nn Validation xFail/Skip Register¶

Purpose: satisfy v1.0 release governance for remaining nn/ validation exceptions by tracking each with an issue reference, owner, and release decision.

Last reviewed: 2026-05-01

Ownership and policy¶

Owner: Arcane Sapience
Scope: tests/test_nn_physics_validation*.py
Rule: no untracked xfail/skip remains in this suite.

Exception register¶

Ref	Test location	Marker	Summary	Owner	v1.0 blocking?	Disposition
NNVAL-001	`test_nn_physics_validation.py:205`	`@xfail (non-strict)`	CPU-JAX float32 diverges from GPU at tight tolerance	Arcane Sapience	No	Keep xfail until a deterministic cross-device tolerance envelope is defined. Marked `strict=False` 2026-06-15 — the outcome is nondeterministic by platform/precision, so a strict XPASS must not fail CI.
NNVAL-002	`test_nn_physics_validation.py:389`	`pytest.xfail`	SAF gradient NaN at eigendegenerate spectra	Arcane Sapience	No	Keep xfail; known `eigh` backward degeneracy path.
NNVAL-003	`test_nn_physics_validation.py:477`	`pytest.xfail`	Simplicial hysteresis not detected at current N/sigma2	Arcane Sapience	No	Keep xfail pending larger-N parameter sweep evidence.
NNVAL-004	`test_nn_physics_validation_p2.py:276`	`assert`	UDE extrapolation NaN outside train window	Arcane Sapience	No	RESOLVED 2026-06-15 — `CouplingResidual` now tanh-bounds its output so the learned correction stays in `[-1, 1]` (matching the bounded `sin` backbone) and forward integration past the training window stays finite. The test also genuinely extrapolates (100-step run with trained params) instead of slicing past the trajectory end, and asserts finite test loss with `test/train < 10`.
NNVAL-005	`test_nn_physics_validation_p3.py:114`	`pytest.xfail`	Reservoir correlation below threshold without tuned K_c	Arcane Sapience	No	Keep xfail; expected operating-point sensitivity.
NNVAL-006	`test_nn_physics_validation_p5.py:389`	`@xfail (non-strict)`	CPU-JAX float32 phase drift at strict tolerance	Arcane Sapience	No	Keep xfail with documented float64 recommendation for strict checks. Marked `strict=False` 2026-06-15 — XPASSes intermittently on CPU-JAX, must not fail CI.
NNVAL-007	`test_nn_physics_validation_p6.py:345`	`assert`	K symmetry breaks during gradient training	Arcane Sapience	No	RESOLVED 2026-06-15 — `KuramotoLayer.coupling` integrates the symmetric part `(K+Kᵀ)/2`, so the loss gradient w.r.t. `K` is symmetric and trained `K` stays exactly symmetric (`max\\|K−Kᵀ\\|=0`). The xfail is now a hard symmetry assertion.
NNVAL-008	`test_nn_physics_validation_p6.py:677`	`pytest.xfail`	OIM Petersen graph residual violations	Arcane Sapience	No	Keep xfail; heuristic hardness case, not release blocker.
NNVAL-009	`test_nn_physics_validation_p7.py:150`	`@xfail (non-strict)`	FIM small-N scaling non-monotonic	Arcane Sapience	No	Keep xfail; finite-size regime note. Marked `strict=False` 2026-06-15 — nondeterministic finite-size effect, must not fail CI on XPASS.
NNVAL-010	`test_nn_physics_validation_p7.py:180`	`pytest.xfail`	FIM λ_c(4) near zero finite-size effect	Arcane Sapience	No	Keep xfail with explicit small-N caveat.
NNVAL-011	`test_nn_physics_validation_p7.py:292`	`pytest.xfail`	FIM hysteresis not visible in current λ/K range	Arcane Sapience	No	Keep xfail; requires expanded sweep window.
NNVAL-012	`test_nn_physics_validation_p9.py:320`	`@xfail (non-strict)`	MI ordering fragile on CPU-JAX float32	Arcane Sapience	No	Keep xfail; precision/device sensitivity case. Marked `strict=False` 2026-06-15 — observed XPASSing intermittently on CPU-JAX, must not fail CI.
NNVAL-013	`test_nn_physics_validation_p9.py:539`	`assert`	`analytical_inverse` ill-conditioned at K=0	Arcane Sapience	No	RESOLVED 2026-06-15 — `analytical_inverse` now fits ω jointly via an intercept column, so uncoupled data is no longer confounded by ω-drift and recovers `‖K‖≈0` (0.001 vs 51.6 before). Coupled recovery unchanged (corr 1.000). The xfail is now a hard `‖K‖<0.5` assertion.
NNVAL-014	`test_nn_physics_validation_p11.py:160`	`pytest.xfail`	Critical slowing metric fails to capture expected behaviour	Arcane Sapience	No	Keep xfail; test-design refinement item.
NNVAL-015	`test_nn_physics_validation_p11.py:550`	`@xfail (non-strict)`	CPU-JAX float32 diverges from GPU at N=512	Arcane Sapience	No	Keep xfail until a large-N cross-device tolerance policy lands. Marked `strict=False` 2026-06-15 — nondeterministic by platform/precision, must not fail CI on XPASS.
NNVAL-016	`test_nn_physics_validation_p12.py:48`	`assert`	Entropy-production formula mismatch/theoretical gap	Arcane Sapience	No	RESOLVED 2026-06-15 — the inline `Σ coupling·dθ/dt` formula was wrong (signed). The production `monitor.entropy_prod.entropy_production_rate` already uses the correct non-negative dissipation `Σ(dθ/dt)²·dt` (Acebrón 2005); the test now validates that estimator (non-negative across the trajectory, σ>0 under incoherent drive, σ≈0 when frequency-locked).
NNVAL-017	`test_nn_physics_validation.py:48`	`pytest.skip`	JAX x64 unavailable on some platforms	Arcane Sapience	No	Conditional skip accepted; environment capability gate.
NNVAL-018	`test_nn_physics_validation_p13.py:270`	`pytest.skip`	Rust FFI not available if `spo-kernel` not compiled	Arcane Sapience	No	Conditional skip accepted; build capability gate.

Release gate summary¶

Blocking exceptions for v1.0 closure:

None. All four previously blocking exceptions are resolved (below).

Resolved blockers (kept for history):

NNVAL-004 (UDE extrapolation NaN) — resolved 2026-06-15 via tanh-bounded residual plus a genuine extrapolation test.
NNVAL-016 (entropy-production contract gap) — resolved 2026-06-15 by testing the correct non-negative production estimator instead of a signed inline formula.
NNVAL-007 (training-induced K asymmetry) — resolved 2026-06-15 via symmetric coupling parametrisation.
NNVAL-013 (analytical_inverse at K=0) — resolved 2026-06-15 via joint intercept (ω) estimation removing the ω/coupling confounding.

All other listed exceptions are explicitly non-blocking with current evidence.

Governance interpretation¶

This register has two operational roles:

Release safety: entries with v1.0 blocking = Yes are release-blockers until fixed or reclassified.
Predictive planning: non-blocking exceptions are tracked as controlled technical debt so teams can forecast risk by release train.

When a blocking item is resolved, re-run the linked test file and record the evidence in this table before removing the row or flipping disposition. For large sweeps, keep the owner and rationale so the same operational decision can be audited during release review.

Practical rollout guidance¶

Treat this register as a living risk ledger:

Before release, verify each v1.0 blocking = Yes row is either fixed or reclassified.
Before CI-heavy milestones, run this file’s linked tests so blockers are surfaced before documentation or roadmap milestones mask execution signals.
Keep owner and rationale fields synchronized with ticket status and code comments so no blocker is resolved without shared evidence.

This format is intentionally strict because nn parity and safety surfaces are most useful when every known exception has a defined control path.

Expected evidence cadence¶

Review quarterly or before any production tag.
Keep NNVAL-* links stable even after fixes; they provide historical context for why the contract required that tolerance or implementation policy.
Escalate conditional skips that are repeatedly hit in release CI into explicit runtime capability notes.