How a read is made

A read is built from what you give it: pasted signal, documents, and —when the subject has shared one—their own measured Stance360 result. What you paste stays with the read—refining and re-validating need the original inputs; What we store spells that out. The read is an inference from that material, never a measurement of the person—every page that shows a read says so, and the one exception is labeled: a stance line backed by the subject’s own answers is marked measured.

The material is read through separate categories, each answering a question no other category answers: public signal (how they position themselves), training signal (how they develop themselves), the work record (how they have performed), their own answers (how they approach), what you know firsthand (how they deliver), and their wiring (how they’re built).

Splitting the evidence is the method: where categories agree, the read says so; where they disagree, the disagreement is surfaced rather than averaged away—the performed self in a bio and the delivered self in your firsthand notes are different evidence, and the seam between them is usually the most useful line in the read. A category the material can’t support is named as missing, never filled in: wiring, for instance, reports itself unavailable on every read without a measured instrument behind it, because a claim requires the inputs the claim is made of.

Not all evidence weighs the same. The subject’s own measured answers outrank what you know firsthand, and firsthand outranks the public record—public signal leans toward the performed self. Every read carries its confidence near the top, and the chip opens into why it sits at that level and the single move that would raise it.

Asking a model to follow rules is not the same as making the rules hold, so the ones that matter are enforced after generation, in code, with tests pinning each guard:

Internal codes never render. The framework’s internal shorthand is checked for and blocked before a read reaches anyone. Messages obey the voice rules—openers are checked, not hoped for. A score never ships without its reasoning: when a fit number comes back with no path explaining how the person could be a fit, the number is withheld rather than the reasoning invented. And a gauge that would show the wrong number shows nothing—an objective without its own fit says so instead of borrowing another objective’s score.

A read is a defensible provisional conclusion—it starts from what you have and is built to be revised, not defended. A living subject is an advantage no archive offers: sources you can question, behavior you can watch land, a person you can simply ask. So every read carries a refine door: add what you learned, and the next read leads with what moved—which calls the new evidence changed, and which held.

Every read asks its owner the only question that grades it: “Was this read accurate?”—on the page, and again by mail a month later, answered by the person who actually knows the subject. Those verdicts are the product’s calibration record.

The honest state today: the loop is live and the tally publishes when real outcomes exist. This page will cite the real distribution—recognizably them, partly, not them—and until there is one, it claims no accuracy number. A method page that quoted a figure it hadn’t measured would fail its own test.

What has been measured, as of September 2026—instrument evidence, which is a different claim from outcome accuracy: across four scored pre-registered prediction sets on historical figures—every prediction committed in writing before any material entered the engine, and checkable rather than taken on trust: the locked registrations are commits 1d38082, 3383ca8 and 6f99650 in the product’s own history, each dated before its rolls—18 of 19 figure-level calls landed in their pre-stated bands, and 10 of 10 aggregate predictions held. Under the same locked-in-advance discipline, test-retest reliability: eleven of eleven fresh regenerations landed the same stance family as their figure’s standing read (across two figures in two different families), with one figure’s exact signature identical three of three. Separately, where the house’s independently ruled designations and the engine’s own fresh reads of the same 23 figures are compared, the two methods agree on the stance family 20 of 23 times (87%) and on the exact signature 13 of 23 (57%)—concordance between two methods reading shared material, not validation against outcomes, and the disagreements are published to ourselves, not filtered.

One deliberate visual choice, stated so it isn’t mistaken for an omission: reads are documents, and documents are paper. The products render in a single light, print-like theme rather than following the operating system’s theme.

Role assessments add three rules. Consent stays specific: the record of how a candidate’s result reached the assessment names the role; the candidate’s own acknowledgement covers the general terms and names no role—and the two claims are never merged into one sentence. A structural ceiling binds the score when the architecture says the fit can’t be natural, and the ceiling note travels with the number. And every assessment ends the same way, on screen and in every copy of it: it is one input to a decision a human makes.