Troubleshooting and known limitations

This page covers v1.0.3. Preserve the project before experimenting with recovery steps, and never edit a project or checkpoint archive in place.

A file imports as one column

Reopen the import dialog and set the delimiter explicitly. Check whether the file uses comma as a decimal separator; if so, choose semicolon, tab, or whitespace as the field separator. Also verify encoding, quote character, and header detection.

If rows have inconsistent field counts, inspect the diagnostics. Equarith can pad or truncate malformed row shapes, but broken quotation prevents safe parsing.

Rows disappear from search

A search row must be finite in the target and every selected input. Missing, malformed, NaN, and infinite values are excluded before splitting.

History functions also remove warm-up rows and rows whose required preceding values are invalid. In Demo mode, only the first 200 source rows are considered before this filtering, so fewer than 200 can remain.

Reduce the selected inputs, repair values, or choose a split that fits the eligible count. Do not replace missing values with zero unless zero is scientifically correct.

Search will not start

Confirm:

• exactly one target and at least one input are selected;
• at least three eligible observations remain and both requested partitions can be formed;
• Demo uses at most four inputs;
• minimum bounds do not exceed maximum bounds;
• at least one valid operation set exists;
• function costs and options are in range; and
• enough memory is available for the dataset, selected operations, and processors used by the search.

The readiness message identifies the owning setting. Importing a saved configuration does not bypass validation.

Search is slow or the computer is unresponsive

Desktop search uses every processor exposed to the application; the CPU workers preference does not limit it. Reduce enabled function families, maximum depth, formula complexity, coefficient-optimization effort, and history window. Use Auto progressive evaluation for large datasets. For a hard search-thread limit, run Equarith Headless with --threads=<n>. A short search with focused functions is often more informative than an unrestricted search.

Evaluation rates cannot be compared fairly across different metrics, operations, row counts, sampling strategies, or machines.

No useful formula appears

• Verify target, units, and column alignment.
• Plot and summarize the data before searching.
• Start with operations justified by the domain.
• Increase the time limit or try a different search seed.
• Compare coefficient presets.
• Try normalization for poorly scaled variables.
• Load a plausible expression as a seed.
• Allow more complexity only when validation justifies it.

Symbolic regression cannot recover information absent from the inputs, and a noisy or stochastic process may not have a compact deterministic expression.

Training is good but test is poor

This commonly indicates overfitting, leakage, a distribution shift, or an unstable small test set. Prefer a simpler Pareto neighbor, reduce functions and complexity, inspect chronological ordering, and repeat with a defensible split.

Do not tune repeatedly against the same test set. Reserve external validation data when decisions are important.

A prediction is invalid or extreme

Inspect logarithm, root, inverse, strict division, tangent, power, gamma, and conditional domains. Compare the input with its observed training range. Protected functions prevent some undefined values by changing semantics and can still produce scientifically implausible extrapolation.

Custom points cannot evaluate history formulas because they do not define preceding rows.

A referenced project does not restore its data

The external source may be missing, moved, edited, or decoded differently. Restore the exact source file or reopen a portable project with embedded data. Equarith checks fingerprint and schema so saved solutions are not silently applied to different observations.

Recovery archives are not backups. If the application stopped abnormally, use the offered recovery, inspect it, and immediately save a named project.

A checkpoint does not resume

Resume requires compatible data, split, metric, functions, complexity costs, constraints, options, and checkpoint version. A completed checkpoint is not resumed. Any intentional model change should start a fresh run.

A forced shutdown can lose the latest checkpoint. Earlier published results remain usable.

Activation or update fails

Check system date and time, network access, proxy or firewall policy, and the entered license key. A license already active on another machine must be deactivated there or resolved through the license service.

Certificates expire within 30 days. An eligible certificate can be refreshed using the installation identity without re-entering the commercial key; reactivate with the key only if refresh is unavailable or rejected. Built-in service calls cover updates, activation, refresh, deactivation, explicit feedback, and the enabled-by-default encrypted anonymous error reports. Any of these can fail without stopping local Demo search; automatic reports can be disabled in Settings.

Executable export is unavailable

Standalone code export does not support history functions, user-defined function bodies, template placeholders, or unknown named constants in v1.0.3. Use a presentation export, evaluate against an ordered dataset, or choose a formula built only from exportable built-ins.

Important v1.0.3 limitations

• Numerical regression only; no classification and no causal inference.
• Expression-defined custom metrics, executable user functions, and formula templates are not executed by EquarithEngine.
• History functions need ordered dataset rows, remove warm-up rows, do not work for custom points, and have no executable source export.
• The Pareto archive holds at most 256 entries and one representative per complexity.
• Search-generated power exponents are non-zero integers from -3 through 3.
• Maximum usable data depends on memory, width, operations, and workers; hard bounds reject unsafe dense inputs.
• Interactive charts decimate to the configured point budget.
• Exact seeded reproducibility assumes an equivalent runtime, worker count, machine environment, and version.
• Eligible license certificates can be refreshed without key re-entry; activation with the key is needed only when refresh is unavailable or rejected.
• The local headless protocol supports one active search per runtime process.
• Official Python, TypeScript, JavaScript, Java, C#/.NET, C++, and Lua SDK packages are coming soon; use the available CLI or local stdio protocol until they are published.

Information to collect for support

Record:

• Equarith version and license mode;
• operating system, architecture, and Java version when using the development JAR;
• exact steps and the visible error;
• dataset dimensions and selected columns, without sharing confidential values;
• relevant configuration and seeds; and
• a redacted structured log around the failure.

Logs can contain file paths, column names, formula text, or error context. Review them before sharing. Start again from Getting started or consult Licensing, privacy, and network access.