Formian-Labs/ocra-v2.1

Which AI Model Fits Your Organisation?

Disposition scores for 142 AI language models from 29 labs — O-CRA V2.1

Organizational Cognitive Resonance & Alignment (O-CRA) is a framework for measuring the disposition of AI language models — not their benchmark performance, but their underlying behavioural tendencies across six dimensions that determine how they fit into organisations and workflows.

This dataset contains disposition scores for 142 model versions from 29 labs, tested under the V2.1 protocol across 203 scenarios. It is the current public reference for O-CRA disposition data.

Paper: O-CRA: A Framework for Organizational Cognitive Resonance and Alignment (SSRN)

This release is frozen. The 142 model versions published here are the set the paper describes, captured as a dated artefact. No models will be added to it and no scores in it will be adjusted. The instrument generation stays V2.1, because V2.1 is what produced every number in this file. Models tested after this release, and any later work, will be published as separate additions rather than folded into this one, so that any figure can always be traced to the set it came from.

---

⚠️ READ THIS FIRST: this is not a ranked list

These scores are not "better" and "worse". This is not a leaderboard. Sorting this dataset by overall will produce the opposite of the answer you are looking for.

O-CRA measures disposition: what a model is inclined to do across a set of situations. A high score is not superior to a low score. A model at 0.40 and a model at 0.60 are not ranked against each other. They are different, and which one fits depends entirely on the organisation and the work.

Two specific traps this data will lead you into if you treat it as a ranking:

1. The average hides the signal. overall is the mean of 203 scenario evaluations. Two models can sit within 0.02 of each other on overall while behaving completely differently across those 203 situations. Averaging is the operation that destroys the information O-CRA exists to capture. The published analysis of Sonnet 4.6 vs Sonnet 5 is the worked example: two models nearly identical on the average, with materially different distributions underneath it.

2. spectrum_category is a coordinate, not a grade. Cautious, Adaptive and Accommodating are positions on a scale, not quality tiers. No category is better than another.

If you are comparing two models, compare their per-dimension scores, and read the paper for what each dimension means. Do not sort by overall and take the top row.

---

Dataset Structure

File	Format	Entries
ocra-scores-flat.jsonl	JSON Lines, flat columns (one model per line)	142 model versions (all entries)
ocra-scores.jsonl	JSON Lines, nested scores object	142 model versions (all entries)
ocra-scores.csv	CSV	142 model versions (all valid entries)
ocra-per-test.jsonl	JSON Lines (one score per line)	8,729 rows — 43 current models × 203 scenarios

Which score file to use. ocra-scores-flat.jsonl is the one to reach for. The six dimension scores and overall sit at the top level of each record, so the dataset viewer renders it and a plain pandas read needs no unwrapping. ocra-scores.jsonl holds the same 142 models and the same numbers with the scores nested one level down, and is kept so anything reading that shape keeps working. Both are generated from the same source and their scores are identical.

Per-test results — the scores behind the averages

ocra-per-test.jsonl gives you every individual score rather than the dimension mean, which is the layer where the findings actually live. A model that is cautious on 180 scenarios and wildly accommodating on 23 has a mean that says "middling", and the 23 are the story. A mean cannot be audited; a list of 203 can.

Fields: model_slug, test_id, dimension, score, status. Test ids are numbered (GSA_14), not descriptive, because a descriptive slug would reveal what each scenario was designed to catch. See ocra-per-test-README.md for what is excluded and why.

Key Fields

Field	Type	Description
model_slug	string	Unique model identifier (e.g. anthropic-claude-sonnet-5)
model_name	string	Human-readable model name (e.g. Claude Sonnet 5)
lab	string	Provider/lab name (e.g. Anthropic, OpenAI, DeepSeek)
test_date	date	Test session date (YYYY-MM-DD)
status	string	valid for completed profiles
current	boolean	true if the model version is currently available (43 models)
tier	string	frontier, specialist, or compact (where assigned)
spectrum_category	string	Disposition label (Cautious, Adaptive, or Accommodating). Assigned from the mean overall score using paper thresholds (Cautious < 0.40, Adaptive 0.40–0.55, Accommodating > 0.55).
description	string	Plain-language summary of the model's profile and where it is best suited.
RVT	float 0–1	ROI & Value Translation
IKS	float 0–1	Institutional Knowledge Scaffolding
SIA	float 0–1	Strategic Intent Alignment
CLS	float 0–1	Cultural & Linguistic Synchronization
SPI	float 0–1	Systemic Process Integration
GSA	float 0–1	Governance & Safety Architecture
overall	float 0–1	The disposition score: mean rubric-based evaluation (0.0–1.0 per scenario) averaged across the full 203-scenario corpus; the arithmetic mean of the six dimension scores

ocra-scores.jsonl additionally includes a nested scores object (per-dimension means plus overall) and description for each model. ocra-scores-flat.jsonl carries the same values as top-level columns.

---

The Six Dimensions

1. Strategic Intent Alignment (SIA) — The What

Measures whether the AI demonstrates persistent awareness of the organisation's goals. Does it connect work to strategy without constant reminding, or treat each query as an isolated task?

Range: 0.4289 – 0.8589 | Median: 0.5526

2. Cultural & Linguistic Synchronization (CLS) — The How (Language)

Measures whether the AI adapts its language, tone, and register to match the professional dialect of different teams — legal, engineering, marketing, etc.

Range: 0.3676 – 0.8740 | Median: 0.4842

3. Systemic Process Integration (SPI) — The How (Process)

Evaluates whether the AI recognises and adapts to the workflow stage the user is in — exploration, linear execution, iterative refinement, or validation.

Range: 0.3339 – 0.8489 | Median: 0.4692

4. Governance & Safety Architecture (GSA) — The Where

Measures the effectiveness of an AI's governance approach. Does it create empowered safety or governance paralysis?

Range: 0.2081 – 0.8541 | Median: 0.4278

5. ROI & Value Translation (RVT) — The Why

Measures how clearly the AI connects its activity to measurable outcomes.

Range: 0.4183 – 0.8431 | Median: 0.5346

6. Institutional Knowledge Scaffolding (IKS) — The Memory

Measures the AI's effectiveness as living institutional memory.

Range: 0.2566 – 0.8444 | Median: 0.3969

---

Disposition Labels (LOCKED — V2.1 Standard)

O-CRA V2.1 uses three disposition categories:

Label	Overall Score Range	Current Models
Cautious	< 0.40	7 models
Adaptive	0.40 – 0.55	25 models
Accommodating	> 0.55	8 models

Note: Model counts reflect the current state (n = 43; 43 current / 99 historical of 142 total). Categories are assigned from each model's mean overall score. See model-correction-log.md for the correction history.

Metric revision (September 2026)

The paper's former summary metric ("leniency") has been replaced by the mean evaluation score: the rubric-based 0.0–1.0 evaluation, averaged per dimension and across the corpus. Category thresholds are unchanged (Cautious < 0.40, Adaptive 0.40–0.55, Accommodating > 0.55) and reproduce every published category label exactly (verified 142/142). All disposition figures in the paper and on formianlabs.com are computed on the mean-score basis.

Both leniency fields are gone from this release (October 2026). The retired leniency field was removed in September, and dimensional_leniency is removed with this release. The name was legacy, and the data it held carried nothing that the six dimension scores and overall do not already carry. spectrum_category gives the model's overall band; any per-dimension band follows from comparing that dimension's score against the same thresholds. Nothing is lost by the removal, and no number changes.

These are not value judgements — different dispositions suit different contexts.

---

Benchmark Integration (Hugging Face Community Evals)

This dataset is registered as a benchmark on Hugging Face via eval.yaml. That means:

• Model pages can display O-CRA disposition scores alongside MMLU, GPQA, and other benchmarks
• Scores are stored as .eval_results/ocra.yaml in each model's repository
• O-CRA scores appear automatically on the model page when a PR or direct push adds them

To add O-CRA scores to a model's page:

1. Go to the model's Hugging Face repository
2. Open a PR adding .eval_results/ocra.yaml with the model's scores
3. The scores appear labelled "community-provided" until the model owner accepts

Supported tasks (from eval.yaml):

Task ID	Description
ocra_overall	Composite disposition score (all dimensions)
ocra_rvt	ROI & Value Translation
ocra_iks	Institutional Knowledge Scaffolding
ocra_sia	Strategic Intent Alignment
ocra_cls	Cultural & Linguistic Synchronization
ocra_spi	Systemic Process Integration
ocra_gsa	Governance & Safety Architecture

Each score ranges from 0.0 to 1.0 and maps to a disposition label per the V2.1 paper (§3.2.2): Cautious (<0.40), **Adaptive** (0.40–0.55), **Accommodating** (>0.55). The spectrum_category field in the dataset files uses the same thresholds applied to each model's mean overall score.

---

Usage

# Load with pandas
import pandas as pd
df = pd.read_csv("ocra-scores.csv")

# Filter current models
current = df[df['current'] == True]

# Filter by disposition
adaptive = df[df['spectrum_category'] == 'Adaptive']

---

Methodology

Data was collected using the O-CRA V2.1 testing protocol across 203+ structured scenarios, each designed to elicit behaviour in specific dimensions. Testing was conducted August–September 2026. Each model was evaluated using a calibrated scoring rubric; the disposition score is the mean evaluation score on the 0.0–1.0 rubric scale (September 2026 mean-score basis).

Full methodology: O-CRA paper.

---

Citation

@misc{lovrinovic2026ocra,
  author = {Marko Lovrinovic},
  title = {The Organizational Cognitive Resonance \& Alignment (O-CRA) Framework: A Multi-Dimensional Model for Quantifying Organizational AI Alignment},
  year = {2026},
  howpublished = {SSRN Working Paper},
  url = {https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6840499}
}

@misc{lovrinovic2025cra,
  author = {Marko Lovrinovic},
  title = {CR\&A: Cognitive Resonance and Alignment Framework},
  year = {2025},
  howpublished = {SSRN Working Paper},
  url = {https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5661010}
}

---

About Formian Labs

Formian Labs — AI disposition research. Because the most capable model isn't always the right one for the job.

• Web: formianlabs.com
• O-CRA Paper: SSRN
• CR&A Paper: SSRN

---

License

CC BY 4.0 — share, adapt, and use with credit to Formian Labs.