GenLayer Intelligent Contract — Execution Performance Benchmark

How long does it actually take for different types of Intelligent Contracts to finalize on GenLayer?

This repo documents a performance benchmark measuring execution time across 5 contract categories on the GenLayer Studio testnet. All contracts, raw data, and methodology are included.

Results Summary

Test Environment

• Network: GenLayer Studio hosted testnet (studio.genlayer.com)
• Chain ID: 61999
• Total validators: 115 active
• LLM models running simultaneously: 8 different models across 2 providers

Contracts Tested

1. Pure Python (baseline)

File: contracts/01_pure_python.py

Basic computation with no external calls — sorts a list and stores the result. Used to isolate the consensus overhead from any LLM or web latency.

2. LLM with strict_eq

File: contracts/02_llm_strict_eq.py

Single LLM call for sentiment classification. Uses gl.eq_principle.strict_eq — all validators must return byte-identical JSON. Prompt is tightly constrained to three possible values.

3. LLM with prompt_comparative

File: contracts/03_llm_prompt_comparative.py

Open-ended text summarization. Uses gl.eq_principle.prompt_comparative — validators agree if outputs are semantically equivalent. Requires an additional LLM call per validator to evaluate equivalence.

4. Web access only

File: contracts/04_web_only.py

HTTP fetch with no LLM. Fetches live data from Coinbase API and stores response body length. Uses strict_eq since the numeric result should be identical across validators if fetched close in time.

5. Web + LLM combined

File: contracts/05_web_llm_combined.py

Fetches live crypto price from Coinbase API, then passes it to an LLM for a HIGH/MEDIUM/LOW assessment. Chains two non-deterministic operations in a single transaction.

Key Findings

1. Consensus is the bottleneck, not computation

Even pure Python with trivial logic takes ~38 seconds. That's the consensus process — 115 validators independently executing and agreeing — not the code itself. Optimizing contract logic has minimal impact on user-perceived latency.

2. LLM calls add surprisingly little overhead

Adding a single LLM call with strict_eq only added ~6 seconds over baseline. LLM inference happens in parallel across validators and gets partially absorbed into the consensus window.

3. Web access and LLM strict_eq are roughly equivalent

Web-only (~43s) and LLM strict_eq (~44s) came in nearly identical. Both operations have similar latency profiles because the bottleneck is the same — waiting for all validators to complete independently.

4. prompt_comparative costs more than strict_eq

The ~6 second gap between strict_eq and prompt_comparative reflects the additional LLM call validators make to evaluate semantic equivalence. If your output can be constrained to a fixed schema, strict_eq is faster.

5. Chaining web + LLM compounds latency

Web + LLM combined (~62s) is more than the sum of web alone (+5s) and LLM alone (+6s) — suggesting compounding effects when two non-deterministic operations need to converge across 115 validators.

Deployed Contracts (Verifiable on Explorer)

All contracts were deployed on GenLayer Studio testnet and can be verified on the explorer:

Methodology

• Each contract was deployed once and executed 3 times
• Time measured manually from transaction submission to FINALIZED status
• All transactions used default validator selection (no custom configuration)
• Tests run on GenLayer Studio hosted environment in March 2026

Raw data: data/raw_results.csv
Environment details: data/environment.json

Recommendations for Developers

1. Use strict_eq by default — if output can be constrained to fixed JSON schema, saves ~6s vs prompt_comparative
2. Avoid chaining web + LLM — separate into two transactions if latency matters
3. Design for the ~38s floor — that's consensus overhead, no amount of code optimization gets below it
4. Account for variance — individual runs varied 5-15s within the same contract type

Limitations

• 3 runs per contract is a small sample
• Results specific to Studio environment (115 validators, 8 models)
• Local setup or Testnet Bradbury may show different numbers
• Network conditions and validator load affect results

Tested on GenLayer Studio testnet, March 2026. Part of the GenLayer community research initiative.