scalability: 8 concurrent student organisms executed simultaneously across multi-core workers in 6.14 seconds wall-clock time, achieving 0.00% process variance ($87.33% \pm 0.00%$ Cold, $90.00% \pm 0.00%$ Post-Grind) and zero observed deadlocks.
  3. Exact Quantitative Physics & Equation Solving ($0.00e+00$ Residual): Analytical constraint satisfaction across 6 fundamental physical equation systems (Hydrostatic equilibrium, Jeans mass, $E=mc^2$, Stefan-Boltzmann luminosity, Schwarzschild radius, Carnot efficiency) with zero numerical approximation error (input parameters fully disclosed).
  4. Capability-Specific Component Ablation Suite: Controlled lesion experiments isolating key cognitive mechanisms, proving that removing the equation engine drops solved equations to 0/6, removing early grounding drops stability to 0.35, and removing dialectical struggle remediation causes a 10.52 percentage point drop ($\sim$50% relative reduction) in struggle recovery.
  5. 30-World Autonomous Causal Discovery Benchmark: Across 30 procedurally generated micro-worlds in 120 controlled benchmark runs under strict budget matching ($B_{\text{active}} \equiv B_{\text{random}} = 50$), Autonomous Active Inference achieved 97.21% confounder rejection with 21.39% Confounder Detection Recall (CDR) and an SHD of 42.60, compared to Passive Observation (SHD 83.00, 78.27% rejection) and Static Random Exploration (SHD 50.70, 9.44% CDR).
  6. Delayed Retention After Cross-Domain Interference: Retained 91.11% of causal structures (retained F1 = 0.1016) across all 30 micro-worlds following active distractor interference and sharp-wave ripple sleep consolidation.
  7. Deduplicated Isomorphic Transfer: Evaluated across 20 unique held-out diagnostic transfer probes across disjoint vocabularies, achieving 30.00% Binary Pass Rate (36.00% Mean Rubric Score) with zero probe duplication.
  8. Extreme Green Edge Efficiency: The entire student substrate executes on under 5 Watts on a single consumer CPU core with zero discrete GPUs required, consuming approximately 430 MB of RAM.

1. The AI Scaling Bottleneck & The Biological Alternative

Modern artificial intelligence has reached an unsustainable physical, financial, and architectural overhead:

  • Brute-Force Energetic Waste: Trillion-parameter transformer architectures require dedicated multi-megawatt data centers, hundreds of gigabytes of expensive high-bandwidth memory (HBM), and hundreds of watts per single prompt.
  • Catastrophic Forgetting: Standard neural networks cannot learn continuously from sequential experience. Fine-tuning on new distributions overwrites previously trained weights, destroying earlier capabilities.
  • Stochastic Fabrication (Hallucination): Because transformers sample statistically probable tokens rather than verifying causal physical invariants, they inherently fabricate plausible-sounding untruths without metacognitive self-awareness.
  • Cognitive Passivity: Large language models have no internal clock or continuous cognitive agency. Compute halts entirely between prompts; there is no autonomous curiosity, no invariant monitoring, and no offline consolidation.

Nature took a radically different path. A human child learns continuously from sparse, real-time experience on a biological metabolic budget of under 15 Watts—without ever wiping past memories or needing to be retrained from scratch.

Today, under Project Earthos, earthoslab announces Synapse Public Release 2.0, advancing the sovereign biomimetic architecture into university-level causal discovery, exact physical reasoning, and adversarial dialectics.


2. An Honest Distinction: From Toddler to University Scholar (Zero Clout)

Before examining empirical metrics, an objective scientific distinction is stated: earthoslab does not claim that Synapse today matches the encyclopedic conversational breadth, essay writing, or coding fluency of a 70B Large Language Model.

A modern 70B LLM is like the Library of Alexandria: an adult repository of nearly all recorded human text, trained by brute-forcing millions of books. It can generate conversational prose and write Python scripts because it has ingested tens of trillions of tokens.

In contrast, Synapse is a sovereign, biologically inspired cognitive reasoning engine. In Release 1.0, Synapse was an early cognitive infant (666 grounded words). In Release 2.0, the seedling enters the Academy. It is not an essay writer; it is a scientific reasoning substrate.

How Does This Show True Intelligence Rather Than An LLM Answering and Acting the Same?

  1. Causal Invariance vs. Word Association:
    An LLM predicts words based on statistical co-occurrence. If two terms frequently appear together in textbooks (such as "tectonic strain" and "mantle convection"), an LLM generates agreeable text linking them regardless of true directionality. Synapse maintains explicit directed causal graphs. When asked if strain causes convection, it checks directionality and explicitly rejects the reverse-causality fallacy (Strain does not cause Convection), rather than generating agreeable filler.
  2. Active Physical Intervention ($do(X)$) vs. Passive Text Processing:
    An LLM is passive—it can only read and write strings. Judea Pearl proved mathematically that passive statistical observation can never untangle true causes from common confounders. Synapse actively formulates hypotheses and performs targeted physical interventions ($do(X)$) in its environment, achieving 97.21% confounder rejection and discovering more than $2.2\times$ as many true confounders as random exploration under an identical budget ($B=50$).
  3. Exact Continuous Physical Constraints vs. Arithmetic Hallucination:
    Transformers treat numbers as text tokens, frequently approximating or hallucinating arithmetic and physical equations. Synapse couples qualitative causal topology directly with continuous analytical constraints, solving governing stellar and thermodynamic equations to 0.00e+00 analytical residual without discrete GPUs.
  4. In-Stride Learning & Metacognitive Calibration vs. Sycophancy & Frozen Weights:
    LLM weights are frozen at inference. When an LLM fails, it cannot learn from the test. When Synapse struggled on difficult adversarial probes, an external Socratic teacher provided counterfactual challenges, and the organism actively recovered 21.05% of its failed struggle items (4/19) in-stride, while calibrating its confidence (Brier score dropping from 0.4223 to 0.2213).
  5. Sleep Consolidation vs. Catastrophic Forgetting:
    Fine-tuning an LLM on new data erases prior distributions. Synapse enters autonomous offline sleep cycles (sharp-wave ripple replay), retaining 91.11% of acquired causal structures even after active distractor interference across completely disjoint domains on under 5 Watts.

3. Evolution of Evaluation Batteries: Release 1.0 vs. Release 2.0

The following table provides the direct empirical comparison between the initial demonstration (Release 1.0) and the definitive Release 2.0 benchmark battery:

Benchmark Dimension Public Release 1.0 (Oct 31) Public Release 2.0 (Nov 4) Measured Capability
Curriculum Scope 155-module sequential run 6-Domain Academy & 30 Micro-Worlds Multi-domain scientific ontogeny
Cold Examination Baseline 24.15% **8...
Read more

Synapse Architecture Technical White Paper & Empirical Benchmark (October 2026)

Choose a tag to compare

@rk-roshan-kr rk-roshan-kr released this 31 Oct 16:06
c09fff1

Beyond Brute-Force Scaling: Announcing the Synapse Cognitive Architecture

Author: Roshan Kumar Gupta
Organization: earthoslab
Website: earthoslab.com
Contact: roshan@earthoslab.com
Date: October 31, 2026
Status: Official Public Release Announcement


The AI Scaling Bottleneck

The modern artificial intelligence industry is racing down a path of brute-force scaling: constructing multi-trillion-parameter Large Language Models (LLMs) that require dedicated power plants and massive GPU server farms.

Yet despite their linguistic fluency, modern transformers suffer from fundamental structural limitations:

  • They are frozen in time: An LLM cannot learn continuously from real-world experience without catastrophic forgetting—new learning destroys earlier knowledge.
  • They hallucinate: Because they predict statistically probable words rather than reasoning over verified causal models, they frequently invent plausible-sounding falsehoods.
  • They are computationally unsustainable: Generating a single answer burns hundreds of watts across data center servers.

Nature took a radically different path. A 3-year-old child learns continuously from sparse, real-time experience on a biological metabolic budget of under 15 Watts—without ever wiping past memories or needing to be retrained from scratch.

Today, earthoslab presents Synapse, an autonomous, bio-inspired cognitive architecture developed under Project Earthos by earthoslab, designed for continuous lifelong learning, deterministic causal reasoning (zero stochastic fabrication), and microscopic computational footprint.


The Breakthrough: The 5-Hour Autonomous Distillation Run

To test the developmental architecture, an unbroken 5-hour autonomous learning benchmark was executed on a single consumer workstation.

Starting from complete tabula rasa—zero initial vocabulary, zero pre-training, and zero pre-configured world knowledge—the Synapse organism was taught continuously by an external local language model acting as an omniscient Socratic tutor across 155 sequential curriculum modules (spanning physics, chemistry, thermodynamics, genetics, computing, and geology).

Benchmark Summary

Metric Quantitative Result Empirical Significance
Duration 5.05 Hours (18,171 seconds) 100% continuous, uninterrupted autonomous loop
Curriculum Modules Covered 155 modules Covered astrophysics, computing, biochemistry, geology, etc.
Curriculum Episodes Absorbed 874 multi-modal episodes Ingested via continuous metric grounding
Autonomous Socratic Dialogues 472 inquiries Organism formulated questions for uncertain concepts
Diagnostic Comprehension 24.15% → 71.33% 107/443 pre-test → 316/443 post-consolidation across 443 probes
Net Distillation Gain +47.18% (+47.18 pp) Robust, statistically verified developmental progression
Vocabulary Acquired 0 → 666 grounded concepts Complete emergence from blank slate
Causal World Model Edges 0 → 64 verified transitions Deterministic factor graph structure
Modular Memory Persistence High (Sparse Edges) Acquired pathways persist without dense gradient overwrite
Inference Latency < 1.5 ms / 1,580 ms < 1.5 ms isolated graph traversal; 1,580 ms median full probe pipeline
Power Consumption < 5 Watts Single-core CPU execution without discrete GPUs

Evaluation Methodology & Diagnostic Protocol

To rigorously measure capability emergence without data leakage, the organism was evaluated across an independent evaluation benchmark consisting of 443 held-out diagnostic probes across all 155 curriculum modules (comprising 166 unique query templates re-evaluated across module iterations):

  • Forward Causal Prediction ($N=299$): Assessing forward physical consequence prediction ($A \to B$, e.g., "molecule causes what?"). Post-consolidation accuracy reached 71.91% (vs. 23.75% pre-test baseline; 215/299 vs. 71/299).
  • Abductive Inverse Retrodiction ($N=144$): Assessing backward causal diagnosis ($B \leftarrow A$, e.g., "reaction happens due to what?") where the inverse direction was never taught explicitly. Post-consolidation accuracy reached 70.14% (vs. 25.00% pre-test baseline; 101/144 vs. 36/144).
  • First-Encounter vs. Repeated Module Probes: On the 166 unique queries upon first encounter, accuracy rose from 4.22% (7/166) pre-test to 69.28% (115/166) post-consolidation (+65.06% net gain). The remaining 277 longitudinal repeated module probes exhibited an elevated pre-test baseline of 36.10% (100/277) rising to 72.56% (201/277), confirming earlier concept persistence alongside authentic developmental competition.
  • Held-Out Protocol Definition: Probes were strictly held-out from the tutor's declarative instruction sentences. The tutor provided declarative facts ("Molecules undergo chemical reactions"), while the independent diagnostic harness queried causal dependencies ("molecule causes what?").
  • Latency Measurement Boundaries: In the embedded dataset, latency_ms measures the complete end-to-end evaluation harness (prompt parsing, lexical matching across cortical columns, uncertainty computation, candidate ranking, and disk logging), which yields a median of 1,580 ms on a single consumer CPU core without GPUs. In contrast, core factor graph traversal ($T(s_j \mid s_i) &gt; 0.50$) executes in < 1.5 ms via direct sparse hash lookups.
  • Embedded Dataset Verification: The complete dataset containing all 443 prompts, expected ground-truth causes, latencies, and verbatim student outputs is physically embedded directly inside the PDF document as an accessible attachment (distillation_evaluation_probes.csv) and tabulated in the Appendix.
  • Interpreting Pass & Fail Results: The transition from FAIL $\to$ PASS indicates successful causal concept grounding and sleep consolidation (+47.18% net learning gain). Lingering FAIL entries in abstract domains (e.g., astrophysics) provide empirical confirmation of an authentic developmental learning curve requiring longitudinal reinforcement, rather than artificial lookup memorization.

An Honest Distinction: A Cognitive Seed vs. A Linguistic Library

Before examining empirical metrics, an objective scientific distinction is stated: earthoslab does not claim that Synapse today matches the encyclopedic conversational breadth, essay writing, or coding fluency of a 70B Large Language Model.

A modern 70B LLM is like the Library of Alexandria: an adult-level repository of nearly all recorded human text. It can write sonnets, generate Python scripts, and converse in 40 languages because it was trained by brute-forcing millions of books.

In contrast, Synapse is currently at the developmental stage of a 2.5 to 3-year-old child—an early, living cognitive core. It knows 666 words and reasons over physical causes and effects. It cannot write a novel, and it does not possess encyclopedic trivia.

So why compare them at all?
Because the evaluation compares the fundamental mechanism of intelligence and learning, not surface linguistic volume:

  • An LLM is an inanimate library: it cannot learn new facts on the fly without catastrophic forgetting, it cannot tell you when it is ignorant (hallucinating instead), and it requires the electrical power of a small town.
  • Synapse is a living seedling: it possesses the genuine machinery of cognition—lifelong continuous learning, epistemic curiosity, sleep consolidation, and deterministic causal truth (zero stochastic fabrication)—operating on less than 5 Watts.

Curriculum Distillation Efficiency: 25 Kilobytes vs. Multi-GB Fine-Tuning

Unlike end-to-end foundation pre-training from raw internet crawl data, Synapse was distilled via active Socratic interaction from an existing local teacher language model.

The empirical significance is therefore not token equivalence to web-scale pre-training, but extreme curriculum distillation efficiency: while traditional architectures require gigabytes of fine-tuning text to acquire domain adaptations, Synapse established a coherent, multi-domain topological world model from an extraordinarily compact 25 Kilobyte Socratic curriculum (874 short sentences). From this sparse 25 KB stream, the organism:

  • Grounded 666 distinct physical and scientific concepts.
  • Connected them into 64 verified cause-and-effect transitions.
  • Ingested representations across 155 curriculum modules with an average of only 5.6 sentences per module.

Verbatim Empirical Proofs: Actual Responses from the Organism

These are actual, verbatim responses generated by the 5-hour trained Synapse model, demonstrating four distinct cognitive mechanics that no traditional LLM can replicate at this footprint:

Proof 1: Deterministic Causal Extraction (< 1.5 ms Core Lookup)

When queried on a physical cause, the organism traverses its factor graph and extracts the exact proven effect in less than 1.5 milliseconds:

User Prompt: "molecule causes what?"
Synapse Response:

🕸️ Active Causal Transitions for `molecule`:
* Molecule causes Reaction.

Why this proves intelligence: It does not generate probabilistic filler or conversational waffle. It computes a direct mathematical graph proof ($T(s_j \mid s_i) &gt; 0.50$). If a path is factually proven, it outputs it with certainty; if not, it cannot hallucinate one.


Proof 2: Bayesian Abductive Retrodiction (Backward Medical/Scientific Diagnosis)

In the real world, doctors and scientists must solve inverse problems: obser...

Read more