0-AI-MANIFEST.a2ml

# STOP - CRITICAL READING REQUIRED

**THIS FILE MUST BE READ FIRST BY ALL AI AGENTS**

## WHAT IS THIS?

This is the AI manifest for **phronesiser** — an ethical constraint engine that
adds provably safe deontic logic guardrails to AI agents via Phronesis. It declares:
- Canonical file locations (where things MUST be, and nowhere else)
- Critical invariants (rules that must NEVER be violated)
- Repository structure and organization

## PROJECT SUMMARY

**Phronesiser** compiles agent behaviour specifications into formally verified
ethical constraint sets using deontic logic (obligations, permissions, prohibitions).
Constraints are proven sound via Idris2 dependent types and enforced at runtime
through a Zig FFI bridge with zero overhead. Every agent action produces an
auditable decision: permitted, denied, or escalated.

Key concepts:
- **DeonticModality**: Obligation | Permission | Prohibition
- **EthicalConstraint**: Named constraint with modality, scope, and formal proposition
- **ValueAlignment**: Partially ordered value hierarchy (safety > helpfulness > efficiency)
- **HarmPrevention**: Severity-classified harm boundary (physical, psychological, financial, privacy)
- **AuditDecision**: Decidable proof — Permitted, Denied, or Escalated with formal justification

## CANONICAL LOCATIONS (UNIVERSAL RULE)

### Machine-Readable Metadata: `.machine_readable/` ONLY

These 6 a2ml files MUST exist in `.machine_readable/6a2/` directory ONLY:
1. **STATE.a2ml** - Project state, progress, blockers
2. **META.a2ml** - Architecture decisions, governance
3. **ECOSYSTEM.a2ml** - Position in ecosystem, relationships
4. **AGENTIC.a2ml** - AI agent interaction patterns
5. **NEUROSYM.a2ml** - Neurosymbolic integration config
6. **PLAYBOOK.a2ml** - Operational runbook

**CRITICAL:** If ANY of these files exist in the root directory, this is an ERROR.

### Anchor File: `.machine_readable/anchors/ANCHOR.a2ml` ONLY

Canonical authority and semantic-boundary declaration MUST exist at:

` .machine_readable/anchors/ANCHOR.a2ml `

Do not place `ANCHOR.a2ml` at repository root.

### Maintenance Policies: `.machine_readable/policies/` ONLY

Canonical maintenance/governance files MUST exist under:

` .machine_readable/policies/ `

Minimum required files:
- `MAINTENANCE-AXES.a2ml`
- `MAINTENANCE-CHECKLIST.a2ml`
- `SOFTWARE-DEVELOPMENT-APPROACH.a2ml`

Do not place maintenance policy files in repository root.

### Bot Directives: `.machine_readable/bot_directives/` ONLY

Bot-specific instructions for your automated agents.

### Contractiles: `.machine_readable/contractiles/` ONLY

Policy enforcement contracts (k9, dust, lust, must, trust).

### AI Configuration & Guides: `.machine_readable/ai/` ONLY

- `AI.a2ml` - Language-specific or LLM-specific patterns
- `PLACEHOLDERS.adoc` - Bootstrap guide

### Community & Forge Metadata: `.github/` ONLY

- `CODEOWNERS` - Review assignments
- `MAINTAINERS` - Machine-readable contact list
- `SUPPORT` - Support channels
- `SECURITY.md` - Technical security policy
- `CONTRIBUTING.md` - Technical contribution manual
- `CODE_OF_CONDUCT.md` - Conduct rules

### Agent Instructions

- `0-AI-MANIFEST.a2ml` - THIS FILE (universal entry point)

## CORE INVARIANTS

1. **No state file duplication** - Root must NOT contain STATE.a2ml, META.a2ml, etc.
2. **Single source of truth** - `.machine_readable/` is authoritative
3. **No stale metadata** - If root state files exist, they are OUT OF DATE
4. **License consistency** - All code PMPL-1.0-or-later unless platform requires MPL-2.0
5. **Author attribution** - Always "Jonathan D.A. Jewell <j.d.a.jewell@open.ac.uk>"
6. **Container images** - MUST use Chainguard base (`cgr.dev/chainguard/wolfi-base:latest` or `cgr.dev/chainguard/static:latest`)
7. **Container runtime** - Podman, never Docker. Files are `Containerfile`, never `Dockerfile`
8. **Container orchestration** - `selur-compose`, never `docker-compose`
9. **Deontic soundness** - Obligation and prohibition must never apply to the same scope simultaneously
10. **Audit completeness** - Every evaluated action must produce exactly one AuditDecision

## REPOSITORY STRUCTURE

This repo follows the **Dual-Track** architecture:

```
phronesiser/
├── 0-AI-MANIFEST.a2ml         # THIS FILE (start here)
├── README.adoc                 # High-level orientation (Rich Human)
├── ROADMAP.adoc                # Phase 0-6 ethics-specific roadmap
├── TOPOLOGY.md                 # Module map and data flow
├── CONTRIBUTING.adoc           # Human contribution guide
├── Justfile                    # Task runner
├── Cargo.toml                  # Rust CLI manifest
├── Containerfile               # OCI build (Chainguard base)
├── LICENSE                     # PMPL-1.0-or-later
├── src/                        # Source code
│   ├── main.rs                 # CLI entry point
│   ├── lib.rs                  # Library API
│   ├── manifest/               # phronesiser.toml parser
│   ├── codegen/                # Phronesis constraint compilation
│   ├── abi/                    # Rust ABI stub types
│   └── interface/              # Verified Interface Seams
│       ├── abi/                # Idris2 ABI (Types, Layout, Foreign)
│       ├── ffi/                # Zig FFI (constraint engine)
│       └── generated/          # C Headers (auto-generated)
├── container/                  # Stapeln container ecosystem
├── docs/                       # Technical depths
└── .machine_readable/          # ALL machine-readable metadata
```

## SESSION STARTUP CHECKLIST

Read THIS file (0-AI-MANIFEST.a2ml) first
Understand canonical location: `.machine_readable/`
State understanding of canonical locations

## ATTESTATION PROOF

**"I have read the AI manifest for phronesiser. All machine-readable content
(state files, anchors, policies, bot directives, contractiles, AI guides) is
located in `.machine_readable/` ONLY, and community metadata is in `.github/`.
I will not create duplicate files in the root directory. I understand that
phronesiser enforces deontic soundness (no obligation/prohibition contradictions)
and audit completeness (every action produces exactly one decision)."**