What is this?

Quantitative evidence for the YON™ format. Measures structural reliability, cognitive economy, streaming properties, fault isolation, and emitter faithfulness across 58 local suites and 12 LLM suites.

Install

npm install @younndai/yon-benchmarks

Quick Start

# Local suites only (no API keys needed)
npm run bench:local

# Full run (local + LLM suites if keys available)
npm run bench

# LLM suites only
npm run bench:llm

# Single provider
npm run bench -- --provider openai

# Multiple providers
npm run bench -- --provider openai,google

# Filter by suite name
npm run bench -- --filter "generation"

API Key Setup

Copy .env.example to .env.local in this package directory and fill in your provider keys:

# OpenAI — required for most LLM suites (default provider)
OPENAI_API_KEY=sk-proj-...

# Anthropic — used for multi-model comparison suites
ANTHROPIC_API_KEY=sk-ant-api03-...

# Google — used for multi-model comparison suites
GOOGLE_GENERATIVE_AI_API_KEY=AIza...

Missing keys are not errors. Suites that need a missing provider will skip with a message explaining which key to add. Local suites never require API keys.

Which keys unlock which suites?

| Suite | OpenAI | Anthropic | Google | | ------------------------ | ------ | --------- | ------ | | Cognitive Load | ✅ | — | — | | Generation Quality | ✅ | — | — | | Shot Curve | ✅ | — | — | | Information Preservation | ✅ | — | — | | Format Comprehension | ✅ | ✅ | ✅ | | Format Traps | ✅ | ✅ | ✅ | | Density Comparison | ✅ | ✅ | ✅ | | Prompt Compression | ✅ | ✅ | ✅ | | Multi-Model Generation | ✅ | ✅ | ✅ | | Report Enrichment | ✅ | ✅ | ✅ |

Report Enrichment is a post-run analysis/synthesis step over the suite results, not one of the 9 counted LLM suites. Suites marked with a single ✅ default to OpenAI but will fall back to any available provider. Multi-model suites run across all available providers.

CLI Reference

npm run bench [flags]

Flags:
  --local                 Run local suites only (no LLM)
  --llm                   Run LLM suites only (skip local)
  --provider <name>       Restrict LLM to specific provider(s)
                          Values: openai, anthropic, google
                          Comma-separated for multiple: openai,google
  --filter <term>         Run only suites whose name contains <term>
  --report                Force report generation (default for full runs)

Examples

# Quick local check during development
npm run bench:local

# Test with just OpenAI
npm run bench -- --provider openai

# Multi-model comparison (OpenAI + Google)
npm run bench -- --provider openai,google

# Run a specific LLM suite
npm run bench -- --llm --filter "cognitive"

# Full run with all providers
npm run bench

What It Measures

Six Pillars

| Pillar | What it validates | Example suites | | ------------------------ | -------------------------------------------------------------------------------- | ------------------------------------------ | | Streaming | Line-oriented processing, first-record latency | Streaming Properties, Streaming Latency | | Lossless | Zero information loss through format conversions | Format Fidelity, Payload Fidelity, Hedging | | Cognitive Economy | Token efficiency at compressed densities (min/ultra), context window utilization | Token Efficiency, Context Utilization | | Cross-cutting | Structural reliability, error recovery, throughput | Error Recovery, Comparative Throughput | | Emitter Faithfulness | LLMs generate valid YON without fine-tuning | Generation Quality, Multi-Model Validity | | Sapir-Whorf | Whether notation shapes model cognition — comprehension, salience, priming | Notation Alignment, Profile Priming, Value Amplifier |

Suite Breakdown

• 58 local suites — deterministic, no API keys needed
• 12 LLM suites — require API keys, measure AI comprehension and generation

Report Output

Reports are written to reports/<timestamp>/:

reports/
  2026-02-14-13-17/
    summary.md              # Human-readable summary
    summary.json             # Machine-readable results
    enriched-summary.md      # AI-polished version (if LLM available)
    <suite-name>/
      result.json            # Per-suite detailed results
      result.md              # Per-suite human summary

Documentation

• HOW-TO-USE.md — task-oriented usage guide.
• TESTING.md — test strategy and coverage.
• CHANGELOG.md — release history.

The YON Project

YON is an open block format and toolchain.

• Specification — @younndai/yon-spec — the normative YON v2.0 standard.
• Toolchain — YounndAI/yon — parser, generator, runner, converter, examples, benchmarks, domains, ai-relay.
• Editor support — yon-vscode (VS Code Marketplace) · @younndai/yon-textmate (TextMate grammar).

Testing

npm test

Deterministic vitest suites run without API keys. The benchmark suites above (npm run bench) are separate from the unit tests.

About YounndAI

YounndAI™ — You and AI, unified. (pronounced "yoon-dye")

A philosophy of intelligence: building with intention, so humans and machines think together without losing what makes either whole.

License & Attribution

Apache-2.0. © 2026 MARLINK TRADING SRL (YounndAI). See LICENSE and NOTICE.

"YON" and "YounndAI" are trademarks of MARLINK TRADING SRL — see TRADEMARK.md.

Created by Alexandru Mareș.

Website: yon.younndai.com

| | | | ------------- | ------------------------------------------------------- | | Spec | YON v2.0 | | Author | Alexandru Mareș | | Company | MARLINK TRADING SRL · YounndAI™ | | License | Apache 2.0 — © 2026 MARLINK TRADING SRL | | Trademark | YounndAI™ Trademark Guidelines |