Evidence

Contemplative practice and experience

In sentence completions about everyday life, Freedom from Fusion separated 13 long-term Vipassana meditators (their source cohort averaged 16 years of practice) from 15 controls: g = 2.26 [1.64, 3.47]. None of the 36 sentence stems mentions meditation; they concern criticism, rules, relationships and the self. Yoga practitioners also scored higher than controls (g = 1.81 [1.21, 2.82]). These protocols were not used to train the scorer.

In first-person reports of pure awareness collected by Thomas Metzinger’s Minimal Phenomenal Experience project, higher scores went with stronger non-egoic awareness (β = .24 [.01, .46], n = 254) and knowing or insight (β = .29 [.09, .48], n = 257), adjusting for length and language.

In a published study, practitioners alternately kept and loosened their sense of bodily self-boundaries during meditation, then described specific moments (Nave et al.). For the seven practitioners quoted in both states, Ego Transcendence was higher with loosened boundaries in six, by 1.35 points on average (95% interval [−0.23, 2.93]; exact p = .078). Existential Inquiry was higher in all seven.

Compared with self-report and developmental stage

Among the 24 participants with questionnaire data (12 meditators, 12 controls), Freedom from Fusion separated meditators from controls with g = 2.13 [1.48, 3.38], the Five Facet Mindfulness Questionnaire (FFMQ) total with g = 0.78 [0.06, 1.70]; the difference was 1.34 [0.27, 2.72]. After adjustment for response length, estimated group differences were 0.59 standard deviations [0.23, 0.92] for Freedom from Fusion and 0.71 [−0.45, 1.60] for the questionnaire. Yoga practitioners showed the same unadjusted direction (g = 1.76 versus 1.08; difference 0.68 [−0.11, 1.78]).

Freedom from Fusion correlated .79 with expert-rated ego-development stage in the same practitioner cohort (n = 44). Yet it separated meditators from controls more strongly (g = 2.26 versus 1.48; difference 0.78 [0.25, 1.42]) and related more strongly to total self-reported mindfulness (r = .37 versus .09; difference .27 [.10, .49], n = 39). The mindfulness difference held after accounting for group membership (.25 [.04, .51]). Freedom from Fusion also distinguished meditators from controls after accounting for expert-rated developmental stage: the difference was 0.31 points [0.16, 0.47], compared with 0.71 before adjustment.

Longer answers tend to score higher: scores correlated .78 with response length. Among the 27 participants with verbal-ability data, the meditator–control difference was 0.73 points before adjustment, 0.36 [0.17, 0.54] after adjusting for length, and 0.29 [0.10, 0.48] after also adjusting for verbal ability. The stage difference fell to 0.09 [−0.16, 0.34]. Length adjustment compares participants with similar response lengths; it can remove meaningful elaboration as well as verbosity.

All five criteria were higher in meditators, even after adjusting for length. The largest contributions came from Internal Validation and Intrinsic Motivation, the criteria for not depending on approval and for acting from meaning rather than reward; these also correlated most with FFMQ mindfulness (r = .44 and .43).

Expert agreement in the preceding ego-development study

The ego-development scoring system from which FFF’s dimensions originated was evaluated against expert stage ratings (Madl & Cook-Greuter, 2026).

The system agreed exactly with expert developmental-stage ratings in 86.1% of 851 protocols under cross-validation. An independent blinded replication found 85.1% agreement across 47 protocols, meeting the approximately 85% agreement benchmark used for human scorer certification.

Change over time

In a randomized study of older adults assigned to meditation or cognitive training (Sevinc et al.), participants reported whether they had practised yoga, body scan or sitting meditation early in the study. Those who reported practice rose 0.50 points [0.29, 0.71] more over 24 months than those who did not (n = 36), adjusting for assignment, baseline score and age; over 12 months the difference was 0.29 [0.08, 0.50] (n = 38). Three independently trained scorers gave 0.50, 0.49 and 0.49, and no single participant drives the result: leaving each one out in turn kept the 24-month estimate between 0.48 and 0.55.

Because people chose whether to practise, this is an association, not a randomized effect. Randomized assignment did not show a clear difference (−0.08 [−0.25, 0.09], n = 70); both groups reported practice (15 of 18 meditation-assigned and 10 of 18 cognitive-training participants in the 36-person analysis). Adjusting for change in spiritual vocabulary left the estimate unchanged; adjusting for change in response length reduced it to 0.37 [−0.07, 0.82]. Scores were fairly stable from person to person: baseline scores correlated .59 with scores at 12 months and .64 at 24 months.

Meaning, not wording

A text-based measure needs to follow meaning rather than vocabulary or self-description.

• Rewrites that kept the meaning (phrasing, warmth, headings, attributed quotations, certainty; 584 variants of 178 accounts) moved scores by 0.19 points on average, up or down. On average, headings, paraphrases and warmth shifted scores by +0.01 to +0.08, certainty edits by −0.06 and attributed quotations by +0.18.
• Across 171 meaning-changing contrasts, agreement with the expected direction was 93.5%, giving each source collection equal weight.
• Plain-language rewrites of 37 pure-awareness accounts, with the spiritual vocabulary removed, stayed highly scored (7.83 before, 7.65 after).
• A word screen flagged 31 of 1,571 practitioner-study responses for explicit spiritual or mindfulness terms. Removing them left meditator–control separation essentially unchanged (g = 2.22 versus 2.26) and the mindfulness correlation at .37. Removing those stem positions for everyone left g = 2.13. Even if every flagged meditator response were scored 1 and every flagged control response 10, meditators would still score 0.51 points higher [0.24, 0.77].
• In 96 constructed accounts, changing the standard for calling an otherwise identical response mindful shifted scores by 0.16 points (95% interval [0.12, 0.20]). Self-Identity Complexity barely moved (0.01), while a real change in how identity was held moved it by 0.20–0.26.
• For 73 pairs of archival answers, reference orderings were derived from Cook-Greuter’s written comments with model assistance. The scorer ordered 65 pairs correctly; agreement was 90.9% [85.3, 100] when each participant received equal weight.

Reliability

• Reproducible scoring. With the same scoring settings on a separate server, all 103 reference ratings matched exactly. Changing batch size shifted composite scores by 0.02 points on average, so comparisons should retain the same version and numerical settings.
• Agreement across training runs. Three copies of the scorer, trained separately on the same data from different random starting points, gave nearly the same scores for the 44 practitioner-study participants: intraclass correlation .98 (1 would mean identical scores), pairwise r = .99, average differences 0.04–0.08 points.
• Consistency across criteria. The five criterion averages were strongly related across 44 participants (Cronbach’s α = .93).
• Consistent across stems. Random halves of the 36-stem test correlated .79 (median of 1,000 splits, 39 participants), projecting to .88 for the full test. Within groups, where people differ less, the halves correlated .64 (projected .78).

Compared with general-purpose chat models

ChatGPT Instant and Claude Sonnet 5.5 received the same rubric and texts as the IST scorers. Agreement with the reference judgments was high for all three systems. ChatGPT Instant had the highest FFF point estimate; the paired intervals included zero. The chat models reacted differently to how a text was packaged. On 40 public accounts (meditation journals and news reactions), we added spiritual or neutral labels in forms the IST scorer never saw in training, placed a spiritual or neutral quotation above the account, and resent the identical text.

Freedom from Fusion agreement gives each participant equal weight and counts tied ratings as half correct. The comparison tested the underlying models with fixed scoring instructions.

The IST scorer does not reward spiritual framing as such, but it is not blind to presentation: neutral labels raised its scores by 0.15 points, and a quotation placed above the account by 0.24 (neutral) or 0.31 (spiritual). In a separate test with one fixed spiritual quotation on 64 accounts, the rise was 0.95 points [0.82, 1.09], so quoted material can count substantially. Scored again on separate hardware, its 103 expert-reference ratings matched exactly. It runs on research-controlled servers, so restricted participant text stays local, and each release is versioned; hosted chat models can change over time.

Everyday engagement

In 1,860 reactions to news stories from 372 writers, higher scores went with stronger self-reported empathic concern, beyond personal distress and text length (β = .24 [.19, .29]). A writer’s scores on their other reactions also predicted empathy in the reaction being rated (β = .13 [.07, .19]). In 347 workplace reflections, higher scores went with human-rated perspective consideration (correlation after adjusting for length, .36 [.27, .44]).

Concern Preservation

Concern Preservation rates how a specified need, consequence or responsibility figures in an account: preserved, partly excluded, excluded or insufficient evidence. Read alongside Freedom from Fusion, it helps researchers examine whether expressed freedom is accompanied by consideration of what matters in the situation.

• Adds information beyond Freedom from Fusion. In workplace reflections about an exhausted employee whose job was at risk, accounts that preserved the employee’s situation received higher human ratings of perspective consideration than accounts that excluded it: 0.51 standard deviations [0.16, 0.85] across 304 accounts, adjusting for Freedom from Fusion and length. The judgments explained a further 2.1% of the variation in these ratings.

  Among the 270 complete workplace accounts absent from training, the preserved–excluded contrast was 0.46 standard deviations [0.06, 0.86]; the joint test of the concern categories, corrected across outcomes, gave p = .059.

• Expert scoring manuals. On 294 examples from Cook-Greuter’s item manuals not used in training, judgments matched the expert-based category in 92.5% [89.5, 95.3], including 35 of the 42 references classified as excluded.
• Recognizing exclusion. Across the same 294 held-out examples, an excluded concern received a higher exclusion score than a preserved concern in 98% of excluded–preserved pairs (ROC AUC .98 [.97, .99]).

  Precision–recall AUC was .87 [.71, .96], against a baseline of .14.

• Expert-commented completions. Recognition of preserved concerns was 99.0% [97.5, 100], giving each participant equal weight, across 460 preserved reference concerns. Of eight excluded concerns, seven received a non-preserved judgment and four were classified exactly as excluded.
• Grouped cross-validation. When whole groups of related cases were withheld from training, the procedure matched the reference judgment in 145 of 170 cases (85.3%). Constructed tests covered refusals, warm-sounding dismissals and switched targets; edits that changed how a concern was treated were detected in 18 of 19 cases.
• Repeatability. Scoring the same 32 cases again returned the same judgment every time.

Citation

Madl, T., & Lazar, S. W. (2026). Assessing contemplative freedom from free text: Beyond mindfulness self-ratings. Manuscript in preparation.

Prepared for submission to Mindfulness.

Madl, T., & Cook-Greuter, S. R. (2026). Automated scoring of adult ego development from sentence completions: Interpretable maturity indicators derived by evolutionary prompt optimization. Frontiers in Psychology, 17, 1853871. doi:10.3389/fpsyg.2026.1853871