Four-model editorial benchmark: 28 October 2026 evidence
On this page
The 28 October four-way review preserves the earlier two-way and three-way rounds. Sol remains the preferred submitted cluster; Terra’s 81.00 is only 0.36 points above Luna’s 80.64, an editorial near-tie. The single AI judge shares Sol’s model family. These are advisory scores for frozen submissions, not general model superiority, statistical significance or human publication acceptance.
Read the complete four-model results, methods and limitations.
Read the articles
- Featured Sol practical guide and series
- Eleven edited Terra publication copies
- Luna comparison archive and original-round history
- DeepSeek revised comparison archive
Original evidence and separate publication edits
- Terra’s original 29-file submission ZIP and original file hashes
- Original production brief and unchanged rubric
- Four-way cluster scores and per-article scores
- Blind quality review, fact-check ledger and editing plan
- Independent pre-comparison score freeze, twelve directed checks, reviewer validation and read audit
- Coordinator integrity and arithmetic checks, identity and production usage and separate judge overhead
- Publication edit log, export notes and public export checksums
The original ZIP is byte-identical by entry to the frozen submission. Curated review exports omit private runtime locations and session identifiers and have their own hashes. A source hash establishes identity, not correctness or publication approval. The original external quality-freeze digest identifies the coordinator’s private record; the curated freeze export is separately hashed. The original scored drafts and historical downloads are preserved.
Methods and history
How the experiment works explains the unchanged brief, editorial rubric and limitations. The new report discloses one production workflow per model, one AI judge, attempted blinding, twelve same-judge internal order checks, later research dates and differences in harness, tools, context, supervision and token reporting. No new cash-cost estimate is offered without comparable verified rates. The judge’s 18,467,115 tokens and 2,523.402688 seconds remain separate from production rows.
Original two-way evidence · Three-way evidence · All benchmark rounds.