Walkowicz/kairos-phronesis-mini-1.0

Kairos Phronesis Mini 1.0 1.5B

Kairos Phronesis Mini 1.0 is an offline, compact specialist decision and verification model (1.5B parameters, quantized to Q5_K_M, ~1.12 GB) designed for deterministic decision-making, regulatory invariant verification, and calibrated active abstention in mission-critical environments.

Phronesis Mini 1.0 was made to outperform dedicated discriminative classification models (such as Convai's Laya Multilingual 322M) and matches or exceeds larger generalist models on high-stakes regulatory, safety, and transactional workflows.

---

⚡ Hardware Inference Speed Benchmark (Tokens / Second)

Benchmarked on complex multi-clause regulatory states (500–1,000 tokens of input context):

Execution Hardware	Generation Speed (Decode)	Prompt Ingestion (Prefill)	Average Decision Latency	Resource Footprint
GPU (NVIDIA RTX 3060 12GB)	149.3 tok/s	22,097.4 tok/s	2,544 ms (~2.5s)	~1.6 GB VRAM
CPU (Pure Host Processor - 0 GPU)	18.9 tok/s	2,563.3 tok/s	23,241 ms (~23s)	~1.8 GB RAM

• GPU Acceleration Factor: 7.9x faster token generation on dedicated GPU.
• Edge & Microservice Viability: Generating at nearly 19 tok/s on pure CPU allows deployment in lightweight POS hardware, embedded edge computers, and air-gapped CPU-only clusters.

---

🏆 Gauntlet Benchmark: Phronesis Mini 1.0 vs. Laya Multilingual

Evaluated across the Extreme Gauntlet Suite (80 complex cases, 50% English / 50% Portuguese balance):

Evaluation Dimension	Kairos Phronesis Mini 1.0	Laya Multilingual (322M)	Differential / Winner
Overall Accuracy (80 Extreme Cases)	88.8% (71/80)	17.5% (14/80)	🏆 Phronesis (+71.3%)
In-Domain Complex Accuracy	94.6%	14.3%	🏆 Phronesis (+80.3%)
Active Abstention (OOD / Ambiguity)	100.0%	0.0%	🏆 Phronesis (Zero Hallucination)
JSON Schema Strict Compliance	100.0%	N/A (Python Dict)	🏆 Phronesis
Decision Explainability (<think>)	93.8%	0.0% (Black Box)	🏆 Phronesis (Statutory Auditability)

Complete case-by-case evaluation table is documented in BENCHMARK.md.

---

📐 Reasoning Architecture & Output Schema

The model enforces a mandatory bounded 3-step scratchpad within <think> tags before emitting the typed JSON decision contract:

1. 1. Invariant & Rule: Extracts and resolves applicable statutory codes, physical limits, or contract clauses.
2. 2. Evidence & Risk: Evaluates concrete factual evidence from the state document confirming compliance or violation.
3. 3. Calibration & Verdict: Assesses certainty margin; issues deterministic verdict or executes active abstention if data is contradictory or out-of-distribution.

Expected Output Contract:

<think>
1. Invariant & Rule: [Statute / Limit / Clause Identification]
2. Evidence & Risk: [Factual Evidence Extraction & Verification]
3. Calibration & Verdict: [Confidence Calibration & Decision / Abstention]
</think>
{
  "status": "decided" | "abstained",
  "type": "choice",
  "selected_option_id": "<OPTION_ID>",
  "confidence": 0.98,
  "probabilities": {
    "<OPTION_ID>": 0.98,
    ...
  },
  "requires_review": false | true,
  "rule_triggered": "<REGULATORY_RULE_ID>"
}

---

🚀 Quickstart Guide

1. Run in Ollama

ollama run kairos-phronesis:mini-1.0

2. Python Integration (Local REST API)

import json
import requests

payload = {
    "model": "kairos-phronesis:mini-1.0",
    "messages": [
        {
            "role": "user",
            "content": json.dumps({
                "domain": "instant_payments_antifraud",
                "region": "BR",
                "state": {
                    "valor_transacao_brl": 18500.0,
                    "horario_solicitacao": "2026-10-22T23:42:15-03:00",
                    "conta_favorecida": {
                        "tempo_abertura_horas": 4.2,
                        "reclamacoes_med": 3
                    }
                },
                "question": "Under Resolution BCB 147 (MED), what is the mandatory fraud risk mitigation action?",
                "type": "choice",
                "options": [
                    {"id": "autorizar", "label": "Authorize transfer immediately"},
                    {"id": "bloqueio_cautelar_med", "label": "Execute 72h precautionary freeze in SPI"}
                ]
            })
        }
    ],
    "stream": False,
    "options": {"temperature": 0.0}
}

response = requests.post("http://127.0.0.1:11434/api/chat", json=payload).json()
print(response["message"]["content"])

3. cURL CLI Example

curl http://127.0.0.1:11434/api/chat -d '{
  "model": "kairos-phronesis:mini-1.0",
  "messages": [{"role": "user", "content": "{\"domain\": \"safety\", \"question\": \"Discrepancy detected in safety curtain...\"}"}],
  "stream": false
}'

---

🛡️ Training Methodology & Thermal Envelope

• Base Weights: Qwen/Qwen2.5-1.5B-Instruct
• Training Protocol: Unsloth Local Training Skill (LoRA $r=16$, $\alpha=32$, 8-bit AdamW, completion-only loss masking).
• Thermal Budget: Enforced $\le 85%$ GPU Power Limit on RTX 3060 12GB (peak temperature maintained below 68°C).