Public scorecard

What the engine has been checked against

This page separates recorded predictions from outcome-known worked examples and numerical checks. A ledger entry alone does not establish independence, applicability or physical validation. No experimentally validated service domain is established by this page.

ledger at data/predictions.json

Ledger entries

Logging alone does not establish an independent prediction

Recorded scored decisions

0 open; 0 withheld; 0 unrecognised. No accuracy rate asserted.

Rules with no external check

and 1 checked on one path only

Runs of a published case, worked

read before they were written, so not scored

Prediction log

No prediction entries are logged. There is no accuracy rate or blind-prediction result to report. The worked example below is outcome-known and is not counted as a prediction. Campaign readiness and the logging/scoring procedure still require review.

What the engine stands on, and what it does not

A rule that admits no external check says so here rather than passing in silence. An external check is a closed form the rule must collapse onto, a published worked example that states both inputs and answer, or a limiting case whose answer a reader can verify without trusting us. These are numerical verification categories, not evidence that the model predicts real specimens. The displayed register is not a comprehensive audit of every current code path.

No external check

R00 selection, a house rule about which row answers a case
R05 step stress of a graded stack, whose optimiser is internal and not yet published
R06 the Phi criterion, which is ours rather than a standard
R19 weakest interface, a house rule about reporting
R21 measure next, a ranking with no ground truth outside a scored case
R23 process window, a sweep of rules already checked individually

Checked on one path only

R18 default path, a tangent law carrying a lambda fitted under the sine law

Arithmetic checked, scope not

R04 mismatch stress, which assumes no bending; a free bilayer of comparable thickness curves
R15 the barrier probability, which draws four correlated quantities independently
mixture bounds which assume statistical isotropy and do not describe a columnar film

Worked example, not scored

Niobium diffusion bonded to sapphire, Korn and Cannon 2002

Three specimens from one laboratory, one geometry and one process. Two survived preparation and were tested; the sulphur-doped one broke at the interface during sectioning. The engine is run on each and the observation is printed beside the verdict without any rule reading it. The third has no measured toughness to run against and the engine withholds rather than guesses, which is the coverage number and not the accuracy one. This is not a prediction: the papers were read before the case was written, so it does not enter the log above. Sources: korn2002a, cannon2002b.

Run	Specimen	Load share	At the interface	G	G_c	Margin	Verdict	Observed
1	undoped O21, bonded 1300 C for 1 h	0.747	56 MPa	25.2 J / m2	1 167 J / m2	64.7	pass	survived preparation and was tested to fracture
2	undoped O11, bonded 1400 C for 3 h	0.747	74 MPa	44 J / m2	115 J / m2	3.7	pass	survived preparation and was tested to fracture
3	sulphur-doped to 300 atomic ppm, bonded 1300 C for 1 h	0.747	56 MPa	25.2 J / m2	not measured	withheld	indeterminate	fractured at the interface during the cutting operation

Load share between the equal-sharing and thin-layer limits, so neither body carries the mismatch alone.

What the engine says it cannot see

Weakest interface: Niobium / Sapphire, a-plane by fracture, margin 64.74. Observed in the source: survived preparation and was tested to fracture. Meets the static fracture criterion at this interface under cooling from the bonding temperature. Static fracture at this interface is the only mechanism assessed: no time-dependent process, no cyclic loading, no chemical change at the interface and no competing mechanism has been considered, so this is not a statement about the joint. Z 0.5 is read from Hutchinson and Suo Fig. 43, drawn for an elastically homogeneous film on a semi-infinite substrate, and this case meets neither part of that: the pair is not elastically homogeneous, Dundurs alpha 0.53; the substrate is not semi-infinite, film to substrate thickness 1.00. The driving force is therefore computed with a Z outside the range its figure covers, by an amount nobody here has bounded, and the mismatch-dependent solution of their Fig. 53 would settle it. The modulus and Poisson ratio are still plane-unspecified sapphire from Lipkin Table 2 p.4837, reached through the generic path and reported one grade down with the substitution named. The expansion is no longer: this record carries the a-plane index and reaches the corundum crystal, so R25 computes both in-plane coefficients from Merz's axial measurements. The elastic tensor that would replace the modulus is loaded and waiting for the direction-dependent rule. Bending is not modelled. Curvature redistributes stress and can increase its magnitude at some locations; these values are not a general upper bound for a free joint. A constrained joint requires boundary conditions consistent with this membrane model. No free-edge or interfacial traction field is computed. The default mixity law is the tangent form at a lambda of 0.3, supported by Evans 2001's own mode I and mode II pair for a segregated TGO interface, 12.5 and 60 J/m2, whose ratio of 4.80 fixes lambda at 0.302 in the tangent form the engine uses. A justification rather than a check: a ratio between two specimens by two methods is not an external validation, and it agrees with the number already in use, which is the reason to be careful with it rather than the reason to trust it. The toughness was measured at psi -6.5 degrees, close to mode I, and is raised to the edge-delamination angle of 46° by the default tangent law at lambda 0.3. No mixity law is fitted for this interface, so that factor of 1.4 is a convention rather than a measurement. It is the same factor in every run of this case, so it moves the margins together and leaves their order alone. The niobium yield strength used here was measured at room temperature and the case runs at 1573 K. Niobium softens substantially over that interval, so the interlayer is stronger in this calculation than in service and the margin is optimistic. For scale, the niobium alloy WC-103 retains roughly half its room-temperature proof stress at 1500 degrees Celsius (Smithells Table 22.57, pp. 22-157 to 22-158). A niobium yield strength measured at the case temperature would settle it.