JSON Smart Diagnostics

JSON formatting and minifying: Format for readability or minify optional layout whitespace. Choose one mode at a time. Strings, numeric spelling and comments are retained. XML/HTML mixed text and preformatted content, and YAML/TOML required line breaks and indentation, stay intact. Review the preview before applying it.

How to use: Paste your text and choose Diagnose & repair. Auto preview is off by default; enable it to update after editing. Choose formatting or minifying, review the changes, then copy the result or apply it to the original. Diagnosis runs in your browser; the site uses Google Analytics for traffic statistics.

How the model works: A trained neural network with 6,753 parameters ranks repair candidates. A strict JSON parser checks the results. Candidate scores are relative rankings, not probabilities of a correct repair. A recovery scanner also continues after unresolved fields. It makes conservative local repairs and leaves unknown values visible for manual review.

Supported cases: Common missing commas or colons, extra commas, single or curly quotes, unquoted keys, True/False/None, full-width punctuation, comments, and some bracket and escape problems. Repairs can take multiple steps. Formatting preserves the exact text of numbers and strings. Also recognizes leading zeros, hexadecimal numbers, a leading plus, non-finite numbers, incomplete Unicode escapes, missing values and array holes. Numeric conversions preserve decimal precision; non-finite and missing values are not replaced with null. Duplicate keys in the same object are highlighted at both occurrences.

What you should review: Valid syntax does not guarantee the intended data. Missing values, truncated strings and ambiguous numbers are not invented; amounts, IDs and business rules cannot be verified. Bracket and escape repairs can be ambiguous. Compare alternative valid repairs when available. Duplicate keys are reported and preserved.

Tests and limitations: On independent synthetic tests, 5,583 / 5,598 single-error cases and 245 / 250 multiple-error cases recovered the original data meaning. Training and tests use similar synthetic distributions; these are not accuracy rates for arbitrary real JSON. One other multiple-error result parsed successfully but did not recover the original meaning. These figures describe the original model-search tests, not the accuracy of the added recovery scanner.

Processing limits: Up to 131,072 UTF-16 code units and 128 nesting levels. Model search uses up to 24 steps; recovery allows up to 256 edits and 512 diagnostic items. If a limit is reached, the incomplete check is reported for manual review.