10 domains. 500 held-out tasks. The gap is published.
This bundle is the evidence for the suite. It is hashed, not signed; integrity is checkable, independent witnessing is not claimed.
In-distribution library: 500/500 (100.0%). No-procedure floor: 72/500 (14.4%). Fixed procedure: 112/500 (22.4%). Distribution shift: 5/100 (5.0%).
Gate G3: PARTIAL. 10 domains x 500 tasks and 10/10 in-distribution domains >=80%, but the library is hand-written for those families, the 100-task shift probe scores 5/100, and there is no external verifier. The target is not met.
Per-domain results
| Domain | In-distribution (50) | Shift (10) |
|---|---|---|
| software repair | 50/50 (100.0%) | 5/10 (50.0%) |
| security triage | 50/50 (100.0%) | 0/10 (0.0%) |
| data analysis | 50/50 (100.0%) | 0/10 (0.0%) |
| business decision | 50/50 (100.0%) | 0/10 (0.0%) |
| science calculation | 50/50 (100.0%) | 0/10 (0.0%) |
| planning | 50/50 (100.0%) | 0/10 (0.0%) |
| memory retrieval | 50/50 (100.0%) | 0/10 (0.0%) |
| tool orchestration | 50/50 (100.0%) | 0/10 (0.0%) |
| recovery | 50/50 (100.0%) | 0/10 (0.0%) |
| governance | 50/50 (100.0%) | 0/10 (0.0%) |
Verify
python3 verify_offline.py . python3 verify_shift.py .
Both verifiers import no CAIN code. Expected: 49/49 checks for the 500-task bundle and 12/12 for the shift bundle.
Files
results.json · rows.jsonl · commitment.json · tasks_public.json · results_shift.json · rows_shift.jsonl · tasks_shift_public.json · verify_offline.py · verify_shift.py · SHA256SUMS · MANIFEST.json
What this is not
Constructed tasks with deterministic checkers. The library arm is a hand-written procedure set; 500/500 is a ceiling on the families it was written for. The 5/100 shift result is the generalization gap. No external verifier, no independent witness. Not AGI, not VAGI, not production, not a certification.
Per-domain table on the VAGI page · Daily VAGI evidence ledger