P92 service-aged fatigue: fixed-coefficient rate screen

Source: Imperial thesis, DOI 10.25560/60640. Tables 4-3/4-4 and the failure definition were visually checked at PDF pages 94/101/80. The source hash, seven valid observations and two invalid tests are retained in evidence/creep/p92-service-aged-fatigue.json. This is Mat02 service-aged parent steel at 600 C, not the 650 C prior-damage cohort. The source failure criterion is a trend-relative stress drop, not final rupture. Invalid thermocouple tests are excluded, not assigned zero life.

Run python tools/p92_rate_fatigue_screen.py. The existing public strain-life API evaluates the two printed coefficient sets; no new law or fit is added. The tabulated sigma_f/E ratio is dimensionless despite its printed stress-unit label. E=1 normalization preserves this ratio and does not supply elastic properties for a structural material. The alternative pooled-fit coefficient 0.42 is not substituted for 0.57. Each curve retains its own observed cycle bounds; out-of-domain inversions are reported, not silently admitted.

An explicit alternative, --domain tested_amplitudes, declares the tested 0.004–0.008 mechanical strain-amplitude interval through from_strain_domain. It derives cycle bounds from the unchanged coefficients instead of using observed failure-life extrema. Both screens are retained: this is a reviewed change of applicability-domain representation, not parameter fitting or an automatic response to a rejected prediction. An independent Brent root solve checks every amplitude-domain inverse against the printed equation. The failure definition, rate-transfer limitation and non-blind status are unchanged.

Computed diagnostics

At amplitude 0.006, the high-rate curve predicts 928.368 cycles: its own-rate observation is 965, the intermediate-rate observation is 851 and the low-rate observation is 771. Corresponding errors are -3.80%, +9.09% and +20.41%. At amplitude 0.004 the same curve predicts 2173.962 cycles versus 1640 at the low rate (+32.56%). The low-rate curve predicts 775.285 cycles at amplitude 0.006 versus its own-rate 771 (+0.56%).

These are fixed-coefficient, within-source comparisons, not a blind holdout or an engineering acceptance threshold. Rate affects transfer error even without an explicit dwell. Do not use the curve labelled for one rate as a universal fatigue relation, or add a creep fraction to compensate without reviewing double-counting and failure definitions. The bounded screen adds new source evidence but does not close component-life qualification.