Every score and warning should be explainable.

Delimiter detection

The detector tests comma, tab, semicolon, and pipe candidates against up to 120,000 characters and 40 logical rows. Candidates are scored primarily by consistent row width, with smaller weights for useful column density and sample size. Quoted separators and quoted line breaks are parsed as field content.

Quality profile and readiness score

The profile measures missing cells, exact duplicate rows, mixed observed types, empty rows, duplicate JSON keys, formula-like spreadsheet cells, and columns with likely PII. The score begins at 100 and applies capped penalties: missingness up to 25 points, duplicate rows up to 20, mixed types up to 20, PII up to 20, structural warnings up to 14, and formula risk by 8. Labels are Strong (90–100), Workable (75–89), Needs review (50–74), or High risk (below 50).

The score is a triage aid. It is not a statistical quality measure and does not evaluate truth, bias, provenance, representativeness, or suitability for a particular decision.

Likely PII detection

The scanner combines column-name hints with deterministic value patterns for common emails, US-style SSNs, IPv4 addresses, Luhn-valid payment-card numbers, and phone-like values. Name, address, date-of-birth, credential, and passport warnings rely on field-name hints. Values are sampled per column, so rare identifiers can be missed and ordinary values can be flagged.

Value-format normalization

The Format tab applies deterministic, column-level rules for dates, phone numbers, numbers, text, names, email, and booleans. Name formatting trims surrounding whitespace, uses name-aware title case for apostrophes, hyphens, and Mc prefixes, and restores missing apostrophes in a conservative list of common compact O' surnames. Ordinary O-names such as Oscar, Oliver, and Owen are left as standard title case. Ambiguous numeric dates remain invalid until an MDY, DMY, or YMD input order is selected. Phone national formats use an explicit United States, Canada, United Kingdom, or Australia default. The preview reports values that will change, remain unchanged, or fail the selected rule before the working copy is updated.

Formatting checks representation, not truth. It does not verify that a date reflects a real-world event, a phone number is assigned, or an email address can receive mail.

Schema inference and validation

Schema generation describes types, nullability, required fields, and common formats observed in the loaded records. It cannot predict unseen values. JSON Schema validation targets Draft 2020-12 and supports local fragment references; external references are not fetched.

JSON selection

JSONPath querying follows RFC 9535 and returns an ordered list of matching values with normalized paths. Queries are limited to 1,000 characters, 10,000 matches, and 5 MB of serialized output; the interface displays the first 200 matches while copy and dataset actions use the full bounded result. RFC 6901 JSON Pointer remains available for exact single-node extraction.

Verification and change control

Parser, transformation, browser-layout, build, and Worker behavior are covered by automated regression checks before deployment. Method changes are recorded in the public changelog and source history.