Scoring, interpretation and limits

The instrument is intended for executives and senior leaders describing current organizational conditions from their own perspective. It does not independently observe behavior or verify supporting evidence. Results may be affected by role, visibility, recall, response style, recent events and social desirability.

The production bank contains 114 questions across eight domains. A session selects 35–50 questions; the default is 40. Selection first applies any sector and organization-size eligibility rules, then greedily favors questions that improve coverage of constructs below the target of three observations. Remaining positions favor domain balance. Selected questions are interleaved across domains and their order varies by session.

Different respondents may therefore receive different item sets. Scores can be compared descriptively, but no equivalence study has yet established that every generated form has the same measurement properties.

Every response is an integer from 1 to 5, with 5 always representing the healthiest described condition. Each item is tagged to two or more constructs with a weight from greater than 0 through 1. For construct c, the engine computes:

construct score(c) = Σ(item weight × response value) ÷ Σ(item weight)

The score is a rolling weighted mean on the 1–5 response scale. An item tagged to several constructs is an observation for each tagged construct. The current engine counts observations, but it does not estimate standard error, internal consistency, item discrimination or a confidence interval.

A metric first takes the equal-weight mean of its observed input construct scores. Construct scores with no observations are excluded. Item weights affect the construct scores beneath a metric; the constructs themselves are not differentially weighted when combined into the metric.

This means a partially completed Builder Index is not merely a less complete estimate of a fixed composite. Its composition changes as previously unobserved constructs enter the mean. An early reading and a final reading can therefore differ because the set of included constructs changed, even before considering changes in the answers themselves.

Builder Index and AI Readiness Indicator values linearly rescale the construct mean from 1–5 to 0–100. The 0–100 format is an index without a percentage unit. Execution Friction and Leadership Load invert the healthy-oriented mean so that a lower displayed value is better. All other metrics remain on the 1–5 scale.

Bands and unlock thresholds are design thresholds. They have not been derived from outcome data, normative samples, sensitivity/specificity analysis or externally observed breakpoints.

An unanswered item contributes nothing. Within a metric, an entirely unobserved construct is excluded rather than assigned zero. A metric becomes visible only after both its question-number threshold and minimum observation threshold are met. The adaptive session does not offer an “unable to observe” response or impute missing answers.

The self-directed short form does, since version 1.1. Its sixth response, “Insufficient evidence”, is stored as 0, is excluded from every condition mean, and is reported as evidence coverage: the share of the 24 items the respondent could substantiate. A condition scores only when at least three of its six items were substantiated; below that no index is computed and the reading names the conditions that lack evidence. Team Mode administration of the same 24 items is unchanged and offers no such response, so group aggregation, divergence and the widest-item selection are untouched.

Evidence sufficiency is answer coverage, not statistical confidence. The labels are Early for 0–9 answered questions, Developing for 10–20, Established for 21–34, and Complete for 35 or more. “Complete” means the coverage threshold was reached; it does not mean the result is definitive, accurate or validated.

Each metric card also has an evidence-fill display calculated as observations divided by twice the metric's minimum observation requirement, capped at 1. That display is likewise a coverage ratio, not a confidence interval or reliability estimate.

The self-diagnostic does not estimate a causal constraint. It ranks unlocked metrics for which a recommendation exists, excluding the Builder Index. Each metric is normalized to a 0–1 severity value: healthy-oriented metrics use (maximum − score) ÷ range; inverted metrics use (score − minimum) ÷ range. The metric with the highest resulting severity is displayed first and supplies the “Address First” recommendation.

All nine non-Builder-Index metrics currently have recommendation entries, so catalog coverage does not presently narrow the field. Unlock order does. Executive Attention Index can enter the ranking at question 5; Alignment Coherence and Authority Distribution at question 8; Decision Velocity at question 10; Execution Friction and Leadership Load at question 12; AI Readiness Indicator and Risk Judgment at question 20; and Transformation Readiness at question 28. Before a metric unlocks, it cannot be selected regardless of what later answers may show. The displayed result is therefore structurally biased toward earlier-unlocking metrics during partial completion.

Ranking also assumes that a linear normalization makes the different metric scales comparable. That is a design assumption, not an empirically established property. In particular, the operation treats distances between adjacent response-derived values as equal and treats equivalent normalized positions on different composites as commensurable. The current evidence does not establish either claim.

The phrase “primary constraint” should therefore be read as indicated area to investigate first, not as proof that the metric causes other organizational outcomes or that improving it will necessarily produce the greatest effect.

Team reports require at least three respondents. For each unlocked metric, the engine reports the arithmetic mean, minimum, maximum, range and population standard deviation across respondents. Constructs receive the same descriptive summaries.

A divergence zone is flagged when the observed range exceeds 1.5 points on a 1–5 metric, or the equivalent 37.5 points on a 0–100 index. This is a design rule, not an empirically validated threshold. Divergence describes disagreement among respondents; it does not by itself establish that Alignment is the cause of the disagreement.

Since version 1.2 a participant in a team diagnostic may contribute their answers to research on the instrument. The offer is made after they have submitted and seen their own result, is refused at no cost, and is never made on the instruments that run entirely in the browser, whose pages promise that answers are never sent.

A contributed record holds the item responses, the instrument id and content fingerprint, sector, size and stage as bands, function and level from the fixed roster vocabulary, elapsed time, the month of contribution and an opaque organization key. It holds no name, no address, no title, no organization name, no team or response identifier and no timestamp for the answering itself. There is no path from a record back to a person. Retention is five years, and a withdrawal code shown once at contribution deletes the record on presentation; only a one-way hash of that code is stored.

This is what makes the analyses below possible and it does not perform them. Every analysis in the validation protocol needs item-level responses, and until this existed no store held any: team data is deleted at ninety days by promise, and the engagement archive deliberately keeps group means only. A corpus is a precondition, not a result, and nothing in this note's claims changes until an analysis has been run and reported here with its sample, its date and this fingerprint.

Version 1.3 has automated tests for scoring behavior, question selection and aggregation. Those tests establish that the software follows its specified rules; they do not establish measurement validity.

No published study has yet established internal consistency, test–retest reliability, inter-form equivalence, factor structure, convergent or discriminant validity, criterion validity, predictive calibration, measurement invariance, normative bands or meaningful-change thresholds for this instrument. Until that work is completed, every output should be treated as developmental and directional.

The instrument is FROZEN for that work, and the fingerprint above is how a reader tells two versions apart. It is a content hash over every item's stem, construct tags and the text of all five behavioral anchors, so a reworded question or a redrafted anchor moves it even when the item count does not. A build in which the bank changes without this value and the version above changing with it fails, because responses collected either side of an unrecorded change cannot afterwards be pooled for any reliability or structure analysis, and nothing in the data would reveal that they had been. The analyses that will be run, the minimum sample each needs and, for each, the result that would count AGAINST the instrument are fixed in advance in the validation protocol held with this repository, along with the consent terms, the unit of analysis and the expert-review and cognitive-interview plans that come before any data at all. Fixing them before the first engagement is what stops an analysis chosen after seeing the data being reported as a test of it. Future technical-note versions will identify the data set, methods and results behind any validation claim. Changes to items, weights, constructs, thresholds or interpretation rules will be recorded in the version history below. What each analysis needs, and how much of that sample exists today, is published live at validation status, including the two thresholds that would fail: omega below 0.70 for a construct reported as a score, and an intraclass correlation below 0.60 for a measurement repeated with no intervention between.