Formian/ocra-v2.1

O-CRA Model Disposition Scores (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 135 models 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)

---

Dataset Structure

Two formats provided:

File	Format	Entries
ocra-scores.jsonl	JSON Lines (one model per line)	135 models (all entries)
ocra-scores.csv	CSV	135 models (all valid entries)

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 (42 models)
tier	string	frontier, specialist, or compact (where assigned)
spectrum_category	string	Disposition label: Cautious, Adaptive, or Accommodating
leniency	float 0–1	Aggregated leniency score (behavioural permissiveness)
consistency	string	uniform, split, or description of dimensional coherence
tendency	string	Behavioural tendency summary
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	Composite disposition score across all six dimensions

The JSONL file additionally includes dimensional_leniency, description, intended_use, and analysis data for each model — providing per-dimension leniency breakdowns and full behavioural analyses.

---

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: 0.5557

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: 0.3692 – 0.8740 | Median: 0.4910

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: 0.4722

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: 0.4379

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: 0.4006

---

Disposition Labels (LOCKED — V2.1 Standard)

O-CRA V2.1 uses three disposition categories:

Label	Overall Score Range	Current Models
Cautious	< 0.3	4 models
Adaptive	0.3 – 0.6	21 models
Accommodating	> 0.6	17 models

Note: Model counts by spectrum_category reflect the data supplied. Some minor discrepancies exist between data sources — the site build (data/models.json) uses slightly different current flag values for 4 Google models. We recommend using the JSONL source for the most current state.

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: Cautious (<0.3), **Adaptive** (0.3–0.6), **Accommodating** (>0.6).

---

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 in August 2026. Each model was evaluated using a calibrated scoring rubric.

Full methodology: O-CRA paper.

---

Citation

@misc{lovrinovic2025ocra,
  author = {Marko Lovrinovic},
  title = {O-CRA: A Framework for Organizational Cognitive Resonance and Alignment},
  year = {2025},
  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.