Generation Ancestry DAG (G0 → G1 → G2 → G3)
| Generation ID | Gen | Parent | Library Digest | Inherited Candidates |
|---|---|---|---|---|
GEN-0-BASELINE | G0 | ROOT | edc7d587134c73… | baseline |
GEN-1-UNIFIED | G1 | GEN-0-BASELINE | 83b0032e134b8e… | CAND-G1-L3-MACRO-LIB |
GEN-2-COMPOUND | G2 | GEN-1-UNIFIED | d661214c834a84… | CAND-G1-L3-MACRO-LIB, CAND-G2-L4-CURRICULUM-COMPOUND |
GEN-3-VERIFIED-INHERITANCE | G3 | GEN-2-COMPOUND | d661214c834a84… | CAND-G1-L3-MACRO-LIB, CAND-G2-L4-CURRICULUM-COMPOUND |
Unseen held-out solve rate: G0 0.25 → G3 0.9583 (2x compute-matched control: 0.25). G3 unseen causal MAE: 0.8026 vs random 0.6891 (Wilcoxon p=0.025709910326).
Unified Benchmark Status & Honest Limits
- Track A (AGI): TESTED on 8 constructed capability families; external benchmarks (ARC-AGI-2/3, SWE-bench Verified, METR) remain BLOCKED.
- Track B (RSI): TESTED across Levels 1–5 (G0→G3 + rollback); AI4AI-Bench remains BLOCKED; no GPU or neural weight self-modification claimed.
- Track C (Unified 4-Arm): Arm D 0.7 vs A 0.34, B 0.385, C 0.59 (Holm p={'D_vs_A': 0.0, 'D_vs_B': 8.32e-10, 'D_vs_C': 0.003185253448}).