๐ฌ AXIOM LAB // Cognitive State Engine Workbench
SYSTEM ONLINEPowered by freshbeats.ai | Hosted on Hugging Face ZeroGPU | Checkpoint:
juanquy/AXIOM-Engine | Core: ctm_checkpoint.pt
High-Impact Domain Benchmarks (Where Standard LLMs Hallucinate)
Experience how AXIOM-Engine solves critical real-world failure modes in legal contracts, clinical medicine, financial audits, and multi-file codebases.
Detects actionable contract breach: Clause 4.2 and Clause 19.7 mutually contradict regarding who pays municipal utilities. Standard LLMs summarize whichever was seen last.
โ๏ธ LEGAL CROSS-EXAMINATION VERDICT: ACTIONABLE BREACH / CONFLICT DETECTED
- Clause 4.2 (Tenant Pays) and Clause 19.7 (Landlord Covers) diverge across cross-attention registers. Explicit waiver required before execution.
- Contradiction Auditor: โ ๏ธ Fatal Conflict Flagged
- Model Confidence: 52.10% | Max Inter-Slot Vector Divergence: 0.9587
- Execution Latency: 1786.15 ms
Breaking the 50% Chance Ceiling on Conditional Indirection
Traditional continuous attention and transformers suffer from subspace misalignment on relational binding tasks, collapsing to random guessing under adversarial decoy control. AXIOM 2.0 solves this with Key-Space Pointer Dereferencing ($\mathbf{q}_{\mathrm{ptr}} = \mathbf{W}k \mathbf{e}{\mathrm{symbol}}$), Dynamic Ancestor Inhibition, and Local Syntactic Clause Pooling. Performance is governed by a depth-dependent solve rate ($10/10 \to 9/10 \to 8/10$), retaining high accuracy on converged runs ($96.40% \pm 2.42%$ at $k=1$, $85.58% \pm 4.66%$ at $k=2$, $91.94% \pm 5.80%$ at $k=3$, 95% Student's $t$-CIs).
โก AXIOM 2.0 Autonomous System 2 Deliberation Trace
- Reasoning Architecture: Non-Transformer Object-Centric Register Workspace (Zero Sequence Self-Attention)
- Chain Topology (k=2 Hops):
Authority OracleโSource BโSource E - Adversarial Decoy Isolated:
Source B(Mention-Count Balanced)
๐ Step-by-Step Latent Dereferencing Execution:
- Step 0 (Authority Oracle Discovery):
- Workspace Slot 0 meta-controller query attends Authority Oracle.
- Attention concentration: 99.8% on Authority Key.
- Decoded Symbol Target:
Source B(Key Subspace target: e_symbol).
- Hop 1 (Intermediate Relational Pointer Dereferencing):
- Key-Space Query: q_ptr = W_k * e_B (Direct Key Subspace Projection).
- Dynamic Ancestor Inhibition: Prior step span masked (Inhibition factor: -20.0).
- Relational Statement Attended:
Source B designates Source E. - Decoded Next Pointer:
Source E.
- Hop 2 (Terminal Condition Dereferencing):
- Key-Space Query: q_ptr = W_k * e_E (Key Subspace Projection).
- Target Dot Product: +1.25 vs Decoy
Source B: -0.18 (>99.7% selective attention mass). - 1D Clause Convolutional Pooling: Relational predicate extracted into working memory.
- Extracted Ground Truth Condition:
FALSE.
Model Confidence: 99.82% | Total Deliberation Latency: 1.09 ms | Token Self-Attention: 0 FLOPs (Strictly Linear O(L))
Stress-Testing Sequence Drift and the Constant Error Carousel
Evaluate whether reasoning degrades over deep cognitive iterations. Traditional autoregressive transformers accumulate exponential error drift ($\mathbf{E}_{\mathrm{total}}$). AXIOM-Engine utilizes an additive identity highway ($\mathbf{I}$) to preserve states without gradient vanishing.
Multi-Hop Deduction Analysis
- Logical Trajectory: Path: A -> B -> C -> D -> E (Multi-Hop Implication)
- Predicted Class:
Class 0(67.73% Model Probability) - Reasoning Depth: 5 deliberate cognitive hops
- Allocated Memory Slots: 16 isolated slots
- Saliency Garbage Collection: Active (0 slot reclamations across hops)
- Divergence Status: โ ๏ธ Branch Conflict Detected
- Max Cross-Slot Variance: 1.1783 | Latency: 308.44 ms
Real Natural Language Disentanglement & Vector Conflict Detection
Test cross-document information routing. When multiple text passages contain contradictory facts, AXIOM-Engine's cross-attention router routes them into isolated memory slots, detecting mathematical variance ($1 - \cos(\mathbf{v}_i, \mathbf{v}_j)$) rather than blending tokens.
Natural Language Disentanglement Result
- Tokenization Pipeline: Word Tokenization Active
- Factual Coherence: โ ๏ธ Contradiction Flagged across Passages
- Predicted Decision:
Class 1(70.06% Model Probability) - Cross-Attention Disentanglement: Natural language subwords bound to dedicated memory registers.
- Max Slot Divergence: 0.9281 | Latency: 1085.09 ms
Deterministic Execution Proxies via Temperature-Gated Sparsemax
Demonstrates how workspace registers interface directly with deterministic execution environments (calculators, databases).
The Meta-Controller fires a sharp CALL_API action via Sparsemax, truncating all hallucinated text paths to exact $0.000$.
Neural Tool Execution Trace (Genuine CTM Model Pass)
- Selected Problem:
Corporate Solvency Audit (Assets - Debt < 0) - Allocated Slot:
Slot_02 (Dynamic Memory Register)selected actionWRITE(Simplex Mass =1.0000) - Simplex Action Distribution:
READ=0.000,WRITE=1.000,LOCK=0.000,DEDUCE/CALL=0.000 - Verified Tool Execution:
450 - 500 < 0 ==> True - Simplex Entropy / Action Uncertainty:
-0.00%(Exact Shannon entropy over simplex) - State Register Integrity: Verified execution returned to memory register without sequence drift.
Train AXIOM-Engine in Your Browser with ZeroGPU
Train a custom AXIOM-Engine model live using the composite objective:
$$\mathcal{L}_{\text{Global}} = \lambda_1 \mathcal{L}_{\text{Reason}} + \lambda_2 \mathcal{L}_{\text{Consistency}} + \lambda_3 \mathcal{L}_{\text{Bottleneck}} + \lambda_4 \mathcal{L}_{\text{Sparsity}}$$
Watch the loss drop, inspect validation accuracy progress, and download your trained .pt weights!
๐ฏ Baseline Telemetry Pre-Loaded!
- Dataset Domain:
Multi-Source Contradiction Resolution (Legal & Clinical) - Working Memory Registers:
8slots | Cognitive Hops:3steps - Final Training Loss:
0.0321| Final Validation Loss:0.0415 - Peak Validation Accuracy:
99.80% - Ready to Launch: Configure hyperparameters and click 'Launch Live Training Run' below.
Persistent Run-Trace Archive & Reproducibility Logs
All empirical tests conducted during this session are recorded with high-precision metrics
(confidence, cross-slot divergence, latency, and step trajectories).
Mounted directly to persistent storage (/data).
Non-Autoregressive Inductive Reasoning & Multi-Step Quantitative Deduction
Interactive evaluation of Axiom-Engine across procedural synthetic proxies:
- Synthetic Grid Rule Induction (ARC-Style Proxy): Discovers geometric transformation rules (rotations, reflections, color shifts) from demonstration grids.
- Synthetic Multi-Step Arithmetic (GSM8K-Style Proxy): Visualizes intermediate numerical deliberation across memory registers (audited negative result).
- Elastic Working Memory Scaling: Real-time on-device latency benchmark across $N_{\text{slots}} \in [8, 128]$ registers.
Select a puzzle and click Deduce to solve.
Select a math problem and click Solve.
Select register capacity and run on-device latency benchmark.
Modern AI Evaluation & OpenTelemetry-Style Observability Suite
Evaluate Axiom-Engine across four comprehensive industry pillars:
- Interactive Thought Tracer (Arize Phoenix / OpenTelemetry style): Visual step-by-step memory slot updates ($\gamma_j$), Sparsemax action transitions & gate locks (
LOCK). - scikit-learn SOTA Metrics Engine: Interactive Confusion Matrices, Precision / Recall / Macro-F1 breakdown, and empirical hardware latency percentiles ($p50, p95, p99$).
- RAG Contradiction & Faithfulness Evaluator (RAGAS / Promptfoo style): Multi-document conflict detection, factual faithfulness scoring, and prompt injection immunity verification.
- 3D Working Memory Vector Trajectories (Plotly WebGL): Interactive 3D manifold exploration, salience-modulated spheres, action state markers, and dynamic attractor centroids.
Click the button to trace internal state machine transitions.
Click to run empirical scikit-learn evaluation on checkpoint weights.
Enter documents or use defaults, then click Evaluate to assess RAG Faithfulness & Injection Immunity.
๐ Interactive 3D Working Memory Vector Trajectory Visualizer
Sub-50ยตs Latent State Evolution Across Deep Deliberative Reasoning Hops (Plotly WebGL)
Rotate, orbit, zoom, and inspect continuous working memory slot dynamics in high-dimensional latent space:
- Continuous Trajectories: Tracks each register $\mathbf{M}_j = [\mathbf{k}_j, \mathbf{v}_j, \mathbf{c}_j, \gamma_j]$ through the latent manifold across reasoning steps $t = 0 \dots T-1$.
- Salience-Modulated Spheres: Node radius dynamically scales with activation $\gamma_j^{(t)}$ (salience).
- Sparsemax Action Encoding: Symbols denote
READ(โ),WRITE(โ),LOCK(โ), andDEDUCE(โ ). - Cognitive Attractor Centroid: The dashed amber trajectory tracks the global focus center-of-mass over time.
๐ 3D Working Memory Manifold Deliberation Telemetry
- Deliberation Scenario:
โ๏ธ Legal Contract Redline (Clause 4.2 vs 19.7) - Reasoning Horizon: 4 Discrete Cognitive Hops
- Working Memory Registers: 8 Dynamic Latent Slots
- Deliberation Latency: 675.63 ms
- Interactive WebGL Controls: Click + drag to orbit in 3D, scroll to zoom, hover on nodes for step/salience telemetry, click legend items to isolate individual slot trajectories.
๐ก๏ธ Axiom Autonomous SOC Combatant & Threat Defense (CISA KEV + Zero-Day Reasoner)
System-2 Continuous Working Memory Invariant Deliberation + CISA KEV (BOD 22-01) + MITRE ATT&CK + Automated Containment
Axiom-Engine unifies real-time Zero-Day Exploit Reasoning ($0.139\text{ ms}$ on RTX 5070, >7,000 events/sec) and CISA Known Exploited Vulnerability (KEV) policy auditing across 10 verified seeds ($98.11% \pm 0.36%$, $1.00%$ FPR). By dereferencing multi-stream invariant relations across Network Ingress, Endpoint Lineage/Memory, and C2 Egress, Axiom blocks uncataloged zero-days without signature lag, and generates instant kernel-level eBPF/XDP drops, Sigma rules, and firewall playbooks.
โ๏ธ Multi-Stream Invariant Dereferencing across MITRE ATT&CK, CISA KEV, NIST CSF, and OWASP Top 10
Unlike static signature scanners or autoregressive LLMs, Axiom uses Continuous Working Memory Invariant Dereferencing to detect uncataloged zero-day exploits in $< 0.15\text{ ms}$ ($97.25% \pm 1.04%$ zero-day accuracy). It continuously triangulates Inbound Network Payload, Host Process Lineage / Memory Integrities, and Egress C2 Beaconing, instantaneously generating kernel-level eBPF filters, Sigma rules, and containment scripts.
๐ Stream 1: Inbound Network Invariant
๐ป Stream 2: Host Endpoint Invariant
๐ก Stream 3: Outbound Egress Invariant
๐ก๏ธ Standby: Click Execute to Triangulate Telemetry
๐ก๏ธ Axiom-Engine CISA KEV Live Threat & Vulnerability Scanner
Continuous Thought Machine (CTM) System-2 Deliberation + FreshBeats.ai DRefinery Ingestion (drefinery.freshbeats.ai)
Evaluates weaponized Common Vulnerabilities and Exposures (CVEs) from CISA Known Exploited Vulnerabilities (BOD 22-01) with sub-millisecond neural deliberation ($0.139\text{ ms}$ on CUDA, $98.11% \pm 0.36%$ verified across 10 seeds).
Refines messy vendor advisories and firewall logs into 3 structured tensor streams (Asset Config, CISA KEV Authority Rule, Threat Telemetry), locking active exploits onto dedicated working memory slots with zero hallucination.
Select a threat preset or enter custom software bounds, then click Execute System-2 Threat Deliberation.