How a reading is made

Worldview Lens doesn’t ask an AI to guess someone’s stage. It asks many small, typed questions about each passage, using written definitions of every stage, and combines the answers with plain arithmetic you can inspect. Here is all of it.

The engine: TypeSafe’s Jev

Scoring runs on TypeSafe’s Jev model. Jev doesn’t write prose; it answers structured questions and returns probabilities: a yes/no question comes back as a probability of yes, and a multiple-choice question as a probability for every option. That makes each judgment measurable and repeatable, and it lets Worldview Lens keep the reasoning (which definitions, which thresholds, how answers combine) in its own open code rather than inside a prompt.

Every passage is sent once, with all its questions asked in parallel. Claude (by Anthropic) is used only for writing: rephrasing a passage for another stage and the optional commentary. Claude never sets a score; when it rephrases, Jev checks where the rewrite actually lands.

What every passage is asked

1. Is there anything to read?

Does `text` contain enough reasoning, justification, or self-expression to infer the writer's way of making meaning? Slogans, quoted maxims or someone else's words, bare links, jokes, greetings, praise or thanks, one-word reactions and pure factual statements do NOT.

Below 35% the passage is left out and the reader sees “No altitude read” rather than a guess.

2. Is it satire?

Is `text` primarily satire, parody, or a joke — mimicking a genre or voice for comic effect rather than sincerely expressing the writer's own view?

Irony imitates the language of later stages (a parody of a self-aware LinkedIn post once scored as a very late stage), so satire is detected separately and left out of profiles.

3. Which stage is it written from?

For the model the reader chose, Jev gets a multiple-choice question whose options are that model’s stages, each defined by its gist, the signs it shows in writing and example sentences. Every definition is visible in the extension’s Learn tab. The question is framed by structure:

Judging by the STRUCTURE of meaning-making (how the writer frames self, others, causality, rules, feelings and perspectives) rather than the topic or opinion, from which ego-development level (Cook-Greuter) is this text most plausibly written?

and every stage question carries the same instruction, refined through testing:

Rate the stage the text is written FROM, not the stage of the people, ideas or theories it describes. Explaining, summarizing or praising a developmental theory, stage, tier or value is TOPIC: a third-person, linear, informational summary of a later stage reflects the writer's own expository reasoning, not that stage. Judge how the writer themselves reasons about a live question, including their own position. Judge structure, not sophistication: later stages are often written in plain, everyday words (seeing one's own part in a recurring pattern, holding two truths at once, noticing a label or story as it forms), while earlier stages can borrow fashionable words (systems, perspectives, growth, holistic) without using them.

Jev returns a probability for every stage. That full distribution is kept, not just the top answer.

4. Which structural signs are present?

Separately, 11 yes/no questions look for signs that lean toward particular altitudes. They become the evidence chips on a reading and the “signs found” under each passage.

From passages to a reading

1. Pooling. Each person’s readable passages are combined into one distribution, weighting longer and more readable passages more.
2. Center of gravity. The probability-weighted middle of that distribution, named as a stage, or a transition (“Achiever / Pluralist”) when it sits between two, or a mix when two separate stages both carry real weight.
3. Range. The neighbouring stages that together hold most of the probability.
4. Confidence and caution. A reading resting on fewer than 40 words or with confidence under 40% is marked tentative. A single passage whose confidence is under 60% is marked uncertain, and the reader can open “Why this stage” to see the stages it was weighing.
5. One scale. Every model’s stages sit on a shared twelve-band altitude scale (after Ken Wilber), so readings can be compared across models and placed on one map.

How accuracy is tested

Before any change to the questions or definitions ships, it is run against a benchmark of 298 passages across all nine models. The cases are written to be hard, in real registers (posts, emails, reviews, memoir), with deliberate traps:

Half of the cases are used for tuning and half are held out, so an improvement has to hold on passages it was never tuned on. A weekly automated run re-checks the scorer against the published numbers, because the underlying model can change.

Benchmark of 2026-10-14, scoring version 8.

Read these numbers honestly. The benchmark labels were written by the Worldview Lens team to the published stage descriptions. They are not protocols scored by certified raters, so the table shows that the scorer is consistent with the definitions it is given, and resistant to common traps. It is not proof of validity against instruments like the SCTi-MAP, WUSCT or the Subject–Object Interview.

What changed, and why

• Written from, not about. An encyclopedic summary of “second tier” thinking once scored as a very late stage. The scorer is now told that explaining a stage is topic, not stage.
• Structure, not sophistication. Later stages written in plain words were being pulled down a stage. Telling the scorer to judge structure lifted those cases from 67% to 100% without hurting the rest.
• Sharper neighbours. In Torbert’s model, the Ironist (who plays with frames while acting) was read as the Individualist (who has just discovered frames are relative). Clarifying the difference took Torbert from 88% to 100%.
• A check that didn’t work. Comparing the stage with the structural signs looked like a free second opinion. On the benchmark, their disagreements never predicted a wrong reading, so it was dropped rather than shown.

The nine models

Stages too rare or state-dependent to read from writing stay on the scale for comparison but are never assigned.

What a reading is not

It reads a piece of text, not a person: the same person writes from different places on different days and topics. It is not a validated psychological instrument. Use it to understand and to bridge, not to judge.

Generated from the Worldview Lens source (scoring version 8). Questions on the method: support@worldviewlens.com.