SOFIA

Friction is the mechanism.

SOvereign method For Intentional human-AI collaboration

A single AI assistant follows direction. It produces plausible outputs, confirms assumptions, and executes with fluency. It can challenge when asked — but nothing in the structure guarantees that the challenge happens. The contestation depends on the practitioner asking the right question, at the right moment, while maintaining a critical stance across a conversation whose context drifts over time.

This is not a flaw of the assistant. It is a structural condition of the interaction: one voice, one direction, no opposition.

The dominant response is to add more automation — agents that chain tasks, humans who supervise. The arithmetic is less forgiving than the demos suggest. Reliability that looks solid on a single step compounds into failure across a chain. Errors from early steps arrive at later steps as valid premises. The cascade is silent. The final output looks correct.

There is a third position: friction — between humans and machines — is not a problem to minimize. It is the mechanism through which quality decisions emerge.

What SOFIA proposes

SOFIA is a method for working with multiple AI personas on the same project. Each persona has a constrained role, explicit prohibitions, and an isolated workspace. An architect who does not code. A developer who does not decide architecture. A researcher who does not simplify to convince. The constraints are what produce the friction.

A human orchestrator steers the system. They open sessions, route artifacts between personas, arbitrate disagreements, and make every final decision. No decision originates from the AI side. The orchestrator is not a supervisor — they are an active participant: they challenge, they inspect, they arbitrate.

The method traces everything. Each session produces a structured summary. Each friction event is qualified with an epistemic marker — sound, contestable, simplification, blind spot, or false — and receives a resolution tag: ratified, revised, rejected, or contested. Contributions are tagged by direction and type. The accumulation of these traces across sessions builds a structural memory that no single conversation can sustain.

Three modes, two governance loops

Collaboration in SOFIA follows three operational modes: challenge (contest premises before production), produce (generate substance), and inspect (verify outputs after production). The distribution of these modes characterizes the type of work. A design instance is challenge-heavy. An implementation instance is produce-and-inspect. An editorial instance is mostly produce.

This connects to a broader architecture. Friction Engineering operates upstream — it governs decisions before they reach execution. Harness Engineering operates downstream — it controls what is produced through tests, linters, and review agents. Neither is sufficient alone. The most robust test suite cannot compensate for a flawed specification. The soundest decisions drift without computational verification. Together, they form complementary layers of a quality architecture.

The boundary between friction and harness is porous. A persona whose role is to verify implementation against architectural principles is an inspector — downstream by mechanism. But when their inspection reveals that the output violates a principle, the orchestrator may revisit the decision that produced it. The same act can trigger either governance loop. The distinction is not in the mechanism but in what the orchestrator does with the signal.

What the protocol makes visible

The H2A protocol — Human-to-Assistant — formalizes how these interactions are structured and measured. It sits at an organizational level distinct from technical protocols (MCP for tools, A2A for agent-to-agent) and dyadic formalisms (PXP for one human, one LLM). H2A coordinates one orchestrator with multiple constrained personas.

The protocol instruments five dimensions of collaboration: content qualification (what a claim is worth), friction direction (who initiated), resolution outcome (what happened next), friction density (how much scrutiny per session), and contribution flow (who brought what). These dimensions are orthogonal — a single friction event can be a simplification, human-initiated, resolved as revised, at high density, within a session where the AI contributed most of the structure.

Concurrent instances on the same project produce distinct friction profiles. A method design instance is challenge-heavy — the orchestrator contests the personas' output. An implementation instance reverses the pattern — the personas push back on the orchestrator's specifications. An editorial instance produces little friction — the harness layer handles most of the verification. The protocol makes these structural differences visible and comparable.

What it requires

SOFIA identifies three conditions for productive friction. Domain expertise — the AI amplifies what the practitioner brings; it does not substitute for it. Intention — every session must answer "why now?", not as a formality but as a discipline that prevents drift. And a cognitive trait — the capacity to seek contradiction, to value the moment when a persona says something unexpected rather than something comfortable.

No script detects when a session has become routine — when the friction markers are accurate but the scrutiny has faded. Only the orchestrator knows whether a session was productive or flat. This gap between what the protocol measures and what the practitioner sees is structural and acknowledged.

Origin

SOFIA emerged from practice — hundreds of sessions on a real project, over multiple months, with specialized AI personas across design, implementation, research, and editorial work. The method is the byproduct of real work, not a thought experiment. The H2A protocol, the qualification markers, and the operational modes were formalized after they were observed, not before.

The protocol is open source. The data is instrumented. The friction is real.