Methodology / Transparency

Transparency

Metryval is built to be inspected, not gestured at.

This page covers how your result is produced, what happens to your data, the provenance of the methodology, and the use cases the product refuses to serve.

How your result is produced

Your result is deterministic

The scoring engine is a deterministic implementation that runs without invoking any language model.

Given your 12-item Lite response set or your full-instrument response set, the engine produces:

  • A 19-dimension PAREN coordinate (Perception 5, Agency 3, Responsiveness 4, Engagement 4, Navigation 3).
  • A distance to every reference centroid in the launching Vectyr registry, computed via Euclidean distance at launch and via Mahalanobis distance at N≥500 empirical responses.
  • A tier classification (which Vectyr-distance band you fall into) and the Anti-Barnum gate pass/fail verdict.
  • A library-assembly key set that the pre-rendered narrative content is looked up against.

The same response set produces the same result every time.

No prompt is sent to a language model.

No temperature parameter changes the output.

The engine is computed, not generated, and it is open source and inspectable.

How the narrative content is authored

Every Vectyr chapter is authored against a fixed discipline: the framework canon, the voice rules at Methodology / Language constraints, and the Mara Vidal narrator register.

Each draft passes a 5-layer QA stack before it is committed to the narrative library:

  • Banned-vague-qualifier linter (the L1 voice linter).
  • Anti-Barnum content check (no claim renders without a threshold-pass coordinate plausibility).
  • Somatic-specificity verification (body location, texture, and intensity all present).
  • Reference-centroid framing check (no “you ARE X” language; no categorical assignment).
  • Stylometric integrity (sentence-length standard deviation, hapax ratio, proper-noun density).

Drafts that clear all five layers are committed to the narrative-library substrate.

At request time the engine looks up the relevant library entry.

It does not generate prose on demand.

Methodology provenance

The methodology behind the content is documented, versioned, and public.

Three sources ground it.

  • The framework reception ledger. The curated set of synthesized peer-reviewed psychological frameworks, with per-framework academic-reception status, license, and public-domain status. See Framework Reception Ledger for the current set and version history.
  • Voice canon and narrator register. The Mara Vidal narrator register is documented in internal voice-canon files; the public-facing voice rules are at Methodology / Language constraints.
  • Synthetic calibration distribution. The pre-pilot calibration sample is a synthetic distribution constructed from published parameters of Big Five, HEXACO, and Self-Determination Theory measures, projected into the PAREN coordinate system. Sample size N=50,000; provenance documented in the centroid-coverage audit (2026-04-26).

The framework reception ledger is the canonical methodology summary.

It is versioned, dated, and updated whenever the framework set changes (most recently Polyvagal Theory removal on 2026-05-07, after the 39-expert consensus paper).

Personal-data handling

Respondent data is handled under the following constraints.

  • Lite Assessment and full-instrument responses are stored encrypted in Supabase, with row-level security policies that restrict access to service-role queries.
  • Respondent responses are not used to train or fine-tune any model. The narrative-library substrate is authored ahead of time from the framework reception ledger and voice-canon documents; it does not learn from respondent inputs.
  • Respondent data is not sold to third parties, brokered, or transferred outside the platform without explicit consent for a documented purpose.
  • Deletion-on-request is supported. A respondent who requests deletion has their account, response data, and derived artifacts removed within the regulatory window required by their jurisdiction (GDPR for EU respondents, PDPA for Singapore, CCPA for California).

The privacy policy at /privacy documents the full data-handling regime, including retention windows, sub-processor disclosure, and the consent regime that governs each data category.

Synthetic data and cleaning

The synthetic calibration distribution exists because pilot data does not yet.

Constructed from published Big Five, HEXACO, and Self-Determination Theory measure parameters and projected into the 19-dimension PAREN coordinate system, it gives the launch-set Vectyr placement, the Anti-Barnum threshold value, and the tier-classification cutoffs a defensible pre-pilot grounding.

It is a defensible proxy, not the empirical pilot distribution.

At N≥500 empirical responses with adequate dimensional coverage, the Mahalanobis covariance matrix replaces the synthetic Σ in the distance computation. At N≥1,500, the Anti-Barnum threshold value re-calibrates against the empirical standard-deviation distribution. Both transitions are documented in advance, dated, and published in the Framework Reception Ledger when they ship.

Synthetic data is never represented as empirical evidence.

The published claims that depend on it are framed as such throughout, and the substitution to empirical data is part of the documented post-pilot roadmap.

Versioning and ledgers

Every methodology decision that affects the instrument or its output is tracked in one of two public ledgers.

  • The Framework Reception Ledger tracks per-framework academic-reception status as the psychological-measurement literature evolves. Polyvagal Theory was removed on 2026-05-07 after the Grossman et al. consensus paper; that change is dated and rationaled in the ledger.
  • The Policy Decisions Ledger tracks load-bearing decisions about use-case scope and product boundaries. P-01 documents the B2B-selection exclusion (hiring, promotion, performance review, comparative ranking) and is dated and rationaled in three layers.

Both ledgers are also served as machine-readable JSON at fixed URLs, so external researchers, replication auditors, and downstream tooling can fetch the canonical state without scraping the prose:

Both ledgers are versioned; version increments include a change-log entry naming what changed and why.

The JSON artifacts are internally cross-checked at build time against the headline counts they declare, and mismatches fail the build before deploy.

Compliance and use-case limits

Use-case exclusion (EU AI Act Annex III item 4)

The EU AI Act classifies AI systems used for employment, worker management, and access to self-employment as high-risk under Annex III item 4. Metryval is not such a system.

The instrument is voluntary self-administration by individual respondents in a self-reflection context.

Selection use cases (hiring, promotion, performance review, comparative ranking) are structurally excluded from the product scope, the terms of service, and institutional-tier agreements per Policy Decisions Ledger entry P-01. The exclusion is named in any sales conversation, design-partner pitch, advisor outreach, press interview, or product copy where the use case might otherwise be assumed compatible.

No content is generated at request time

Metryval does not generate content in response to a respondent.

The Vectyr chapter content is pre-authored narrative-library substrate, looked up deterministically at request time.

California AB 2013 (training-data disclosure)

Where any generative tooling is used in content production, California AB 2013’s training-data disclosure is satisfied by the public framework reception ledger, which is the documented, versioned methodology provenance described above.

Cross-references

← Back to Methodology