agdr-aki 1.8.11

AgDR-AKI: Sub-microsecond AI accountability SDK. Court-admissible forensic records (PPP Triplet) with cryptographic proof for regulated AI (EU AI Act, Canada Evidence Act).

AgDR-Phoenix

Atomic Kernel Inference SDK for Phoenix v1.8.9

Developed by the Genesis Glass Foundation (Fondation Genèse Cristal), a federally incorporated Canadian not-for-profit. [e-mail]

"AgDR satisfies a CISO. AgDR satisfies a judge."

What It Does

AgDR-Phoenix seals every AI decision at the exact moment it occurs. Each record carries a PPP Triplet (Provenance, Place, Purpose), a coherence score, a reputation scalar, and a HumanDeltaChain linking machine output to human oversight or escalation.

The result is a tamper-evident, court-admissible decision record with a clear chain of custody from the instant of inference. Zero-knowledge proofs allow selective disclosure to regulators and courts without exposing the underlying reasoning.

Installation

Python (PyPI)

pip install agdr-aki

Rust (crates.io)

cargo add agdr-aki

Build from Source

pip install maturin[patchelf]
git clone https://github.com/aiccountability-source/AgDR-Phoenix
cd AgDR-Phoenix
maturin develop --release

Quick Start

from agdr_aki import AKIEngine, PPPTriplet, HumanDeltaChain

engine = AKIEngine("records.db")

ppp = PPPTriplet(
    provenance="Phoenix v1.8 ACME",
    place="Toronto CA PIPEDA",
    purpose="AgDR"
)

hdc = HumanDeltaChain(
    agent_decision_ref="aki_001",
    resolved=True,
    terminal_node="autonomous"
)

record = engine.capture(
    ctx={"claim": "CLM-12345"},
    prompt="Process claim #12345",
    reasoning_trace={"steps": ["analyze", "decide"]},
    output="Approve",
    ppp_triplet=ppp,
    human_delta_chain=hdc
)

print(record.id)
print("Coherence:", record.coherence_score)
print("Reputation:", record.reputation_scalar)

Zero-Knowledge Proofs

AgDR-Phoenix supports lightweight on-demand zero-knowledge proofs on sealed records.

zk_proof = record.generate_zk_proof(mode="full_validity")
court_pkg = record.to_court_package(include_zk_proof=True)
print(court_pkg)

Properties

• Proves the record is contemporaneous, tamper-evident, PPP-compliant, and correctly signed
• Reveals nothing about the original context, prompt, reasoning trace, or output
• Verification uses only the public Merkle root
• Designed for regulators and courts under PIPEDA and the Canada Evidence Act

Core Concepts

Concept	Description
PPP Triplet	Provenance, Place, Purpose immutable legal anchor recorded at inference time
Atomic Kernel Inference (AKI)	Sub-microsecond sealed decision capture at the exact moment of inference
HumanDeltaChain	Required fiduciary link tracking human oversight or escalation and enforced at API level
Coherence Score	Normalized cosine similarity against the agent's historical spine (0.0 - 1.0)
Reputation Scalar	Exponentially weighted moving average of past coherence, provides a longitudinal reliability signal
Sealed Record	Tamper-evident post-encryption object with BLAKE3 hash and Ed25519 signature
Zero-Knowledge Proof	Selective disclosure of record integrity without revealing decision content

Evidentiary Architecture

Layer	Implementation
Hashing	BLAKE3
Signing	Ed25519 (downstream of AKI hot path)
Chaining	Forward-secret Merkle tree
Persistence	Write-Ahead Log (WAL)
Degradation	Limitations Act 2002 (Ontario): 2-year basic, 15-year ultimate
Admissibility	Canada Evidence Act, portable to EU AI Act (August 2026)
Privacy	Zero-knowledge proof layer for selective disclosure

Distinction from Structural Monitoring

AgDR captures semantic decision provenance: what the agent decided, why, under what authority, and with what confidence, sealed at the moment of inference.

Structural container monitoring (e.g., the AGA standard) records whether the system was running correctly. AgDR records what the system actually decided. Both serve different functions in an AI governance stack.

Specification and Standards

Resource	Reference
Reasoning Capture Methodology v1.0	ISBN 978-1-7389042-1-1
Library and Archives Canada	Deposited under CC-BY 4.0
Steward	Genesis Glass Foundation, federally incorporated, royalty-free anchor in articles of incorporation

License

Dual-licensed under Apache 2.0 OR CC0 1.0.

The AgDR open standard is stewarded by the Genesis Glass Foundation (Fondation Genèse Cristal), a federally incorporated Canadian not-for-profit. The royalty-free grant is anchored in the Foundation's articles of incorporation.

Contributing

See CONTRIBUTING.md. All contributions are welcome. The standard is open.

Contact

Genesis Glass Foundation

[e-mail]

https://accountability.ai

Resources

• AgDR Specification
• AKI Formal Definition
• Human Delta Chain Spec
• Technical Details
• PPP Pillars
• Changelog
• Contributing Guide
• Rust API Documentation
• PyPI Package
• Crates.io Registry