Research hub:
recomputable data, defects included

Research assets

Research archive (citable essays)

The full experiment write-ups behind the boards, published on the blog — each with sample sizes, dates and a method section.

• Which model clones your novel's style — The nine-model fantasy arena, in full
• Three ways LLMs fail at long fiction — The failure-mode taxonomy essay
• Can AI imitate an author's style? — 2024-2026 literature × our experiments
• Benchmarked by our own benchmark — The 46.5%→89.9% cache-fix story
• Kimi K3, tested within 48 hours — The incremental-eval template

Annual data report

Not published yet, and we won't fake a placeholder. The raw material is accumulating automatically: the model price table archives a dated snapshot every day (since 2026-07), and every benchmark's chains and verdicts land in the research archive. Once a full year is banked, the first report appears on this page.

How we do research

• Same-condition setups: in comparative benchmarks every side runs the same model, the same key, byte-identical scripts, and every AI call passes through one metering gateway. Numbers collected under different conditions never share a table.
• Rankings come from blind review: ordering conclusions use double-blind judging with two shuffled mappings to cancel position bias; verdicts that flip between orders are discarded, not averaged.
• Judges take an exam first: AI reviewers need ≥80% on known-answer samples to earn a vote. We've published our own study of judges being unanimously wrong — so key conclusions either get human final reads or are labeled as leanings.
• Structural metrics are gates, not verdicts: sentence-length, dialogue-rate and repetition detectors catch regressions and cross-check the blind review; “does it read human” is never delegated to statistics.
• Our own defects go on the record: when a benchmark catches our app failing (the cache-hit embarrassment, the instruction-wording bug), we publish it, with the fix trail in the open.
• Recomputable: full 20-round chains, blind-pack mapping keys, judge verdicts and per-call metering logs are archived; data pages ship a downloadable JSON snapshot.