R and Python¶

credgate is one tool in two packages. Both read the same specification, and a shared set of conformance cases - run by both test suites on every change - holds them to the same results. This page lists what is the same, how the names map, and the few places they differ, and why.

The same in both¶

Given the same input, R and Python give the same output, and each reads what the other writes:

Detection: the same matches, confidence scores and reasons, and the same explanations of what was not reported.
Redaction and tokens: the same placeholders and tokens; keyed tokens agree across languages, and a vault saved by one opens in the other.
Column scans and PII inventories: the same findings, samples and inventory files.
Policies, policy tests and the audit log: the same decisions, test results and audit records.
The policy editor: the same app, tabs and roles, on the same files. A policy saved by either editor is written byte for byte the same, with the same change log.
Custom detectors, proxy keys, the users file and the access log: the same file formats, read and written the same way.

Names¶

Each language follows its own conventions: R uses functions with a cg_ prefix where a bare name would clash, Python uses classes and methods.

	R	Python
Detect	`detect_pii()`	`credgate.detect_pii()`
Preview, redact	`preview()`, `redact()`	`credgate.preview()`, `credgate.redact()`
Tokenise, restore	`tokenise()`, `restore()`	`credgate.tokenise()`, `credgate.restore()`
Vault	`cg_vault()`, `vault_save()`, `vault_open()`, `vault_merge()`	`credgate.Vault`, `.save()`, `Vault.open()`, `.merge()`
Keyed tokens	`cg_new_token_key()`, `cg_key_fingerprint()`	`credgate.new_token_key()`, `credgate.key_fingerprint()`
Detectors	`cg_detectors()`, `new_detector()`, `register_detector()`, `reset_detectors()`, `load_detectors()`, `bounded()`	`credgate.detectors()`, `credgate.Detector`, `register_detector()`, `reset_detectors()`, `load_detectors()`, `bounded()`
Normalise	`cg_normalise()`	`credgate.normalise()`
Mask a value	`mask_value()`; `options(credgate.mask = "full")`	`credgate.mask_value()`; `credgate.set_mask("full")`
Checksums	`verhoeff_valid()`, `verhoeff_check_digit()`, `luhn_valid()`, `luhn_check_digit()`, `gstin_valid()`, `gstin_check_char()`	the same names in `credgate`
Column scan	`scan_pii()`	`credgate.scan_pii()`
Inventory	`pii_inventory()`, `write_inventory()`, `read_inventory()`	`credgate.pii_inventory()`, `.write()`, `credgate.read_inventory()`
Databases	`inventory_database()`, `redact_query()` (DBI)	`credgate.database.inventory_database()`, `redact_query()` (SQLAlchemy)
BigQuery	`bq_scan_pii()`, `bq_pii_inventory()`, `bq_inventory_dataset()`	`credgate.bigquery.scan_pii()`, `pii_inventory()`, `inventory_dataset()`
Parquet	`redact_parquet()`	`credgate.arrow.redact_parquet()`
Policies	`cg_policy()`, `policy_decide()`, `policy_apply()`	`credgate.Policy.load()`, `.decide()`, `.apply()`
Saving policies	`policy_write()`, `policy_diff()`, `policy_changes()`	`credgate.write_policy()`, `policy_diff()`, `policy_changes()`
Policy tests	`run_policy_tests()`	`credgate.run_policy_tests()`
Audit log	`cg_audit_log()`, `audit_record()`, `audit_verify()`, `audit_read()`, `audit_prune()`	`credgate.AuditLog`, `.record()`, `audit_verify()`, `audit_read()`, `audit_prune()`
Policy editor	`policy_editor()`	`credgate.editor.policy_editor()`, `credgate editor`
Proxy keys	`proxy_add_user()`, `proxy_rotate_key()`, `proxy_remove_user()`, `proxy_users()`, `proxy_access()`	`credgate.proxy_keys.add_user()`, `rotate_key()`, `remove_user()`, `list_users()`, `read_access()`
Users file	`cg_users()`	`credgate.users.read_users()`
Evaluation, test data	`cg_evaluate()`, `cg_read_corpus()`, `synth_indian_pii()`	`credgate.evaluate()`, `credgate.load_corpus()`, `credgate.synth_indian_pii()`
The spec	`cg_spec_path()`	`credgate.spec_path()`

Only in Python¶

The proxy (credgate proxy serve). It is a server, run once for a team or per machine; R and every other tool use it through its URL - for example ellmer::chat_openai(base_url = "http://127.0.0.1:8790/v1").
Managing users and unused keys on the server: credgate users ... (who may use the app, and their roles) and credgate proxy expire-unused. Whoever runs these on the server is the super admin; the R package reads the users file but does not change it.
The command line: credgate detect, redact, scan, policy test and the rest, for shells, CI and pre-commit hooks.
Polars, pandas and DuckDB SQL functions: Python data tools without an R counterpart. In R, the same work goes through data frames, Arrow, dbplyr and DBI.

Small differences¶

Field names: a policy decision's rewritten text is decision$text in R and decision.texts in Python, and a match's data class is the column class in R and span.data_class in Python. Python raises an error for a misspelt field; R gives NULL, so check the name if a field is unexpectedly empty.
Positions: R counts documents and positions from 1 with the end included, Python from 0 with the end excluded, as each language indexes strings. A policy decision's matches are a data frame in R and a tuple of PolicyMatch in Python.
Masking everything: R takes an option (options(credgate.mask = "full")), Python a function (credgate.set_mask("full")); both also read the environment variable CREDGATE_MASK.
The tests file (bank.tests.yaml): each editor writes it with its language's own YAML library, so quoting, indentation and line breaks can differ. The contents are the same, and either language reads either.
The Python policy editor with Nix on Python 3.11: nixpkgs cannot build Shiny for Python for 3.11, so install it from PyPI there (pip install 'credgate[editor]').
Releases: the Python package is on PyPI. The R package's first version is under review at CRAN; until it is accepted, everything since is in the development version on R-universe.

For AI coding agents¶

A plain-text summary for agents, with the main functions in both languages and the differences above, is at llms.txt.