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.