A schema, not a quiz

From the outside, a career assessment looks like a form with opinions. A few dozen questions, some letters, a PDF. That framing is why most assessment tooling dies in a folder: it optimizes the artifact, not the data.

We build ProfileClaw from the opposite direction. The instrument is a measurement schema; the quiz is merely its interface. Everything interesting happens after the last question — in how the data is structured, versioned, and served.

An instrument is a measurement

Take RIASEC. It is not six letters; it is a few dozen answered items collapsed into six dimension scores, each carrying a confidence derived from how consistently and how extremely the person answered. Big Five does the same for working style across five factors; our values and skills instruments add two more tables with their own scoring methods.

Concretely: 48 answered items become 6 interest dimensions. Twenty-two dimensions across the four instruments become one profile. The questions are the collection mechanism. The schema — dimensions, scores, confidence, versions — is what downstream systems actually consume.

Three layers, three audiences

The single most consequential design decision: raw answers, dimension scores, and narrative summaries are different products, stored and served as separate layers.

• Answers are the evidentiary floor — the most sensitive layer, useful for audit and re-scoring, appropriate for almost no consumer.
• Scores are the quantitative layer — comparable, rankable, machine-friendly.
• Summaries are the narrative layer — what most agents and humans actually want, in language they can quote.

An integration does not read "the data." It reads a layer, and the choice of layer is the consent decision. A coach might be granted scores; a Q&A agent gets summaries; nothing gets answers unless the person means it.

Versions are not optional

Instruments improve. Scoring methods get recalibrated, items get refined. If your middleware has no schema version, every improvement is also a silent mutation of history — yesterday's conclusions quietly reinterpreted under tomorrow's method.

Every ProfileClaw profile carries a version. When a scoring method changes, re-scoring old answers is an explicit, visible act — the profile moves to the new version with its history intact, never reinterpreted in place. Anyone downstream can pin, compare, or migrate.

Confidence, or it didn't happen

Self-report has signal quality. Someone answering a values instrument inconsistently is producing a different kind of fact than someone answering cleanly, and a middleware that flattens that difference is lying by rounding.

So confidence ships as a first-class field next to every score. A consumer that reads artistic 0.84 at confidence 0.81 can weight it accordingly — or ask, or double-check. Numbers without confidence invite downstream overconfidence; we would rather ship the uncertainty than hide it.

One envelope, two transports

The same profile reads the same over REST and over MCP. A program gets JSON; an agent client gets native tools. One canonical shape, no per-client dialects, because a middleware that speaks differently to different clients is not a middleware — it is a compatibility problem with a marketing budget.

The quiz was never the product. It is the door. The schema behind it is what makes human data something a harness can build on.