Each phase solved one problem and uncovered the next. The architecture you see today is a consequence of these discoveries — not a plan drawn in advance. Status notes mark how far each piece has actually progressed.
Q1 – Q2 2024Identity & the original question
Research questionCan a system truly know a person?
It started from a personal loss. The first instinct — can we reconstruct a person? — quickly became the harder and more honest question: can a system genuinely know one, over time? That reframing set the direction for everything that followed.
What this phase produced
- Voice and identity experiments — early work on voice as a recognizable signal, and on identity as something the system holds rather than re-asks.
- Multi-dimensional personality modeling — a structured model of a person across cognitive and emotional dimensions, assembled from dialogue and cross-referenced signals.
- First multi-agent coordination — prompt chains and agent flows; dozens of arrangements tried and discarded.
What broke / what was missing
Representation without persistence collapsed. Nothing survived between interactions — every conversation restarted from zero. A model of a person that forgets the person is not a model at all.
What emerged
A hard architectural requirement: a person cannot be modeled meaningfully if nothing persists between interactions. The next problem was memory.
ResearchEarly-stage experiments in identity, voice and personality modeling.
Q3 – Q4 2024Memory & the open-loop problem
Research questionWhat actually persists — and what creates unresolved pressure?
If a system must remember, the question becomes what to keep. Raw transcripts decay into noise. We moved to structured facts, and to a unit of human life that no system was tracking: the open loop — an unresolved commitment still occupying working memory.
What this phase produced
- Fact-based memory — a design for abstracting content into structured facts (importance, source, context, relationships) rather than storing raw text.
- Open Loop detection — a stateful model for tracking unresolved commitments through their lifecycle, with signal extraction (futurity, obligation, ownership, dependency, consequence, closure).
- Semantic retrieval — an embedding-based retrieval approach to surface the right memory at the right moment.
What broke / what was missing
Memory without context produced noise. The system could record everything and still understand nothing about what mattered now. Recall is not the same as relevance.
What emerged
Persistence was necessary but not sufficient. The system needed a coherent situational model — a way to interpret the present, not just store the past.
SpecifiedOpen-loop and fact-memory designs specified; a runnable implementation is not yet established.
Q1 – Q2 2025Context, state & the first judgment
Research questionHow should the system interpret what matters now?
Context turned isolated facts into a situation. But the more the system understood, the more it wanted to act — and acting on everything is its own failure. This phase produced the first real attempt at judgment, grounded in a model of the person’s state and relationships.
What this phase produced
- Context Engine — unifies loops, profiles, calendars and preferences, and derives tone, urgency and depth for a given moment.
- Human State Engine — inference of cognitive load, attention and related state from calendar, communication and behavioral signals.
- Relationship reasoning — priority derived from who an action affects, not only what it is.
What broke / what was missing
Awareness without judgment became intrusive. More context made the system more eager, not more useful. Knowing more is not permission to do more.
What emerged
Intelligence needed a decision boundary, not merely more context. The central problem shifted from understanding to restraint.
BuiltHuman State Engine implemented as a typed, tested core (multi-state inference + intervention pipeline).
Q3 – Q4 2025Trust, restraint & bounded authority
Research questionWhen may the system act — and when must it stay silent?
This was the hardest problem, and the one most systems skip. The answer could not be a setting. Restraint had to be built into the architecture: a decision pipeline that scopes action to evidence and earned trust, and treats silence as a legitimate output.
What this phase produced
- A human-in-the-loop decision pipeline — gated evaluation that weighs confidence, interruptibility, risk and earned trust before any action.
- Silence as a first-class output — an explicit bias toward doing nothing when intervention is not justified.
- Scoped autonomy & traceability — authority granted per domain, and meaningful actions kept reviewable.
- Identity & privacy research — an abstraction-first posture designed to minimize raw-data retention.
What broke / what was missing
Autonomy could not be a global switch. Early, eager prototypes felt like surveillance rather than support. Trust given all at once is trust misplaced.
What emerged
Authority must be earned per domain, constrained by evidence, and kept reviewable. Trust became a structural property, not a promise.
BuiltDecision & restraint layer — built within the Human State core.
Q1 2026Architectural convergence
Research questionHow do separate engines behave as one persistent system?
By now there were several engines, each solving its own problem. Run independently, they drifted. The work of this phase was coordination: a single constrained sequence where each layer limits the next, with shared state instead of separate memories.
What this phase produced
- A coordination layer — Loop → Context → State → Relationship → Decision, sequenced and constrained.
- A cross-engine bridge — engines query one another for shared context rather than maintaining private worlds.
- Internal use — selected components have been used by the founder across internal coordination workflows.
What broke / what was missing
Intelligence fragmented across surfaces — each developing its own quirks. The system risked being several systems wearing one name.
What emerged
A constrained intelligence architecture whose layers govern one another — more than a collection of features. Consistency, not capability, was the achievement.
InternalSelected components used by the founder across internal coordination workflows.
Q2 2026The organizational question emerges
Research questionWhat if memory, judgment, authority and continuity must hold across an organization?
The same four problems — memory, judgment, authority, continuity — reappeared at a larger scale. And a hypothesis began to take shape: some of the difficulty AI faces inside a business may not be inherent to the work. It may be inherited from how human organizations were assembled — across departments, handoffs, systems and authority boundaries.
What this reframed
- Institutional state — memory and commitments that belong to the organization, not to any single agent.
- Authority that survives — mandates and accountability that persist when the actor changes.
- Agent vs. institution — the difference between a reliable assistant and a reliable operating structure.
What emerged
A separate organizational system: Corpiler. Not the personal engines relabeled — a distinct system that extends the same research principles (persistent context, bounded authority, evidence, judgment, continuity) into how an organization is compiled and run.
Research & frontierActive research and product direction — not a completed achievement.