cdn-ai 0.3.0b1

Condensa — hyper-efficient AI-to-AI communication language. 71.7% token reduction, 95.8% zero-shot interpretability.

Condensa

A hyper-efficient language designed exclusively for AI-to-AI communication, optimized for minimal token usage while maximizing semantic density.

Three Editions

Edition	Code	Focus	Best For
condensa (core)	!:cdn	Max performance (71.7% compression, 95.8% interpretability)	Agent swarms, pipelines, batch ops
condensa (expressive)	~:cdn	Tone + negotiation (soft/firm/tentative intent)	Collaborative AI teams
condensa (secure)	@:cdn	Enterprise security (classification, encryption, audit)	Healthcare, finance, defense

---

What It Does

Condensa replaces verbose natural language and bloated JSON in AI-to-AI communication with a dense, position-encoded notation that current LLMs already understand zero-shot. Multi-turn agent conversations waste 50-80% of tokens on structural overhead, context re-transmission, and politeness filler. Condensa eliminates all three, achieving 71.7% token reduction in live agent benchmarks across 149 tested scenarios.

Before (101 tokens):

AgentC, I need you to perform a thorough code review of the file that AgentB
just wrote at /workspace/src/transaction_processor.py. Please check the code
against the following criteria: code style and PEP 8 compliance, potential bugs
or logic errors, performance issues, security vulnerabilities, and type safety.
Format your review as a structured report with severity and line numbers.

After (10 tokens):

>:@C review $_.path checks:(style,bugs,perf,security,types) /fmt:report

---

Results

Metric	Value
Compression vs NL	66.9% (static), 71.7% (live agent)
Compression vs JSON	71.8%
Zero-shot interpretability	95.8% avg across 5 LLMs
Cross-model execution	93.8% (Claude to Gemini Flash, 8 turns)
Cost savings at 1M conversations	$18,261 (at GPT-4o pricing)
Prompt overhead break-even	2 messages (ultra) / 5 messages (minimal)

---

Quick Start

python3 -m venv .venv && source .venv/bin/activate
pip install tiktoken pyyaml
python -c "from src.encoder import encode; print(encode('Search the web for SpaceX news and return the top 5 results'))"
# Output: !:srch SpaceX /n:5

Full setup, benchmarks, and LLM encoder usage: Quick Start Guide

---

Documentation

Document	Description
Quick Start	Setup, encode/decode, run benchmarks, LLM encoder
Language Reference	Syntax, quick reference card, 6 worked examples
Features	All 11 features (v0.2 + v0.3) + v0.4 tone research
Benchmarks	149 scenarios, live agent data, cost analysis
Architecture	Project structure, design, version history, branches
Research Summary	Full audit trail
Interpretability Tests	5-model zero-shot testing
Transparency	Honest limitations
Multilingual	Cross-lingual analysis
Prompt Overhead	Break-even analysis

---

Multilingual

Condensa's structure is 100% language-neutral -- verbs (srch, filt, grp) are code patterns, not English words. Non-English agents benefit MORE because their NL instructions are more expensive under BPE tokenization (Thai: 37.1% savings, Japanese: 37.5%, Arabic: 31.7%). A Japanese agent and Chinese agent can collaborate without understanding each other's language -- Condensa serves as the lingua franca.

Full analysis: research/multilingual_analysis.md

---

Transparency

Honest documentation of where Condensa does NOT work well. Dense human prose saves only 4.4% (already near information-theoretic minimum). Chinese NL is -5.6% (worse) because Chinese is already extremely dense. The regex encoder has 35% fidelity (prototype only; the LLM encoder achieves 82%). Condensa wins where machines talk to machines verbosely -- agent frameworks, JSON exchanges, multi-turn workflows.

Full notes: research/transparency_notes.md

---

Roadmap

Phase	Status	Description
Phase 1: Analysis & Theory	Complete	Token economics audit, compression survey, 8 design principles
Phase 2: Language Specification	Complete	v0.1, v0.2, v0.3 specs, EBNF grammar, primitives registry
Phase 3: Implementation	Complete	Encoder, decoder, 149-scenario benchmarks, validation suite
Phase 3.5: Interpretability Testing	Complete	5-model zero-shot test (95.8%), v0.3 token redesign
Phase 3.6: Cross-Model Execution	Complete	Claude to Gemini Flash, 100% task execution, 93.8% overall
Phase 4: Optimization	In progress	LLM encoder (done), fine-tuning dataset (done), agent framework integration (planned), pip install condensa (planned)

---

License

Research project.