VERIFICATION_V09.md
Streaming Polyphase Factor Decomposition / SKAT
VERIFICATION_V09.md
1 # v0.9 verification
2
3 ## Cumulative regression tests
4 - `test_streaming_skat.py`: PASS (6/6)
5 - `test_approximate_lightcone.py`: PASS (5/5)
6 - `test_budget_scheduler.py`: PASS (4/4)
7 - `test_spectral_tail_bridge.py`: PASS (5/5)
8 - `test_residual_trace_cap.py`: PASS (5/5)
9 - `test_residual_completion.py`: PASS (3/3)
10 - `test_realistic_skat.py`: PASS (3/3)
11
12 ## Realistic simulation battery
13 - scenarios: 4
14 - maximum streamed-kernel relative error: 6.611e-16
15 - maximum phenotype-path relative discrepancy: 8.318e-16
16 - deepest tested log10 probability: -457.973916
17 - worst absolute deep-tail discrepancy: 3.209e-12 decades
18
19 ## Historical correction
20 The historical copy-LD simulator used `np.empty` followed by `+=`.
21 v0.9 repairs this with zero initialization. Previous copy-LD benchmark
22 numbers are not authoritative validation.
23
24 ## Boundary
25 Gaussian SKAT-style quadratic forms under synthetic rare-variant LD models
26 are validated here. External cohort replication, finite-sample logistic
27 calibration, and native-GPU performance remain separate.
28
29 ## Numerical warning boundary
30 Deepest adaptive queries emitted overflow warnings in an intermediate rejected
31 certificate-growth calculation. Accepted results had finite certificates and
32 matched the full-spectrum reference. This is not a demonstrated accuracy
33 failure, but the internal bound arithmetic should be made log-domain in a
34 future hardening pass.