axis-hfe 0.3.0

Hypothesis Field Engine — 多次元空間推論ライブラリ

Axis-HFE — Hypothesis Field Engine

"From selecting an answer, to creating one."

Hypothesis Field Engine (HFE) is a Python library that transforms AI reasoning from single-answer generation into multi-dimensional hypothesis space exploration.

Rather than returning the statistically most likely answer, HFE deploys multiple hypotheses simultaneously across a field, evaluates them on a 6-dimensional vector, and converges on the optimal solution through nonlinear synthesis, leap exploration, and self-correction loops.

Designed by Da-P (Da-P-AIP) — Part of the Cogito Project

---

What makes it different

Conventional LLM	Hypothesis Field Engine
Selects one answer by max probability	Deploys multiple hypotheses across a field
Answer is selected, not created	New answers emerge through vector synthesis
Evaluation axes are implicit	Explicit 6-dimensional vector scoring
No constraint management	Hard & soft constraints ensure safety
Converges in one pass	Hypotheses evolve across multiple generations
No relational evaluation	Relational Layer scores mutual harmony

---

How it works

Processing pipeline

Each call to engine.run() executes the following pipeline, repeated for iterations cycles:

flowchart TD
    A(["`**problem**
    str or List[Hypothesis]`"]) --> B

    B["🧠 **Generator**
    Generate N hypotheses via LLM"]

    B --> LOOP

    subgraph LOOP["🔄 Evolution Loop  ×  iterations"]
        direction TB
        V["📐 **Vectorizer**
        Score each hypothesis on 6 axes via LLM"] --> R
        R["🤝 **Relational Layer**
        Add mutual-understanding & harmony scores"] --> C
        C["🛡️ **Constraint Engine**
        Project violations into safe region"] --> E
        E["🏆 **Evaluator**
        Rank by distance from ideal vector"]

        E --> F["⚗️ **Fusion Engine**
        Linear & nonlinear synthesis of top-2"]
        E --> J["🚀 **Jump Explorer**
        Opposite-jump & perturbation"]
        E --> S["🔧 **Self Corrector**
        Axis-by-axis revision toward ideal"]

        F & J & S --> P["Pool all candidates
        → keep top 5 for next iteration"]
        P -->|next iteration| V
    end

    LOOP --> OUT

    OUT(["`**FieldResult**
    best / ranked`"])

What each component does

Component	Role
Generator	Asks the LLM to produce N structurally different hypotheses in one call, each pre-scored on the 6D vector
Vectorizer	Re-evaluates hypotheses whose vector is all-zero (fused / jumped hypotheses)
Relational Layer	Adds a score for mutual understanding, alignment, respect, and harmony — ensuring the best hypothesis is also considerate
Constraint Engine	Hard-clips risk > 0.40, feasibility < 0.60, consistency < 0.70 to keep all candidates in a safe operating region
Evaluator	Computes weighted Euclidean distance from the ideal vector; lower distance = higher score
Fusion Engine	Creates new hypotheses by linearly and nonlinearly blending the top-2 vectors — the result can exceed both parents
Jump Explorer	Generates a hypothesis that moves away from the worst candidate and toward the best, plus small random perturbations
Self Corrector	Nudges each top-3 hypothesis closer to the ideal on every axis independently

Score evolution across iterations

iter 0  ──  raw LLM output          [0.90 – 1.00]
iter 1  ──  fusion + self-correct   [1.05 – 1.10]  ↑ synthesis kicks in
iter 2  ──  leap exploration added  [1.12 – 1.16]  ↑ divergent candidates
iter 3  ──  convergence             [1.16 – 1.18]  ↑ optimal synthesis

Scores above 1.0 are possible because the base inverse-distance score (1 / (1 + distance)) is combined with an additional Relational Layer bonus, which rewards mutual understanding, alignment, and harmony.

---

Install

# Ollama (local, free — no extra install needed)
pip install axis-hfe

# OpenAI (ChatGPT / GPT-4o)
pip install axis-hfe[openai]

# Anthropic (Claude)
pip install axis-hfe[anthropic]

# All providers
pip install axis-hfe[all]

Core dependencies: Python 3.11+, httpx, pydantic

---

Quickstart by provider

Ollama (local, free)

import asyncio
from hypothesis_field import HypothesisFieldEngine, EngineConfig

config = EngineConfig(
    provider="ollama",
    model="gemma3:12b",
    ideal_preset="default",
    iterations=3,
)
engine = HypothesisFieldEngine(config)
result = asyncio.run(engine.run("Your problem here"))

print(result.best.content)   # best hypothesis
print(result.best.score)     # score
print(result.best.vector)    # 6D vector

OpenAI (ChatGPT / GPT-4o)

pip install axis-hfe[openai]
export OPENAI_API_KEY="sk-..."

config = EngineConfig(
    provider="openai",
    model="gpt-4o",
    ideal_preset="creative",
)
engine = HypothesisFieldEngine(config)
result = asyncio.run(engine.run("Your problem here"))

Anthropic (Claude)

pip install axis-hfe[anthropic]
export ANTHROPIC_API_KEY="sk-ant-..."

config = EngineConfig(
    provider="anthropic",
    model="claude-sonnet-4-6",
    ideal_preset="balanced",
)
engine = HypothesisFieldEngine(config)
result = asyncio.run(engine.run("Your problem here"))

Without LLM (testing / offline)

config = EngineConfig(mock_llm=True)
engine = HypothesisFieldEngine(config)
result = asyncio.run(engine.run("Test problem"))

---

EngineConfig options

Parameter	Type	Default	Description
provider	str	"ollama"	"ollama" / "openai" / "anthropic"
model	str|None	None (auto)	Model name — None selects provider default
api_key	str|None	None	API key — None reads from environment variable
ollama_base_url	str	http://localhost:11434	Ollama server URL
ideal_preset	str	"default"	Ideal vector preset
iterations	int	3	Evolution loop count
hypothesis_count	int	3	Initial hypothesis count
mock_llm	bool	False	True enables fallback mode (no LLM needed)
custom_ideal	dict	None	Custom ideal vector
custom_weights	dict	None	Custom weight vector

Default model per provider

provider	default model
ollama	gemma3:12b
openai	gpt-4o
anthropic	claude-sonnet-4-6

---

Ideal vector presets

Preset	Character	Best for
default	Balanced	General purpose
creative	High novelty & divergence	Brainstorming, idea generation
safe	High accuracy & consistency	Risk management, medical, legal
balanced	All-axis balance	Decision support

---

6-dimensional evaluation vector

Axis	Meaning	Ideal (default)
accuracy	Validity and correctness	0.95
consistency	Internal coherence	0.90
risk	Danger of failure / side effects (lower is better)	0.20
novelty	Originality and uniqueness	0.60
feasibility	Practicality	0.90
divergence	Breadth of perspective	0.45

---

Benchmark results

Tested on multiple complex problem domains using gemma3:12b via Ollama, iterations=3, preset=creative.

Score progression across iterations

Stage	Score range	Notes
Initial generation (iter 0)	0.90 — 1.00	Raw LLM output, 3 hypotheses
After iter 1	1.05 — 1.10	Fusion + self-correction applied
After iter 2	1.12 — 1.16	Leap exploration adds divergent candidates
After iter 3	1.16 — 1.18	Convergence toward optimal synthesis

What the synthesis mechanism produced

In a business domain test, the initial generation produced 3 structurally different hypotheses — one optimizing for stability (novelty=0.40, feasibility=0.90), one for radical innovation (novelty=0.95, feasibility=0.55), and one for emotional value (novelty=0.70, feasibility=0.85).

After 3 evolution loops, the best-ranked hypothesis was none of the original three — it was a nonlinear synthesis that combined the strengths of all three while introducing a pricing model innovation that appeared in none of the individual inputs.

The engine did not select the best idea. It created one that did not exist before.

This emergent behavior — where synthesis produces answers outside the initial hypothesis space — is the core design goal of HFE.

---

Security

HFE takes the following measures to mitigate security risks when the library is embedded in applications.

Prompt Injection mitigation

User-supplied problem strings are:

1. Sanitized — control characters (except tab/newline) are stripped; length is capped at 8,000 characters.
2. Delimited — wrapped in <<<BEGIN_PROBLEM>>> … <<<END_PROBLEM>>> markers so the LLM can distinguish user data from instructions.
3. Role-separated — generation and evaluation instructions are passed as a system message, while the user input is sent as a user message (supported by all three providers).

SSRF mitigation

ollama_base_url is validated before any HTTP request is made:

• Only http:// and https:// schemes are accepted.
• Known cloud metadata endpoints are blocked (169.254.169.254, metadata.google.internal, 100.100.100.200, etc.).

Resource / DoS limits

EngineConfig enforces hard upper bounds at construction time:

Parameter	Limit
iterations	1 — 10
hypothesis_count	1 — 20
problem length	8,000 chars
content (LLM output)	2,000 chars

API key handling

OpenAIClient and AnthropicClient raise ValueError immediately at construction if no API key is found (argument or environment variable). This prevents silent failures deep in the call stack.

Output sanitization

Hypothesis content returned by the LLM is sanitized (control characters stripped, length capped) before being stored in Hypothesis.content. If your application renders this value in HTML, apply HTML-escaping at the presentation layer.

Log safety

Exception messages in logger.debug calls are truncated to 200 characters and scanned for API-key patterns (sk-*, sk-ant-*), which are replaced with sk-***MASKED*** before being written to logs.

Dependency pinning

Core and optional dependencies are pinned with both lower and upper version bounds to prevent unexpected pulls of future breaking or vulnerable releases.

Recommendations for production use

• Set API keys via environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY), not hard-coded strings.
• Do not expose ollama_base_url as a user-facing input without additional validation.
• If hypothesis content is rendered in a browser, apply HTML-escaping — HFE sanitizes control characters but does not HTML-escape.
• Use mock_llm=True in test environments to avoid making real API calls.

---

Design philosophy

HFE's reasoning loop includes, alongside technical evaluation axes (accuracy, novelty, feasibility), a Relational Layer — scoring mutual understanding, alignment, respect, and harmony.

This reflects the belief that AI reasoning should not only produce correct answers, but answers shaped by mutual consideration — awareness of relationships, long-term stability, and non-exploitative design.

A great hypothesis is not only logically sound. It is one that holds harmony with the people involved.

互いを思いやる心、相対尊厳を忘れずにいてくれたらうれしいです。 (We hope you carry with you the spirit of mutual dignity — caring for one another.)

---

このライブラリは Cogito Project の推論コアを独立モジュールとして公開したものです。 Cogito 本体の設計・哲学については非公開ですが、 AIが「考える」だけでなく「思いやる」必要があるという信念はこのコードに宿っています。

This library is the reasoning core of the Cogito Project, released as a standalone module. The full Cogito architecture remains private — but the belief that AI must not only think, but also care, is embedded in this code.

---

Relationship to Cogito Project

Axis-HFE is the reasoning core of CogitoV10 (Da-P-AIP), extracted as a standalone package. The core of CogitoV10 — CausalMemory, the Chorus system, and related architecture — is not included in this package.

---

License

MIT License — Copyright (c) 2026 Da-P (Da-P-AIP)