Where the fit could be certified — and where it could not.
A development-run evaluation of Caldrivon R176 on three published real-data sources. This note distinguishes reported outcomes from independently established claims.
01 / SCOPE
What was tested
R176 is designed for fitting constrained models in which numerous local groups share a small set of global parameters. It can enforce bounds such as nonnegative coefficients, provide an optimality certificate for the specified model, and bound rounding error when its search closes. Exhausting the search budget results in abstention.
The development-run trial used 40,223 observations from selected subsets of three publicly available datasets. The user-provided trial account reports zero synthetic inputs. These subset sizes must not be confused with the full source-dataset sizes.
02 / RESULTS
Observed-data outcomes
| Dataset | Rows tested | R176 result | Time |
|---|---|---|---|
| Household electricity | 20,156 | Certified | 219 ms |
| Capital Bikeshare rentals | 17,379 | Certified; 12 local constraints binding | 150 ms |
| Steel plant energy | 2,688 | Abstained after search budget exhaustion | Not reported |
A repeat run reportedly reproduced both certificates and the steel abstention. It also reproduced the SciPy tolerance outcomes.
03 / COMPARISON
What the SciPy comparison means
On the Bikeshare fit, the tuned SciPy run reported convergence while missing the trial's preset objective tolerance. On the household-electricity fit, it met that tolerance. This shows a meaningful difference on one specified case; it is not evidence that SciPy, as a library, cannot solve that class of problems, nor a general speed or accuracy superiority claim.
To permit a full independent comparison, the model formulation, solver options, stopping rules, arithmetic settings, benchmark hardware, exact result certificates and runner code would need to be released or independently inspected.
04 / OPEN QUESTIONS
What has not been proved
- Independently selected datasets and external customer workloads have not yet been evaluated.
- Certificate files and a complete independently executable benchmark bundle have not yet been published with this note.
- Runtime figures are case-specific and should not be treated as guaranteed production latency.
- R176 did not certify the steel case within its fixed budget.
- Numerical fitting results do not validate Caldrivon's separate Python code-repair workflow.
Increase the range of real constrained fits that can be certified under a fixed search budget—beginning with the steel-energy case—while preserving correct abstention whenever a certificate cannot be established.
05 / DATA PROVENANCE
Public dataset sources
These links establish the existence and nature of the public sources. They do not independently substantiate the R176 timings or certificates.
06 / CONTACT
Have a comparable fitting problem?
We can discuss a limited, non-sensitive case with clear objective, model assumptions, constraints, baseline and acceptance criteria.