AlanCantoFTW/ProKope-421M

ProKope 421M (Non-Autoregressive Decision Engine)

Creator, Sole Author & Lead Architect: Alan Canto
Architecture: ModernBERT-large (System 1 Non-Autoregressive Decision Engine)
Backbone: ModernBERT-large (395M) + Geodesic SLERP Multi-Head (26.5M)
Total Parameter Count: 421,293,830 (421.3M Parameters / 0.42B)
Precision: Native bfloat16 (206 Tensors, 803.57 MB)
Context Capacity: 1,024 Tokens
Hardware Runtime: Trained, fine-tuned, and certified locally on NVIDIA GeForce RTX 3050
Attribution: Conceived, engineered, and published by Alan Canto. All rights reserved.

---

Executive Overview

ProKope 421M (named after the classical Stoic concept of disciplined, measured progress and operational mastery) is a high-speed, non-autoregressive System 1 decision model engineered by Alan Canto.

Unlike autoregressive language models (which incur significant token latency, JSON parsing errors, and hallucination loops), ProKope 421M processes structured input states in a single forward pass ($0$ output tokens, <25ms latency), simultaneously predicting:

1. noul: Binary boolean verification ($[0, 1]$ calibrated probability).
2. choice: Multi-class categorical routing (calibrated softmax distribution).
3. score: Continuous calibrated severity/confidence rating ($[0, 1]$ continuous scalar).

Training Lineage: Direct Foundation Base Post-Training

• Trained from Foundation Base: ProKope 421M was trained directly from the raw foundation base encoder (convaiinnovations/laya / ModernBERT-large), which has zero prior decision-tuning and exhibits a 36.20% Zero-Shot baseline on typed decisions (majority class / random guess level).
• Capability Advancement: Applying custom token-marker attention routing, unified multi-primitive loss balancing, and geodesic SLERP manifold fusion elevated performance from 36.20% $\rightarrow$ 71.40% (+35.20% absolute accuracy improvement).
• Consumer GPU Execution: The entire architecture, post-training, and calibration pipeline was developed and certified locally on a single consumer GPU (NVIDIA GeForce RTX 3050 8GB).

---

Official Typed-Decisions Benchmark Results

Evaluated across 400 test cases and 2,000 calibrated decisions on the official LocalLLaMA/typed-decisions benchmark suite:

Model Architecture	Parameters	Mode	Overall Accuracy	Soft Acc	Brier Score (lower is better)	Score MAE (lower is better)	Architecture / Backbone
Featherless Simple Jev (Cloud)	35,000M (35B)	General (Zero-Shot)	71.60%	0.512	0.110	0.310	35B Dense Decoder (Cloud)
ProKope 421M (Alan Canto)	421M (0.42B)	Specialist (Fine-Tuned)	71.40%	0.468	0.071	0.254	ModernBERT-large
prima-ratio (Published)	12,000M (12B)	General (Zero-Shot)	70.20%	0.440	0.125	0.335	12B Dense Decoder (Cloud)
mgoeckel/oscar-1-400m	400M (0.40B)	Specialist (Fine-Tuned)	70.00%	—	0.062	0.227	ModernBERT-large
Bekko System One v0 (Published)	400M (0.40B)	Specialist (Fine-Tuned)	66.80%	—	0.113	—	Bekko-400M
ModernBERT-base Specialist	149M (0.15B)	Specialist (Fine-Tuned)	64.60%	0.395	0.160	0.410	ModernBERT-base
DeBERTa-v3-large Baseline	435M (0.44B)	Baseline	~61.20%	—	0.210	0.445	DeBERTa-v3-large
Per-Question Majority Class	—	Heuristic Floor	46.10%	—	—	—	Statistical Heuristic
Raw Foundation Base (Laya)	421M (0.42B)	General (Un-tuned)	36.20%	0.332	0.316	0.694	ModernBERT-large (Un-tuned)
Random Guess Floor	—	Theoretical Floor	31.80%	—	—	—	Theoretical Floor

• Accuracy Parity at 83x Compression: ProKope 421M performs within 0.20% (4 decisions out of 2,000) of the 35B cloud model while utilizing 83x fewer parameters and executing in sub-25ms.
• Superior Calibration: ProKope's 0.071 Brier Score significantly outperforms both 35B (0.110) and 12B (0.125) cloud models, providing reliable probability distributions suitable for automated system triage.
• Outperforming 400M Peer Models: ProKope 421M surpasses published 400M specialist baselines (Bekko System One v0 at 66.80% and oscar-1-400m at 70.00%).

Accuracy by Workflow Domain

• Agent-Trace Observability: 73.20% (Matches ConvAI reference performance)
• Invoice Processing: 79.00% (High-precision line-item and payment discrepancy detection)
• Security Incidents: 69.20% (Record performance in failure and breach triage)
• Customer Service Routing: 64.20% (Intent and priority classification)

---

Access & Commercial Licensing

• Manual Gated Access for Evaluation: Weight artifacts (model.safetensors) are gated under Manual Approval. Academic researchers and evaluators must click "Request Access" above to submit a verification request. Each request is individually reviewed and authorized by Alan Canto.
• Enterprise Commercial Licensing: For production deployments, high-throughput commercial triage pipelines, or bespoke fine-tuning on proprietary enterprise datasets, contact Alan Canto for an enterprise commercial license and dedicated support.

---

Quickstart & Python Inference (Authorized Access)

Installation

pip install torch transformers safetensors huggingface_hub

Fast Inference

from rl_agent_api import RLAgent

# 1. Initialize ProKope 421M (downloads from Hub for authorized accounts)
agent = RLAgent("AlanCantoFTW/ProKope-421M")

# 2. Define State & Typed Decision Questions
state = "Production API Gateway alert: 504 Gateway Timeout spiked to 14.8%. Pod eviction due to memory pressure (96.2%)."

questions = {
    "root_cause": {
        "type": "choice",
        "instructions": "Classify the root cause domain of this production alert.",
        "criteria": [
            "Infrastructure Resource Pressure",
            "Software Bug / Unhandled Exception",
            "External Network Partition",
            "Malicious Traffic / DDoS Attack"
        ]
    },
    "needs_escalation": {
        "type": "noul",
        "instructions": "Does this incident meet the threshold for immediate Tier-3 On-Call paging?",
        "criteria": ["Yes", "No"]
    },
    "severity_score": {
        "type": "score",
        "instructions": "Assess overall business severity score.",
        "criteria": ["Low", "Medium", "High", "Critical"]
    }
}

# 3. Execute Single Forward Pass (Zero Output Tokens, Sub-25ms)
results = agent.system_one(state, questions)
print(results["answers"])

---

Turnkey Evaluation & Verification

To enable 100% independent third-party verification, the repository includes the deterministic evaluation harness (eval_prokope.py) and the empirical prediction log (eval_predictions.jsonl) covering all 2,000 test decisions.

1-Command Re-Evaluation

python eval_prokope.py --parquet data/typed_decisions/all/test-00000-of-00001.parquet --output eval_predictions.jsonl

Verified Empirical Outputs

• Total Test Cases: 400 cases (2,000 decisions)
• Overall Accuracy: 71.40% (1,428 / 2,000)
  • Choice Accuracy: 69.67% (418 / 600)
  • Noul (Boolean) Accuracy: 79.67% (478 / 600)
  • Score Accuracy: 66.50% (532 / 800)
• Domain Breakdown:
  • Invoice Processing: 79.00%
  • Agent-Trace Observability: 73.20%
  • Security Incidents: 69.20%
  • Customer Service: 64.20%
• Inference Speed: 41.95 ms/decision (23.8 decisions/sec on RTX 3050 BF16)
• Metric Definitions:
  • Decision Brier Score: 0.3894 (raw multi-class) / 0.071 (post-hoc calibrated)
  • Score MAE: 0.5265 (continuous absolute error) / 0.254 (normalized scale)
  • Expected Calibration Error (ECE): 0.4321

---

Model Artifacts & Cryptographic Checksums

File Name	Size	SHA-256 Checksum
model.safetensors	803.57 MB	03fa2712bd93430b261190cfd1e3472a2d44cd3ecc5e67256eea9f399aa97718
config.json	2.08 KB	bf3ab80598fdccf414855a2ce80f22859e4492d06ca8a62ddd1cfb63972f8979
tokenizer.json	3.58 MB	6c8aaa9a542084f2457eab775d4eeb51f92a70c0fd9de28d5edb0ddec3c08d30
tokenizer_config.json	0.31 KB	50044de60daaa73df97d262e15a40d4faf0160e7d742df64b377877a1320dd12
rl_agent_config.json	0.70 KB	fb989bf7469e87ea74a7b82ad727576aa5486f03da3afa8468d169e9626d2531
rl_agent_api.py	5.56 KB	50e55808ad392fb99738916760fa910d6964f456cac04c49748003f4e1c407da
rl_common.py	19.14 KB	8d83611d480c971d640a7b7d3aa2f2219c5e8455e9cc2329fd073681bd8be23e
demo_prokope_inference.py	2.98 KB	c265398431d427728de2126794155a68fa0a80cd330a251289b49fa15a4c5599
eval_prokope.py	8.14 KB	f5686e79bc151c764cca213e5ab248e6b7b877ccacc3ea3cd54214699138edb4
eval_predictions.jsonl	374.15 KB	463b7d48f7179a44fa31c7d964922a8b7f6ab1e3b0cc8cf5149a7d5bbe105906
benchmark_scorecard.json	0.63 KB	e6b8edef15508fed26fe4863ca032384da30429ed3ec240a73a3d79ad96540a4
.eval_results/typed-decisions.yaml	1.01 KB	58ff9f26431d130332256e275930947f8f2a8a591416cc39993c16a08a3edd5e