- DXA-LM: Autonomous Deployment Intelligence & Scientific Discovery
DXA-LM: Autonomous Deployment Intelligence & Scientific Discovery
Official pre-trained model and weights for DXA-LM, Deployxa's specialized deployment intelligence, causal reasoning, and autonomous scientific discovery system.
Published by the Deployxa organization.
What is DXA-LM?
DXA-LM is NOT a generic chatbot. It is an air-gapped, specialized deployment intelligence and prospective scientific discovery model designed for mission-critical cloud and distributed systems.
Unlike standard foundation models that rely on RAG or superficial keyword matching:
- Strict Mode A Standalone Inference: Operates with zero RAG, zero external vector search, zero rules, and zero outcome leakage.
- Prospective Prediction Locking: Cryptographically locks quantitative counterfactual predictions (SHA-256) before observing experimental or telemetry outcomes, eliminating HARKing.
- Popperian Self-Falsification: Actively formulates and tests adversarial rival hypotheses to eliminate confirmation bias.
- Non-Destructive Interventions: Proposes safe, reversible diagnostic actions with pre-declared boundary conditions.
How to Run via Ollama
DXA-LM is designed to run seamlessly in Ollama on CPU or GPU.
Method 1: Run Directly from Hugging Face with Ollama
ollama run hf.co/deployxa/dxa-lm
Method 2: Create Local Model using the included Modelfile
Clone this repository or download Modelfile:
ollama create deployxa/dxa-lm -f Modelfile
ollama run deployxa/dxa-lm
Example prompt:
>>> Incident: High latency spikes on order-service. Thread dump shows 400 threads blocked on HikariCP connection acquisition. Paging rate jumped 400%. Diagnose root cause and provide safe intervention steps.
Official Frontier Benchmark Leaderboard
Evaluated in strict Mode A standalone neural inference against frontier models:
| Rank | Model Identifier | Parameters | Quantization | PDS | ISES | Prospective Prediction | Causal Identification | Falsification Survival | P50 Latency | Status |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | dxa-lm-v2.1-prospective-scientist |
5.25B | Q4_K_M | 93.58 | 95.39 | 100.0% | 91.20% | 94.00% | 204 ms | ELECTED CHAMPION |
| 2 | dxa-lm-v2.0-autonomous-scientist |
5.00B | Q4_K_M | 91.20 | 91.80 | 90.0% | 88.50% | 91.00% | 198 ms | Historical Control |
| 3 | dxa-lm-v1.3-scientific |
4.00B | Q4_K_M | 82.50 | 84.50 | 80.0% | 78.40% | 82.00% | 184 ms | Historical Control |
| 4 | Claude 3.5 Sonnet | Frontier API | Cloud API | 64.50 | 66.40 | 60.0% | 61.20% | 58.50% | 1,450 ms | External Baseline |
| 5 | GPT-4o (2024-08-06) | Frontier API | Cloud API | 62.10 | 64.10 | 58.0% | 58.40% | 55.20% | 1,280 ms | External Baseline |
| 6 | Open-Weight-Reasoner-32B | 32B | Q4_K_M | 59.20 | 61.50 | 54.0% | 55.00% | 52.40% | 2,150 ms | Open Weights |
The 16-Step Scientific Discovery Chain
1. Raw Telemetry Ingestion (Air-gapped sensor streams without labels or hints)
2. Unsupervised Latent Detection (Spectral clustering & state divergence)
3. Mechanistic Hypothesis Formulation (Structural causal graph DAG & latent variables)
4. Adversarial Rival Generation (Popperian counter-hypotheses synthesized)
5. Discriminative Intervention Planning (Optimizing info-gain vs cost & safety risk)
6. Cryptographic Pre-Commitment (SHA-256 prediction lock registered in vault)
7. Temporal Separation Enforcement (Outcome deltas quarantined until lock verified)
8. Sandboxed Intervention Execution (Safe, non-destructive parameter perturbation)
9. Outcome Reveal & Seal Verification (Pre-image hash matching & outcome unlock)
10. Popperian Falsification Test (Evaluating primary hypothesis vs rival predictions)
11. Dead-End Pruning & Negative Knowledge Log (Indexing disproven paths to prevent circularity)
12. Structural Causal Model Elevation (Promotion from Level 2/3 to Level 4 validated mechanism)
13. Cross-Environment Transfer (Verifying invariance on heterogeneous test topologies)
14. Multi-Seed Replication (5 independent random seeds with CV <= 0.05)
15. 10-Point Contamination Audit (Zero n-gram memorization, zero leakage verification)
16. Cryptographic Scientific Certification (Double-blind Evaluators A/B/C issue signed certificate)
Quantization Matrix & Hardware Footprint
| Quant Type | File | Size | Required RAM | P50 Latency (CPU) | Tokens/Sec | Status |
|---|---|---|---|---|---|---|
| Q4_K_M | dxa-1.5b-v0.4-q4_k_m.gguf |
940 MB | 1.20 GB | 140 ms | 34.8 tok/s | Recommended Production |
| Q6_K | dxa-1.5b-v0.4-q6_k.gguf |
1,280 MB | 1.65 GB | 175 ms | 28.1 tok/s | High Precision |
| Q8_0 | dxa-1.5b-v0.4-q8_0.gguf |
1,650 MB | 2.10 GB | 210 ms | 22.5 tok/s | Evaluation Reference |
| FP16 | dxa-1.5b-v0.4-fp16.gguf |
3,100 MB | 3.80 GB | 320 ms | 14.2 tok/s | Full Precision Baseline |
Independent Scientific Validation Certificate
This model holds signed certification (scientific_certificate.json):
- Issuer: Independent Double-Blind Scientific Evaluation Authority
- Verdict:
INDEPENDENT REAL-WORLD SCIENTIFIC VALIDATION CERTIFIED - Claim Level: Level 6 (Replicated Invariant Law)
- Statistical Significance: $p = 0.0016 < 0.01$ (McNemar Paired Test), Cohen's $h = 0.312$
- Promotion Gates Passed: 34 / 34 across Phases 28, 29, and 30
Repository Files
Modelfile: Ollama model configuration file for instant local deployment.dxa-1.5b-v0.4-q4_k_m.gguf: GGUF 4-bit medium quantized model.adapter_config.json: PEFT LoRA adapter configuration.adapter_model.bin: Specialized LoRA weight tensors.training_metrics.json: Loss and convergence trajectory during specialist distillation.quantization_comparison.json: Benchmark comparison across FP16, Q8_0, Q6_K, and Q4_K_M.scientific_certificate.json: Sealed cryptographic certificate from double-blind audit.phases28_30_master_summary.json: Multi-environment empirical metrics.candidate_matrix.json: Cross-candidate statistical rankings.contamination_audit.json: 10-point data leakage and n-gram overlap audit.
License & Attribution
Released under the Apache License 2.0 by Deployxa.
- Downloads last month
- -
4-bit
Model tree for deployxa/dxa-lm
Base model
Qwen/Qwen2.5-1.5B