The machinery behind the hard test.

Each dimension is bipolar and reported from −100% to +100%. The midpoint means the weighted evidence landed between the two poles; it does not automatically mean apathy. Because the dimensions are scored independently, combinations that look unusual in a two-axis model are allowed to remain unusual.

The current v2.2 bank contains 350 active items: 108 conceptual and 242 applied. The two registers are intentional, because agreement with a principle and willingness to accept its consequences are different evidence.

Every item pairs its technical proposition with an “In other words” restatement. Many applied items also display an explicit “Assume that” premise, keeping the intended factual setup visible rather than hiding it in a tooltip. The v2.2 expansion adds 50 neutrally worded, overtly contentious items across subjects such as abortion, firearms, immigration, economic systems, gender identity, capital punishment, and nuclear force. These are tagged as a separate controversy-stress layer so their performance can be audited rather than blended invisibly into the bank.

Primary items are balanced toward both poles within every axis to reduce agreement bias. The bank also carries semantic metadata for policy domain, latent conflict, actor level, and policy instrument, supporting content and coverage audits.

Responses use a five-point agree–disagree scale from −2 to +2, plus a separate “Not sure / need more information” option. A numeric zero means you understood the proposition and are genuinely balanced; it remains in the score denominator. “Not sure” is stored as null and excluded from both numerator and denominator, lowering the reported coverage for that axis.

• Order: a session-seeded shuffle is reproducible and actively spaces same-axis items apart.
• Pacing: there is no timer, and progress saves to local storage after every answer.
• Passing: “Skip” moves ahead temporarily; you must return and record an answer—including “Not sure”—before submission.
• Version safety: a saved draft is only resumed against the same question-bank version.

The result is normalized to −1…+1 and displayed as a percentage. A question’s primary link always receives the item’s full weight. Some items also have bounded secondary links to related axes; all secondary weight from one question is capped at half that question’s own weight, so cross-loadings cannot overpower the construct the item was written to measure.

Scores are calculated three ways: conceptual only, applied only, and combined. Per-axis coverage records how much available primary weight received a numeric answer. Coverage below 50%, or fewer than eight numeric primary responses, is marked insufficient; higher bands are labeled low, moderate, or high by explicit thresholds in the scoring code.

The principles/practice alignment rating compares the conceptual and applied score on every adequately covered shared dimension. It converts their root-mean-square gap to a 0–100 scale, where identical profiles score 100 and a complete pole reversal everywhere scores 0. Squaring each gap before averaging gives a large local divergence more influence, so it cannot vanish among many small differences. A single 0.50 gap caps the result below “Highly aligned,” while a 0.75 gap caps it below “Broadly aligned.” The overall rating is withheld unless at least 12 dimensions have 50% coverage in both registers. Near-midpoint profiles are still rated because moderation or counterbalancing convictions are valid results; they receive a note clarifying that high alignment means the two aggregate profiles are similar, not that either profile is strongly directional.

The bank includes 26 deliberate tradeoff scenarios covering 26 mirrored value-tension groups. Each has a primary axis link and a separate tradeoff link to a competing value. The primary link contributes normally to its designated axis; the tradeoff link does not cross-score the competing axis. Instead, it feeds the tension analyzer.

A tension is reported only when at least two relevant probes were answered. Fewer than three is low confidence, three or four is medium, and five or more is high. “Contradiction” is a rule-based signal for reflection, not a diagnosis of hypocrisy.

Your combined axis scores are compared with 38 named archetypes. Each archetype is a weighted pattern across selected axes; affinity is the weighted average alignment between your scores and that pattern. Several partial matches are expected. The full spread is usually more informative than the winner, and no match changes your underlying scores.

When the optional AI-analysis feature is enabled, a respondent can explicitly consent to send their anonymous result data and any context they choose to add to the configured third-party AI provider. The model receives a structured evidence bundle: deterministic axis scores, coverage, conceptual–applied gaps, tradeoff signals, topic signals, question text, responses, and evidence IDs. It does not calculate, alter, or replace the Polyaxis scores.

Provider output must match a strict versioned schema and pass a second evidence validator. Serious claims require linked question, axis, or tradeoff evidence; stronger labels face higher evidence thresholds. A failed output gets one constrained repair attempt and is rejected if it still cannot be verified. Respondents can optionally answer clarifying questions to create a separately stored refined interpretation without overwriting the original.

The repository includes an administrator validation console that computes item response rates, item–rest discrimination, per-axis Cronbach’s alpha, and inter-axis correlations from stored response sets. Those diagnostics make weak, redundant, or cross-loading items visible for revision. The bank also has versioned migrations and documented content, balance, and conflict-coverage audits.

These facilities are infrastructure for validation, not proof that validation is finished. Reliability estimates are unstable at small samples, correlation does not establish construct validity, and item revisions can change the meaning of a score. That is why submitted results retain their bank version.