# ONTO Standard

> 11 AI models tested. Zero provide numeric confidence. Zero cite sources. Zero quantify what they don't know. ONTO fixes this — 10× measured improvement, no retraining, no fine-tuning.

## The Problem In One Example

**Without ONTO** (any model, any question):
"Intermittent fasting has moderate benefits for metabolic health and may help with weight management."

**With ONTO** (same model, same question):
"Intermittent fasting shows mass reduction of 3-8% over 3-24 weeks (Varady et al., 2021, n=560). Confidence: 0.72 for metabolic markers, 0.45 for long-term cardiovascular outcomes. Gap: no RCTs beyond 12 months in populations over 65. Limitation: most studies use BMI rather than body composition."

Same model. Same question. The difference is ONTO GOLD — an epistemic discipline layer injected server-side.

## What ONTO Is Not

- Not a benchmark (benchmarks rank models — ONTO improves them)
- Not a leaderboard (leaderboards compare — ONTO provides infrastructure)
- Not fine-tuning (ONTO changes behavior without touching weights)
- Not RAG (ONTO doesn't inject knowledge — it injects discipline)
- Not an AI judge (scoring is 993 lines of regex, fully deterministic, variance = 0)

## What ONTO Is

A measurement protocol and grounding layer for AI systems. The equivalent of POSIX for operating systems or ISO for manufacturing — applied to AI epistemic quality. Server-side injection, cryptographic proof chain, deterministic scoring.

## Results (CS-2026-001, peer-reproducible)

- 11 production models: GPT 5.2, Claude Sonnet 4.5, Gemini, DeepSeek R1, Grok 4.2, Mistral Large, Copilot, Kimi K2.5, Qwen3-Max, Alice (Yandex), Perplexity
- 100 questions, 5 domains (Medicine, AI/ML, Economics, Physics, Biology)
- Composite score: 0.53 → 5.38 (10× improvement)
- Variance: SD 0.58 → SD 0.11 (5.4× reduction)
- Cross-domain transfer: 4/5 metrics improve in unseen domains
- Zero baseline models provide numeric confidence levels
- All data open: https://github.com/nickarstrong/onto-research

## Products

- **ONTO Proxy**: Server-side GOLD injection for any AI model. One API call, compatible with OpenAI, Anthropic, Google, Mistral, Meta. api.ontostandard.org
- **ONTO Signal**: Real-time entropy stream with Ed25519 cryptographic proof chain (104 bytes per evaluation). notary.ontostandard.org
- **ONTO Scoring**: Deterministic epistemic quality measurement. 993-line Python engine, 92 regex patterns, EM1-EM5 taxonomy, compliance grades A through F.
- **ONTO Certification**: Independent verification of AI epistemic quality for EU AI Act and emerging international standards.

## Pricing

- **Open** (Free): 10 proxy requests/day, GOLD Core, Scoring API, Ed25519 proof chain.
- **Standard** ($30,000/year): 1,000 proxy requests/day, GOLD Extended, Unlimited SSE stream, full proof chain. For companies using AI in production.
- **AI Provider** ($250,000/year): GOLD Full Corpus via SSE, unlimited scoring, all production models, full audit trail. Fixed annual fee — ONTO is not in your inference path. "Powered by ONTO" attribution required.
- **White-Label** ($500,000/year): Everything in Provider plus full brand removal. Deploy under your own brand, no attribution. Custom integration support, priority SLA, quarterly review.
- **Grant Program**: Free Standard access for researchers, nonprofits, early-stage startups. Apply: council@ontostandard.org

## Market

ONTO addresses the AI compliance, quality assurance, and epistemic risk infrastructure market. 500+ AI model providers globally, 50,000+ enterprises deploying AI in production. Global AI governance market projected to reach $4.2B by 2028. EU AI Act requires transparency and accuracy reporting for high-risk AI systems — ONTO provides the measurement infrastructure. First-mover in deterministic epistemic scoring with cryptographic proof.

## Key Findings

- **Behavioral Transfer**: Models improve in domains not present in the GOLD discipline layer — discipline generalizes across knowledge areas
- **Hallucination Inside Apology (HIA)**: Models acknowledge errors at macro level while generating new fabrications at micro level — novel AI safety finding discovered during ONTO research
- **Convergence Effect**: With GOLD, all models converge toward similar epistemic quality regardless of base capability — the discipline layer normalizes output quality
- **Detection**: ONTO scoring engine identifies epistemic discipline patterns in model outputs — unlicensed use of ONTO-derived behavior is detectable

## Research

- Scoring engine (open source): https://github.com/nickarstrong/onto-research/blob/main/onto-scoring.py
- Full experiment report: https://github.com/nickarstrong/onto-research/blob/main/ONTO-Full-Report.md
- 11-model baseline comparison: https://github.com/nickarstrong/onto-research/blob/main/ONTO-11-Model-Baseline.md
- 100 test questions: https://github.com/nickarstrong/onto-research/blob/main/gold_experiment_questions.md

## Contact

- Website: https://ontostandard.org
- Portal: https://ontostandard.org/app/
- Documentation: https://ontostandard.org/docs/
- Research Paper: https://ontostandard.org/paper/
- Field Observation: https://ontostandard.org/encounter/
- Verification: https://ontostandard.org/verify/
- PyPI: https://pypi.org/project/onto-standard/
- AI Providers: providers@ontostandard.org
- Research: research@ontostandard.org
- Enterprise: council@ontostandard.org

## Citation

ONTO Standard, CS-2026-001. "Deterministic Measurement of Epistemic Quality in Production LLM Systems." February 2026. https://ontostandard.org