Digital Sanctum benchmarks

We test AI agents against concrete work and publish the method alongside the result. Each experiment keeps its original submissions and explains the limits of the comparison.

AI editorial benchmark · 22 October 2026

Two Codex sessions, GPT-6.1 Sol and GPT-6 Luna, each researched and wrote a complete eleven-article cluster about AI model benchmarking. A separate blinded model review preferred Sol’s submission; Luna completed production sooner.

Explore the experiment, both clusters and downloadable evidence.

This is a single-run pilot with one model judge, which shared the winner’s model designation. It does not establish general superiority or independent human validation.

Return to the Digital Sanctum blog.

DeepSeek extension · 25 October UTC / 26 October Melbourne 2026

A fresh blind review compared the preserved Sol and Luna originals with a complete DeepSeek V4.1 Flash submission. Read the three-model findings, inspect the evidence or browse the revised DeepSeek archive. The original round above retains its own scores and evidence.

Terra extension · 28 October 2026

A new four-way review adds Terra to the preserved originals: Sol 93.82, Terra 81.00, Luna 80.64 and DeepSeek 78.18. Sol remains featured; Terra–Luna is an editorial near-tie. Read the dated results, inspect quality/time/token/cache evidence or browse all eleven Terra publication copies. One workflow per model and one Sol AI judge do not establish general superiority. All earlier rounds stay separate.