Flight 1, Visualized

KB Document Format — 3 arms × 2 models × 7 cases
42 rows · 0 failed
Run 2026-07-11
Footprint cut −47% Structured Markdown vs. prose
Quality cost None 4.84 vs. 4.73 — inside the noise
Hallucinations 0/42 Every arm caught the Q6 trap
Document footprint by arm Characters in the source document. Same facts in all three arms — only the formatting differs. This is the finding the flight actually proved.
Arm A Prose (control) 12,052 Arm B Structured MD 6,426  −47% Arm C Semantic XML 10,261  −15% 0 3k 6k 9k 12k
Quality composite by arm Blind 5-criterion rubric, 1–5, averaged across both models (n=14 per arm). Drawn from zero — the bars really are this close.
Arm A Prose (control) 4.73 Arm B Structured MD 4.84 Arm C Semantic XML 4.83 0 1 2 3 4 4.5 5
Read this as a null result, not a win. 40 of 42 rows scored at or above 4.5 — the dashed line. On content this easy, quality can't separate the arms, so the footprint result above is the one worth trusting.
Where quality actually moved — by criterion All 42 rows, every arm combined. Four criteria sit at the ceiling; only conciseness has any room in it.
Correctness 5.00 Groundedness 4.95 Completeness 4.93 Audience fit 4.81 Conciseness 4.31 0 1 2 3 4 5
The lowest score is a rubric artifact, not a bad answer. Conciseness at 4.31 is the judge penalizing accurate elaboration as "verbose" — a documented failure mode of the framework, visible in the rationales themselves.

Latency and cost are charted nowhere here on purpose. This flight ran through Claude Code's Agent tool, which inflates latency with orchestration time and cannot split input from output tokens — the score output states what that distorts. Plotting those numbers would give them a credibility the run does not support.