Where laterite and python-ags4 differ

laterite is a clean-room re-implementation of the AGS4 rules, calibrated against the incumbent python-ags4 on its own test corpus (see Cross-surface parity). Two independent implementations of one specification will disagree, so every disagreement is written down rather than smoothed over.

17 of them change what you see. They are not all the same kind of thing, which is why this page is grouped by what actually happened rather than filed under one heading: some are deliberate differences from python-ags4, some are places the two agree and the spec is the outlier, and some are laterite's own false negatives that the comparison caught and closed.

This is the user-facing list. The full catalogue — including the internal NOTE/SPEC entries and the records since resolved — lives in OBSERVATIONS.md in the repo, and this page is generated from the same source, so a record cannot be resolved there and stay live here.

Where laterite differs from python-ags4

#	What you see
O-2	Rule 6 is a no-op in python-ags4 — its body is return ags_errors. laterite implements the embedded-CR check.
O-8	python-ags4's rule_7_2 can raise IndexError on duplicate headings; laterite bounds-guards it and reports.
O-12	The DT/datetime validity engine is chrono-based and stays lenient on semantics for UNIT shapes the AGS dictionary does not define, where pandas would still attempt a parse.
O-30	The dictionary edition is selected from TRAN_AGS. An AGS 3.x file is refused outright, where python-ags4 silently validates it against 4.1.1.
O-34	A non-AGS4 file surfaces as a clean NotAgs4 error, reconciled against python-ags4's "missing mandatory groups".
O-37	The native parser is lenient where python-ags4 raises hard (duplicate GROUP, ragged rows) — findings first, never a crash.
O-41	Rows before the first GROUP are reported as Rule 2 findings, not a parser crash.
O-42	TRAN_AGS="4.0" resolves to 4.0.4, the newest 4.0 patch; python-ags4's static map picks the oldest and over-reports Rule 10c on PMTL.
O-49	A numeric TYPE's count — the n in nDP/nSF/nSCI — is clamped to 30. Read uncapped, a crafted 9999999999SF drives python-ags4 into a ~10 GB string.
O-50	A 0DP value outside i64 converts to Null; python-ags4's conversion keeps full precision. Both validators flag the cell — the difference is in conversion, not validation.

Where both depart from the written spec

#	What you see
O-1	Rule 1's "entirely ASCII" is an FYI for extended ASCII (128–255), not a hard error — matching python-ags4, which the spec text does not.
O-32	Non-UTF-8 input is decoded lossily (python's errors="replace"), not refused.

Where laterite changed to match python-ags4

#	What you see
O-31	Rule 8 flags an empty DT UNIT. This was laterite's own false negative, found by the comparison and closed to match python-ags4.
O-33	DT/datetime validity is bounded to pandas' Timestamp range — again laterite's own false negative (the year 0018 was accepted), closed to match python-ags4.

Checks laterite adds

#	What you see
O-43	A self-declared but non-standard PA abbreviation → a laterite FYI (Related to Rule 16).
O-44	Structural validation of a file-level DICT group → a laterite WARNING (Related to Rule 18).
O-45	An unrecognised TRAN_AGS edition → a laterite WARNING (Related to Rule 14), shown by default.

Reading the tiers

Whether a difference surfaces as an error, warning or FYI follows laterite's severity tiers.