Skip to content
VARZINIndependent research
Menu
Protocol history / June 2026

Core-30 human-rater draft

VPE-001A — Human Rater Study Protocol

Historical Core-30 design draft · Not executed · Registration unverified Date drafted: June 2026

Editorial status note — 15 September 2026. This earlier card-sorting draft is protocol history. It is separate from both the v1.3 N=40 classification design and the later N=59 plan with Control, Affine Priming, and Morphological Priming cohorts. Recruitment for the later plan has not commenced according to the latest available project documentation; no human outcome is established. The YEGJ8/2NKZA registration identity remains unresolved. See the current human-study status.

The original draft used “pre-registered” and made informal power and cognitive-interpretation claims. Those claims are not verified here. Free sorting into participant-defined groups does not by itself define a common category coding for Fleiss κ: an explicit scoring rule and a justified power analysis are still required before execution. The interpretations below cannot establish a hidden cognitive layer or a general human advantage over AI. This annotation preserves the draft for review; the original source remains available.


Research Question

Do naive human raters group Core-30 LUXVAR words by:

  • Morpheme families (ELUZ/SHA/NAR) — as AI systems do (VPE-001B+, ARI=0.812)
  • Semantic axes (LIGHT/REFLECTION/SILENCE/GATE/MOTION) — as designed
  • Something else — neither of the above

All three outcomes are scientifically interpretable.


Participants

Parameter Value
Minimum n 5 raters
Target n 10–15 raters
Inclusion No prior LUXVAR knowledge
Exclusion Any exposure to VARZIN project, conlang research, or this paper
Recruitment Convenience sample — colleagues, online volunteers

Blinding: Raters receive no information about family structure, axes, or frequencies. Only the 30 words are shown.


Materials

The 30 words (Core-30):

ELUZ-MAHAR, SARTHAL, SHARUM-DOZ, XANORAH, MATIR-ELUZ, ANUR,
SHAUREN, RAHMOT-SHA, HALZAR, BANIR-RAHT, SHUKAR-MAH, NAZRUK,
NARAH-SHA, GON-SIL, VARKHAN, SHALOM-HATTAH, NARAH, NEMRUH,
SAR-ELUZ, SAVAR-ELUZ, DARZAN, RAZKAN, ELURIM, RITAM,
ZARTH, ZUN-NAR, KHALTUR, KUZTUR, TARHAN-ELUZ, KHIRZAH

Instructions to raters (verbatim):

"You are looking at 30 words from a constructed language you have never seen before. Please group these words into however many groups feel natural to you (between 2 and 8 groups). There are no right or wrong answers — use whatever feels most natural. You may use sound, spelling, or any other pattern you notice. For each group, give it a name or description in one or two words."

Format: Card sorting (physical or digital). Each word on a separate card/tile.


Scoring

Primary metric: Fleiss κ

κ Interpretation
< 0.20 Slight agreement
0.21–0.40 Fair
0.41–0.60 Moderate — primary threshold
0.61–0.80 Substantial
> 0.80 Almost perfect

Original proposed threshold: κ ≥ 0.40 (registration unverified)

Secondary metrics

Metric Computation
ARI vs true axes adjusted_rand_score(rater_labels, true_axis_labels)
ARI vs morpheme families adjusted_rand_score(rater_labels, family_labels)
NMI vs axes normalized_mutual_info_score(...)
Inter-rater ARI Mean pairwise ARI across all rater pairs
Mean n_groups Average number of groups per rater

Power Analysis

Historical assertions below; no formal power calculation is supplied in this draft.

With n=5 raters and 30 items:

  • Fleiss κ ≥ 0.40 is detectable with reasonable confidence
  • For κ ≥ 0.60, n=5 is sufficient
  • Recommended: n≥10 for publication-grade power

Key Research Outcomes

Result Interpretation
κ ≥ 0.40 + ARI_family > ARI_axes Humans group by morpheme (same as AI)
κ ≥ 0.40 + ARI_axes > ARI_family Axes exist in cognitive layer (not surface)
κ ≥ 0.40 + both ARI low Novel cognitive structure (neither AI nor designer)
κ < 0.40 No consistent human grouping — structure may be idiosyncratic

What This Test Settles

Original proposed interpretations, retained for provenance. These are not established conclusions or validated decision rules.

If humans → morpheme families (like AIs):

Morpheme-family organization is the primary cognitive unit of LUXVAR. Designed axes are a secondary abstraction.

If humans → semantic axes:

Semantic axes exist in a cognitive layer inaccessible to computational analysis. The gap between AI and human structure is real and meaningful.

If κ < 0.40:

LUXVAR's structure is not reliably perceived by naive observers. This is also a valid scientific result.


Timeline

Step Action
1 Recruit ≥5 blind raters
2 Administer card-sort (20–30 min per rater)
3 Score: Fleiss κ, inter-rater ARI, ARI vs axes/families
4 Compare human groupings to AI groupings (VPE-001B+)
5 Write Paper B

Historical protocol draft · VPE-001A · Registration unverified · VARZIN Project Reza Nirouyar · ORCID: 0009-0000-4690-6842