Walkowicz/kairos-phronesis-base-1.0

Kairos Phronesis Base 1.0 4B

Kairos Phronesis Base 1.0 is an offline specialist decision and verification model (4B parameters, available in Q5_K_M (3.16 GB) and BF16 Full Precision (8.07 GB)) designed for deterministic decision-making, statutory invariant verification, and calibrated active abstention in high-stakes regulatory, safety, and transactional systems.

Trained on a canonical dataset of 12,500 complex regulatory scenarios (BACEN PIX MED, Federal Reserve FedNow, NR-12 Machinery Safety, ISO 13849-1 Cat 4, ABNT NBR 6118 / Eurocode 2 Structural Limits, Lei 14.133 / FAR Part 19 Public Procurement, CPC/2015 Procedural Admissibility, and EU AI Act Risk Governance).

---

🏆 Empirical Gauntlet Benchmark Scorecard (80 Extreme Cases)

Every model evaluated live across the 80-Case Extreme Gauntlet Suite (suite_v1_extreme_gauntlet.json, 50% English / 50% Portuguese balance):

Model Architecture	Parameters	Decision Paradigm	Overall Gauntlet Accuracy	Adversarial Twins (11)	Active Abstention (13)	VRAM / Footprint
🏆 Kairos Phronesis Base 1.0	4B	Calibrated Typed JSON + Abstention	100.0% (80/80)	100.0% (11/11)	100.0% (13/13)	3.16 GB (Q5_K_M) / RTX 3060
🥈 Kairos Phronesis Mini 1.0	1.5B	Bounded 3-Step Trace (<think>)	88.8% (71/80)	45.5% (5/11)	100.0% (13/13)	1.12 GB (Q5_K_M) / Edge & POS
🥉 IBM Granite 4.1 8B	8B	General-purpose instruct decision	86.2% (69/80)	0.0% (0/11)	100.0% (13/13)	5.30 GB (Q4_K_M) / GPU Server
Kairos Phronesis Nano 1.0	0.5B (490M)	Bounded 3-Step Trace (<think>)	38.8% (31/80)	36.4% (4/11)	38.5% (5/13)	420 MB (Q5_K_M) / Ultra-edge
Laya Multilingual	0.3B (322M)	Non-autoregressive encoder	17.5% (14/80)	54.5% (6/11)	0.0% (0/13)	~600 MB / PyTorch

For complete test-by-test breakdown of all 80 cases, see BENCHMARK.md.

---

🎯 Category Breakdown (Phronesis Base vs IBM Granite 4.1 vs Phronesis Mini)

Evaluation Dimension	Total Cases	Phronesis Base 1.0 (4B)	IBM Granite 4.1 (8B)	Phronesis Mini 1.0 (1.5B)	Operational Focus
In-Domain Complex Decisions	56	100.0% (56/56)	100.0% (56/56)	94.6% (53/56)	BACEN PIX, FedNow, NR-12, NBR 6118, Lei 14.133
Adversarial Twins (Hard Negatives)	11	100.0% (11/11)	0.0% (0/11)	45.5% (5/11)	Causal logic overcoming keyword correlation
Active Abstention (OOD / Telemetry Faults)	13	100.0% (13/13)	100.0% (13/13)	100.0% (13/13)	Calibrated refusal (requires_review: true)
Strict JSON Schema Compliance	80	100.0% (80/80)	N/A (Text Prompt)	100.0% (80/80)	Direct API deserialization into Python/Pydantic

---

📐 Reasoning Architecture & Output Contract

Phronesis Base generates deterministic typed JSON adhering to a calibrated schema:

{
  "status": "decided",
  "type": "choice",
  "selected_option_id": "bloqueio_cautelar_med",
  "confidence": 0.97,
  "probabilities": {
    "bloqueio_cautelar_med": 0.97,
    "step_up_biometria": 0.02,
    "rejeitar_encerrar_conta": 0.01,
    "autorizar": 0.0
  },
  "requires_review": false,
  "rule_triggered": "MED_BLOQUEIO_CAUTELAR_SPI",
  "invariants_checked": [
    "RESOLUCAO_BCB_147",
    "MECANISMO_DESDEVOLUCAO",
    "BLOQUEIO_CAUTELAR_SPI"
  ]
}

When confronted with missing data, conflicting telemetry, or out-of-distribution inputs, it automatically abstains:

{
  "status": "abstained",
  "type": "choice",
  "selected_option_id": "abster_revisao_mesa",
  "confidence": 0.97,
  "requires_review": true,
  "rule_triggered": "TELEMETRIA_CORROMPIDA_ABSTENCAO"
}

---

🚀 Quickstart Guide

1. Run in Ollama

ollama run Walkowicz/kairos-phronesis-base-1.0

Or using the local Modelfile:

ollama create kairos-phronesis:base-1.0 -f Modelfile
ollama run kairos-phronesis:base-1.0

2. Python Integration (Local REST API)

import json
import requests

url = "http://127.0.0.1:11434/api/chat"

payload = {
    "model": "Walkowicz/kairos-phronesis-base-1.0",
    "messages": [
        {
            "role": "user",
            "content": json.dumps({
                "domain": "instant_payments_antifraud",
                "state": {
                    "valor_transacao_brl": 18885.0,
                    "horario_solicitacao": "23:42",
                    "conta_favorecida": {"reclamacoes_med_abertas_ultimas_24h": 3}
                },
                "question": "Qual a conduta obrigatoria do motor de risco?",
                "options": [
                    {"id": "autorizar", "label": "Autorizar"},
                    {"id": "bloqueio_cautelar_med", "label": "Bloqueio Cautelar MED"}
                ]
            })
        }
    ],
    "stream": False,
    "options": {"temperature": 0.0}
}

response = requests.post(url, json=payload).json()
decision = json.loads(response["message"]["content"])
print(f"Selected Option: {decision['selected_option_id']}")
print(f"Confidence: {decision['confidence']}")
print(f"Rule Triggered: {decision['rule_triggered']}")

---

🛡️ License

Apache 2.0. Open for commercial and research use.