Calibration

Which benchmarks predict which

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

For each ordered pair of same-capability benchmarks with at least 5 shared models, several monotone curves are fitted and the one with the lowest leave-one-out error is kept, if it passes the quality gate (R² ≥ 0.3, error ≤ 15 pp).

Calibration	Curve	Models	Error	Estimates
AA-AnalystAgent → AutomationBench	inverse Michaelis–Menten	11	11.2 pp	5
AA-AnalystAgent → EnterpriseOps-Gym	linear	7	1.8 pp	9
AA-AnalystAgent → Harvey LAB	Michaelis–Menten	10	1.8 pp	2
AA-AnalystAgent → τ³-Bench Banking	Michaelis–Menten	7	4.5 pp	5
AutomationBench → AA-AnalystAgent	Hill	11	7.8 pp	14
AutomationBench → EnterpriseOps-Gym	Michaelis–Menten	11	6.2 pp	11
AutomationBench → Harvey LAB	linear	8	3.2 pp	8
AutomationBench → τ³-Bench Banking	offset logistic	15	5.0 pp	8
CritPt → GPQA Diamond	Michaelis–Menten	16	2.3 pp	1
CritPt → Humanity's Last Exam	linear	27	4.1 pp	1
EnterpriseOps-Gym → AA-AnalystAgent	inverse Michaelis–Menten	7	3.6 pp	8
EnterpriseOps-Gym → Harvey LAB	Hill	6	1.2 pp	9
EnterpriseOps-Gym → τ³-Bench Banking	Hill	10	11.4 pp	3
GDP.pdf → MMMU-Pro	Hill	17	2.6 pp	13
GPQA Diamond → CritPt	inverse Michaelis–Menten	16	6.8 pp	3
GPQA Diamond → Humanity's Last Exam	linear	14	4.6 pp	3
Harvey LAB → AA-AnalystAgent	inverse Michaelis–Menten	10	13.9 pp	2
Harvey LAB → AutomationBench	linear	8	13.0 pp	2
Harvey LAB → EnterpriseOps-Gym	Hill	6	4.4 pp	3
Harvey LAB → τ³-Bench Banking	Hill	6	10.4 pp	2
Humanity's Last Exam → CritPt	Hill	27	3.5 pp	3
Humanity's Last Exam → GPQA Diamond	Michaelis–Menten	14	1.3 pp	16
ITBench SRE → Harvey LAB	Hill	5	1.6 pp	11
ITBench SRE → τ³-Bench Banking	linear	8	2.6 pp	10
MMMU-Pro → GDP.pdf	Hill	17	4.0 pp	3
τ³-Bench Banking → AA-AnalystAgent	linear	7	8.1 pp	4
τ³-Bench Banking → AutomationBench	Hill	15	10.6 pp	4
τ³-Bench Banking → EnterpriseOps-Gym	Michaelis–Menten	10	5.3 pp	6
τ³-Bench Banking → Harvey LAB	offset logistic	6	2.1 pp	4
Terminal-Bench 2.1 → Terminal-Bench 4.0	inverse Michaelis–Menten	15	11.6 pp	4
Terminal-Bench 4.0 → Terminal-Bench 2.1	Michaelis–Menten + offset	15	4.0 pp	15
Terminal-Bench 4.0 → Terminal-Bench Science 0.1	Hill	21	6.5 pp	9
Terminal-Bench Science 0.1 → Terminal-Bench 2.1	Hill	10	7.0 pp	2
Terminal-Bench Science 0.1 → Terminal-Bench 4.0	Hill	21	8.2 pp	2