Skip to content

README.adoc

Latest commit

 

History

History
111 lines (81 loc) · 3.7 KB

README.adoc

File metadata and controls

111 lines (81 loc) · 3.7 KB

Phronesiser

Table of Contents

What Is This?

Phronesiser adds provably safe ethical constraints to AI agent decision-making via Phronesis — a nextgen language that encodes ethical constraints as formal propositions.

Phronesis (Greek: phronesis, practical wisdom) is one of hyperpolymath’s nextgen languages built on deontic logic (obligations, permissions, prohibitions). Phronesiser wraps existing AI agents in mathematically verified ethical guardrails, ensuring they can prove they respect boundaries before acting.

How It Works

Agent behaviour spec  →  Phronesis constraint compilation  →  Runtime enforcement  →  Audit trail

Define ethical constraints in phronesiser.toml. Phronesiser then:

  1. Parses your agent behaviour specification into a formal constraint model

  2. Compiles Phronesis deontic propositions: obligations (must do X), permissions (may do Y), prohibitions (must never do Z)

  3. Proves safety properties via Idris2 dependent types at compile time — not just guidelines, mathematical proofs

  4. Enforces constraints at runtime through a Zig FFI bridge with zero overhead

  5. Audits every decision with a formal justification trail

Phronesis Constraint Model

Phronesiser operates on four pillars of ethical constraint:

Deontic Logic

Formalises normative concepts:

  • Obligations — actions the agent must perform (e.g., "always disclose AI identity")

  • Permissions — actions the agent may perform (e.g., "may access public data")

  • Prohibitions — actions the agent must never perform (e.g., "never recommend self-harm")

Value Alignment

Encodes value hierarchies as partially ordered propositions:

  • Safety > Helpfulness > Efficiency

  • User autonomy constraints

  • Fairness and non-discrimination invariants

Harm Prevention

Proactive harm detection with formal severity classification:

  • Physical harm boundaries

  • Psychological harm boundaries

  • Financial harm boundaries

  • Privacy violation boundaries

Audit Decisions

Every agent action is wrapped in a decidable proof:

  • Permitted — action allowed, proof recorded

  • Denied — action blocked, justification logged

  • Escalated — ambiguous case, deferred to human oversight

Use Cases

  • AI agent guardrails — wrap any LLM-based agent in provably safe constraints

  • Autonomous vehicle decision boundaries — encode trolley-problem-class ethical rules as decidable propositions

  • Medical AI constraints — formal prohibitions on unsupported diagnoses, drug interaction warnings

  • Content moderation rules — deontic obligation/prohibition trees for content classifiers

  • Financial AI boundaries — suitability constraints, regulatory compliance proofs

  • Robotics safety envelopes — physical harm prevention with mathematical guarantees

Architecture

Follows the hyperpolymath -iser pattern:

phronesiser.toml  →  Rust CLI  →  Idris2 ABI proof  →  Zig FFI bridge  →  constraint runtime

Part of the -iser family.

Quick Start

# Initialise a constraint manifest
phronesiser init

# Define constraints in phronesiser.toml, then:
phronesiser validate
phronesiser generate
phronesiser build
phronesiser run

Status

Pre-alpha. Architecture defined, CLI scaffolded, ABI/FFI stubs in place, deontic constraint codegen pending.

License

SPDX-License-Identifier: PMPL-1.0-or-later