Let a node carry a surpyval accelerated-failure-time (AFT) model so that "current life" becomes accumulated exposure and the current load sets the go-forward acceleration φ(load).
surpyval's AFT computes H(x | Z) = H₀(φ(Z)·x) — exactly the Nelson cumulative-exposure / time-scale model — with the acceleration factor available directly via model.phi(Z, *phi_params). So the conditional survival under load ℓ from accumulated exposure e is R₀(e + φ(ℓ)·x) / R₀(e).
Builds on the condition-based layer (#36): with an AFT node, R(x | e, load ℓ) = R₀(e + φ(ℓ)·x) / R₀(e) — the same conditional-survival formula as #36, just with the load driving φ.
Estimate
Effort: M (2–4 days) · Value: High
The forward-reliability math is clear because surpyval exposes phi directly; the real risk is AFT serialisation/reconstruction (bullet 2) — if surpyval has no clean reconstruction API this slips to L, so spike that first. Foundation for load-sharing (#38) and warm standby (#41).
Let a node carry a surpyval accelerated-failure-time (AFT) model so that "current life" becomes accumulated exposure and the current load sets the go-forward acceleration
φ(load).surpyval's AFT computes
H(x | Z) = H₀(φ(Z)·x)— exactly the Nelson cumulative-exposure / time-scale model — with the acceleration factor available directly viamodel.phi(Z, *phi_params). So the conditional survival under load ℓ from accumulated exposureeisR₀(e + φ(ℓ)·x) / R₀(e).Builds on the condition-based layer (#36): with an AFT node,
R(x | e, load ℓ) = R₀(e + φ(ℓ)·x) / R₀(e)— the same conditional-survival formula as #36, just with the load drivingφ.Estimate
Effort: M (2–4 days) · Value: High
The forward-reliability math is clear because surpyval exposes
phidirectly; the real risk is AFT serialisation/reconstruction (bullet 2) — if surpyval has no clean reconstruction API this slips to L, so spike that first. Foundation for load-sharing (#38) and warm standby (#41).