How ModelScope
decides.
A statistical comparison with explicit safeguards. Every finding links back to measurements and calculated evidence.
Read the full methodology- 01
Fit the baseline
Use the configured line or fixed theoretical reference. Compare the same observations.
- 02
Compute residuals
Observed minus predicted. Structure matters more than a high R².
- 03
Search candidate transitions
Compare a continuous hinge at eligible measured inputs. Penalize complexity and the search.
- 04
Require sustained deviation
At least four consecutive post-knee residuals of one sign must exceed twice the reference scale.
- 05
Check influence
Diagnostically omit the largest absolute baseline residual. Support and departure direction must remain.
- 06
Return the evidence
Supported, ambiguous or none. Insufficient sampling is reported separately; no causal claim is made.
Try to fool ModelScope.
An unusual point, ordinary noise, or a sustained departure. The same detector handles each case.
No supported transition
One large residual is not a sustained departure. A single point cannot establish a supported transition.
- Engine outcome
- none
- Penalized improvement
- -11.03 / required 10
- Sustained measurements
- 0
- Influence safeguard
- Not reached
No supported transition does not prove model adequacy.
Brief visual reveal only. Decisions use all 24 observations · assumed scale 0.040 N.
Evidence has limits.
A supported candidate describes statistical disagreement. Its sensitivity range is not a confidence interval, physical failure point, or safe operating boundary.
Assume accurate inputs, adequate sampling, an explicit response scale and one possible transition. These demonstrations use synthetic data.
Inspect a complete analysis