Skip to main content

agi4_core/
threshold.rs

1//! Threshold constants from SPEC.md ยง3.
2//!
3//! Every threshold and floor in the specification lives here as a named const.
4//! No magic numbers in evaluation logic. Every value here maps to exactly one
5//! source line in SPEC.md.
6
7pub mod generality {
8    pub const ARC_AGI_2_PASS: f64 = 0.85;
9    pub const ARC_AGI_3_PASS: f64 = 0.50;
10    pub const ARC_AGI_3_FLOOR: f64 = 0.05;
11    pub const HLE_PASS: f64 = 0.80;
12    pub const GPQA_DIAMOND_PASS: f64 = 0.90;
13}
14
15pub mod economic_substitutability {
16    pub const GDPVAL_PASS: f64 = 0.85;
17    pub const RLI_PASS: f64 = 0.60;
18    pub const RLI_FLOOR: f64 = 0.10;
19    pub const APEX_AGENTS_PASS: f64 = 0.75;
20}
21
22pub mod environmental_transfer {
23    pub const ARC_AGI_3_PASS: f64 = 0.50;
24    pub const ARC_AGI_3_FLOOR: f64 = 0.05;
25    pub const OSWORLD_PASS: f64 = 0.85;
26    // NES thresholds: TBD in v0.1.x
27}
28
29pub mod autonomous_agency {
30    pub const METR_80PCT_PASS_HOURS: f64 = 168.0;
31    pub const METR_80PCT_FLOOR_HOURS: f64 = 8.0;
32    pub const REBENCH_PASS: f64 = 0.60;
33    pub const SWEBENCH_VERIFIED_PASS_AT_5: f64 = 0.85;
34}
35
36pub mod consistency {
37    pub const MARGIN_VARIANCE_RATIO: f64 = 0.5;
38}