The evidence
Four figures from the frozen study — 384 recorded runs, 2,560 hypotheses, eight arms over 48 worlds, heuristic ranker throughout so anyone can regenerate it without credentials. Design reasoning in docs/viz-rationale.md.
loading study…
1 · Cutting measurement costs accuracy — only under independent noise
Held-out direction accuracy per world. Each dot is one world (n=48 per arm); the heavy tick is the arm mean.
Full 0–1 axis: an earlier draft ran 0.4–1.0 and silently clipped two lean-25 worlds at 0.33 — caught by the independent design review, and they are exactly the failure cases the figure exists to show. Strip plot rather than a bar of means because the finding is the variance: lean-25 drops the mean and quadruples the spread — unless the arms share noise, in which case cutting samples is invisibly free and the metric measures nothing.
2 · The confounded worlds refute nearly everything
Hypotheses refuted by the experiment that tested them, by world topology.
Sorted by rate; length on a common scale from a zero baseline. The two confounded topologies were built to punish observational priors, and do.
3 · On the sign-inverting worlds, observation points the wrong way
Population: sign-carrying hypotheses on template T6 across the study. Two parts, so the whole comparison is one boundary — a single 100% bar.
4 · The agent never misses a real edge — and over-claims one in five
Raw counts, zero baseline; precision and recall are annotations, not bars — the ratios are derived, the counts are the evidence. The empty false-negative segment is the recall result.
Live ledger — separate from the study
Totals only, no rates: a live denominator changes under the reader. The figures above never include these runs. Raw data: CSV · JSON · stats · world inspector