What a rule document covers

A rule document and the standard reader handle everything listed here. A preparation hook may add only what this list lacks: the work that brings a table's rows to one hierarchy level. Anything listed here stays in rules.json, even when the table is prepared. How rules-driven conversion works explains the syntax.

Reading

• Formats: .tsv, .txt, .csv, .parquet, and a named .xlsx sheet
• Declared detection of delimiter, decimal and thousands marks, and encoding
• Exact file name inside a vendor directory (input.file_name)
• Header signatures for --software without a parameter file (input.software_only)
• Version selection by software_version_pattern
• Levels that apply only under declared search-parameter values (requires_search_parameters)

Table shapes

• Long: one row per observation–variable measurement
• Wide: observations captured from layer headers by a regex sample group; sample_layer when other columns share the header shape
• Packed fragments: delimiter-packed values with positional or column labels

Axis columns

• Rename a vendor column; mark it required or optional
• Types: string, integer, number, boolean
• Computed: coalesce, join_nonempty
• Vendor markings: apb_Decoy and apb_Contaminant, declared once on every level, never missing
• Sequences: stripped_sequence, proforma_sequence, proforma_ion, proforma_fragment
• Sequence grammars: token_regex, site_list, embedded_site_list, plain_sequence
• Sequence characters: residue letters or declared tokens and markers; anything else fails
• Modification maps to UniMod, with unknown tokens kept, dropped or rejected
• Rows with a missing or empty-text final key component are dropped
• Two raw keys mapping to one final key raise an error

Measurements

• Numeric layers (number, integer) and factor layers with a category map
• missing_values: exact numbers and one <= threshold
• missing_tokens: text written for a missing value, such as - or NA
• value_pattern: one number extracted from structured cells by regex
• Repeated cells: error, keep_first, sum, max, or keep_best (one row per cell by a ranking layer, with optional summed layers); another reduction is a new mode here, never hook work
• sum and max leave a cell null when nothing is present, and reduce each layer separately, so with several layers one feature's cells can come from different repeated rows
• Primary layer, swapped by search_parameter_overrides
• Semantic roles: abundance, protein_assignment, fasta_accessions

Not covered: hook territory

• Joining several files and resolving foreign keys between them
• Dropping rows by a column value, such as FlashLFQ Full Sequences Mapped ≠ 1
• Splitting or rejecting multi-valued identities such as SEQA|SEQB
• Making rows that are not a hierarchy level, such as FlashLFQ peaks, parseable as one; combining their repeated cells stays a duplicate mode
• Giving rows another table's identity through a vendor foreign key, such as MaxQuant evidence's protein group, so levels share one protein identity

Prepared tables are read as text (prepare_source.py); when a rule uses sum or max, their text layer columns are read as numbers, as the direct reader does.