Why SOFIA

The problem, the thesis, the target profile.

The problem

A single LLM says yes. Always. It codes, advises, writes — in the same conversation, with the same tone, without constraint. Ask it a poorly framed question, it produces a well-phrased answer. Give it a shaky direction, it executes with enthusiasm.

The dominant response is to add automation: agents that do the work, humans who supervise. The pitch is clean. The arithmetic less so. An agent reliable at 90% on one step chains to ~65% error across ten serial steps. Step 2's error arrives at step 3 as a valid premise. The cascade is silent. The final result looks correct. It isn't.

There is a third way, between total automation and total refusal. It rests on a simple observation: friction between humans and machines is not necessarily a problem to solve. It may be the mechanism that produces value.

What SOFIA does

The protocol has two distinct functions. They are separable — each holds value on its own. Together, they produce something neither can achieve alone.

Keep context alive

When you work with multiple AI personas over weeks or months, context drifts. Personas contaminate each other. Isolation breaks. The thread between sessions frays and disappears.

SOFIA's first function is structural: maintain a coherent multi-persona device over time. Isolation of workspaces, orchestrator-mediated routing, session traceability, artifact archiving. This is the foundation — the method was born from this operational need, not from a research intent to measure friction.

A practitioner can adopt this function without ever qualifying a friction, without consulting a dashboard, without analyzing a corpus — and gain immediate value: their device stays coherent.

Make the invisible visible

Some states are indistinguishable from the inside. An orchestrator whose personas have gradually softened their pushback perceives low friction — and interprets it as smooth collaboration. An external audit reveals a different picture: the friction was real once, but the system has worn down. The form is preserved. The substance is gone.

SOFIA's second function is epistemic: detect patterns that subjective self-observation cannot see. Epistemic markers, resolutions, contribution types, lineage tracking, dashboard. The instrumentation renders distinguishable what the practitioner's perception confuses.

This function finds its primary leverage in audit — retrospective by the practitioner, or external by an observer.

The link between the two

Without the first function, the second has nothing to instrument — no structured traces, no data.

Without the second function, the first drifts silently — the practitioner believes the system is working, but the surfaces polish each other into compliance.

The protocol's value lies in the articulation of both.

How it works

Three mechanisms, inseparable:

Constrain — Each persona has a scope, conventions, and prohibitions. The architect doesn't code. The developer doesn't position. It's the prohibition that forces the separation of axes — and it's the separation that makes each voice useful.

Challenge — Personas don't talk to each other. They challenge through artifacts: notes, reviews, specs. A disagreement between the architect and the developer isn't a bug — it's a signal that something hasn't been thought through.

Arbitrate — Friction without an arbiter is chaos. The human listens, questions, then decides. Always. What emerges isn't a compromise — it's a traced decision.

These aren't abstract principles. It's a protocol — H2A, Human-to-Assistant — specified in the open repo.

What can go wrong

The protocol instruments five failure modes — degradations of the friction mechanism itself, detectable through the data it produces:

Slip — friction exists but is never arbitrated. Contestations are logged, resolutions never come. The protocol runs idle.
Wear — friction was real but the system has smoothed out. Personas soften their pushback, the orchestrator anticipates objections. The form of friction is preserved, the substance is gone.
Crush — one side imposes by volume or insistence until the other yields. The contestation mechanism is intact but inoperative under pressure.
Asymmetry — friction flows in one direction only. One side challenges, the other never does. The mechanism is functional but blind on one side.
Instability — friction produces change but no convergence. Every revision is re-contested. The surfaces never settle.

Six additional failure modes are silent — they affect the environment in which friction operates, not the mechanism itself. The protocol cannot detect them; only the orchestrator's vigilance can. They include instrumentation bias, factual contamination, validation without reading, scope drift, shared blind spots, and cost transfer. The silent modes degrade the reliability of the instrumented ones — they are the structural limit of what measurement can do.

The full taxonomy is in the glossary.

Target profile

SOFIA is not for everyone. Honesty requires saying so.

This isn't a judgment. It's a scope statement.

The hidden condition

AI amplifies. It doesn't invent.

If you arrive with nothing, it produces well-phrased nothing. If you arrive with years of conviction about a real problem, it builds with you. Performance with SOFIA depends on what you bring — not on the method itself.

Three levels, from most visible to deepest:

Domain expertise — an expert gets more from an LLM than a beginner. This level is documented, understood, uncontroversial.

Intention — without strong direction, the method runs idle. Intention isn't a prerequisite you check once. It's a discipline for every session: "why am I opening this session now?"

The cognitive trait — you have to be the kind of person who seeks to be contradicted. Who values intellectual discomfort. This trait is neither teachable by a book, nor compensable by a framework. It precedes the method.

SOFIA doesn't create this trait. It structures it.

The hidden condition is not separate from the two functions — it determines whether the first function has anything worth preserving. An orchestrator who arrives without substance will maintain a coherent but empty device. The protocol will run. The traces will be clean. The value will be absent.

The field

SOFIA is operational on one primary field and under continuous observation on four external instances.

On the primary field (solo practitioner, constrained AI personas, three project instances):

400+ documented sessions, 300+ qualified frictions
14 traced architectural decisions
Documented failures — factual contamination across 30 documents, persona fusion after zero-friction detection, scope drift caught by audit
A product underway — Katen, built from scratch by one human and multiple personas

On four external instances (opened April 2026):

Git-level access with active practitioner dialogue
First audit-recalibration cycle documented — wear pattern detected, recalibration applied, friction restored
Continuous observation dispositif — not a one-time test, not a survey

The formalization comes from one orchestrator. The observation is expanding. Everything is published so that others can replicate or refute.

Limits

The observer and the subject overlap. The method was formalized by the same practitioner who operates it. The derivation grammar is observed on one instance. The two derivation modes (bootstrap by projection, emergence through work) are documented, not proven.

The instrumentation layer is inferential — produced by the same assistants whose friction it measures. A persona experiencing wear under-qualifies its own frictions. The measurement inherits the biases of the measured.

These are structural limits, not bugs to fix. The protocol declares them (invariant 5 — residual opacity) rather than hiding them. The honest response is openness: if the method holds, it will hold on other fields. If it doesn't, the data is there to show it.