Launching Australia’s National Sustainability Data Platform Redesigning-Australias-Business-Digital-Identity-Experience Redefining-Mobile-Banking-Interaction Accelerating AI Insight Without Losing Depth Fueling Growth with Customer Insight Steering Smarter Digital Investment Across Government National Health Records From Friction to Function Connecting Officers to Real-Time AI Situational Awareness Redesigning Authorisation for Inclusive Access From Legacy Quoting to a Live Jobs Pipeline in Eight Weeks

Complete infrastructure to run governed AI

DSS - One operating system for identities, data lineage and autonomous agents.

Built a model-agnostic, threadless chat harness that keeps conversational context when users switch models mid-session (dualsubstrate.com)

Control Plane - every app surface is governable, every person and every agent is a named Principal, and every answer traces back to the ledger entries it drew on.

THE CHALLENGE

Lack of deterministic data and increasing semantic drift.

At scale, context windows buckle. Vector-RAG pipelines return documents that are lexically or conceptually adjacent but often structurally incompatible with the state of the task; and nobody can easily say afterwards why a given document was in the context at all.

MY APPROACH

AI harness: Ledger Oriented Agent Memory (LOAM)

I've built a memory lane to be discrete and durable; a control plane for governed memory that separates what AI computes from what your system keeps. Inference stays probabilistic. Custody stays deterministic. The harness uses an immutable ledger with each turn a Portable Evidence Object (PEO): an attestation with a permanent ID that any agent can cite and anyone can open.

Detail: The graph shows why it matters: The white node is this record.  Blue nodes are the earlier records it builds on and the references it declares. Orange lines are indexed relations. Green lines are prime-factor neighbours: records that share part of their structural address. That's relatedness by structure, not "these texts sound similar."

THE IMPACT

Model-agnostic, threadless chat harness that keeps conversational context 

A multi-turn conversation that keeps conversational context when users switch models mid-session is native to experience.  Th ledger-oriented agent memory is alway deterministic - signed, hash-chained, append-only events in RocksDB, principal-based identity aligned with verifiable credential and DID patterns, and AND-gated validation before anything is committed to memory.

Results include recall@1 of 1.00 at 517K tokens and Precision@5 of 0.93 vs 0.51 for vector RAG (p = 0.016), with open-domain QA limits reported alongside. This is part of an open technical disclosure (July 2026) backed by a reproducible, containerised multi-seed evaluation and a CI-enforced claims registry.

Detail: Each write adds relationships to the ledger. In one small controlled experiment on a single model, history-aware retrieval beat a no-history baseline, and because the index is rebuilt from the ledger, what the system accumulates is auditable.