v1.2.0

Pointblank v1.2.0 focuses on correctness, consistency, and safety. Every table-level validation step now has a step report, and many step report layouts have been cleaned up. NaN is now treated as a missing value in every validation method on every backend. AI-powered prompt() steps are much more reliable as they no longer report every row as failing when the model's replies can't be used. Pandas 3 is now supported, and to_code()/to_yaml() now round-trip every validation method.

Breaking Changes

• Python code in YAML configs is opt-in — python: blocks and the expr/pre/active shortcuts in YAML configs now run only with allow_python=True (in yaml_interrogate(), YAMLValidator.build_validation(), and execute_workflow()) or pb run --allow-python. Declarative calls such as pb.load_dataset(), pb.col(), pb.ref(), pb.seg_group(), and the column selectors are checked without running any code, so they always work. yaml_to_python() no longer runs any code. (#527)
• data_freshness() now fails when the newest timestamp is later than the reference time. A new allow_future= parameter accepts True (the old behavior) or a duration such as "5m" to allow for clock skew. (#518)
• Pipeline.run() now raises a new ContractViolationError for contracts with on_violation="raise". Its result attribute holds the PipelineResult. Before, the error was caught and the run kept going. (#517)

New Features

• Step reports for table-level steps — get_step_report() now returns a summary report for col_exists(), row_count_match(), col_count_match(), col_pct_null(), col_pct_missing(), data_freshness(), and specially(). Before, it returned None for these steps. Each report shows the value measured during interrogation (counts, null and missing percentages, data age, the closest matching column name, and so on) next to the expected value, with a pass/fail mark. (#488)
• Pandas 3 support — The pandas<3 pin has been removed. A new CI job runs the test suite with every direct dependency at its minimum declared version. specially() also no longer renames columns in place on Pandas results. (#496)

Enhancements

• col_vals_expr() steps are now easier to tell apart in the validation report. The COLUMNS cell lists the columns the expression uses, STEP shows the expression's name if you set one, and on Python 3.11+ the footer shows the expression as you wrote it. (#492)
• col_vals_increasing() and col_vals_decreasing() now support date, datetime, and Decimal columns on every backend and skip MissingSpec codes the same way they skip nulls. (#532)
• The col_sum_*(), col_avg_*(), and col_sd_*() methods now work on Decimal columns on every backend and accept a Decimal value=. (#529)
• Comparison steps now accept Decimal and datetime.time values. Validation methods also accept ISO datetime strings with a UTC offset (e.g., "2023-10-01T13:45:30+05:00"). Thanks @DG47! (#511, #528)
• col_vals_regex(), col_vals_str_len(), and col_vals_within_spec() now check Polars Categorical and Enum columns. Thanks @pcbeingused333! (#481, #482, #483)
• col_summary_tbl() and DataScan now choose each column's profile from its dtype class. Decimal columns are summarized as numbers, Enum columns as strings, and Time columns get a min and max. Columns that can't be summarized are listed instead of throwing an error. (#501)
• emit_otel() gains a meter_version= parameter. OTelExporter now checks its arguments when it's created. (#514)
• generate_dataset() now gives numeric columns their field's dtype (e.g., int_field(dtype="Int32") gives an Int32 column), and time_field() columns now hold real time values. (#515)
• pb run and pb validate now show steps that are inactive or couldn't be evaluated, along with the reason. pb run adds Not evaluated: and Inactive: counts and an incomplete status, and --fail-on any now also fails when a step couldn't be evaluated. (#525)

Bug Fixes

Missing values and NaN

• NaN is now treated as a missing value on Polars, PyArrow, and DuckDB, as it already was on pandas. This covers comparison steps on DuckDB tables, col_vals_between(), col_vals_outside(), col_vals_not_in_set(), col_vals_increasing(), col_vals_decreasing(), and MissingSpec checks. Range bounds given as NaN are now handled too. Thanks @DG47 and @pcbeingused333! (#474, #484, #485, #486, #487, #500, #531)
• A missing value that fails col_vals_in_set() now appears among the failing rows in sundered data and data extracts. Thanks @pcbeingused333! (#480)
• When segments are derived from a Pandas column, missing values now form a None segment instead of causing an evaluation error or being skipped. Thank you @DG47! (#508)
• Aggregates that can't be computed (e.g., the standard deviation of one value) are now reported as evaluation errors instead of raising errors or failing silently. (#529)

AI validation (prompt())

• When no model reply can be used (because of a refusal, truncated JSON, an empty result, or a cut-off reply), the step is now marked as not evaluated instead of reporting every row as failing. The step note quotes the unusable reply. Thanks @pcbeingused333! (#475, #503)
• Results in a reply are only accepted for that batch's own rows. (#503)
• Each batch now starts a new conversation, so requests no longer grow with every batch and earlier verdicts can't influence later ones. Thanks @pcbeingused333! (#479)
• The JSON array is now found when the reply has square brackets before it. Thanks @pcbeingused333! (#471)
• Row verdicts are now matched to pandas rows by position, so frames with a non-default index get the right results. Thanks @pcbeingused333! (#477)
• The batch schema is now inferred from every unique row, so sparse text columns no longer cause an error. Thanks @pcbeingused333! (#478)
• The reply-token limit for Anthropic and Bedrock models has been raised to 16,000. DraftValidation and EditValidation now raise a clear error when a reply is cut off. (#493)
• The draft_validation_plan MCP tool now uses anthropic:claude-sonnet-5-5 by default, replacing a model that's being retired. Thanks @pcbeingused333! (#470)

Step reports

• col_schema_match() step reports now agree with the validation result for every combination of complete= and in_order=, and they're easier to read! (#491)
• Step report headers, assertion boxes, pass/fail marks, and widths are now consistent across report types. (#489)
• Awkward and ungrammatical text in step reports has been fixed. conjointly() and specially() reports now name their expressions instead of showing <lambda>. (#490)
• get_dataframe_report() no longer fails on segmented steps, and checksum specs in col_vals_within_spec() no longer stop interrogate() on Polars LazyFrames. (#495)

Column summaries (col_summary_tbl() and DataScan)

• col_summary_tbl() now explains why it fails when no column can be profiled, instead of raising a bare AssertionError. Thanks @cristianchiriac! (#469)
• List and array columns are now skipped instead of resulting in an error. Thanks @DG47! (#499)
• Empty cells now show - on every backend, HTML is no longer shown as text, and string-length statistics now ignore missing values. (#495, #502, #506)

Serialization (to_code(), to_yaml(), yaml_to_python())

• to_code() and to_yaml() now write out every validation method, and the output gives the same results as the original plan. (#494)
• Validate(reference=...) is now written out when the reference is a file path or connection string, so plans that use pb.ref() can be rebuilt. (#504)
• Decimal, time, and timezone-aware datetime values are now written out with the imports they need. (#528)
• The original active= value is now kept when a plan is serialized after interrogation. (#521)
• yaml_to_python() now writes working code for file paths and connection strings. (#505)

Validation workflow

• Calling interrogate() again now starts over from the steps as they were defined, and brief templates are kept. (#524)
• get_sundered_data() now returns the right rows for segmented steps and leaves out steps that use pre=. (#495, #497)
• Segmented steps are now kept when no segments are found, instead of being dropped from the validation. (#516)
• add_steps() now works when copying steps from a Validate object for methods that don't take value=. (#523)
• emit_otel() now sends traces and log records, the same as OTelExporter.export(). (#514)

Data generation

• With seed=, each generated column now gets its own random sequence, so columns are no longer identical or correlated. Seeded multi-country data is now the same in every session. (#526)
• Prefixed profile_fields() profiles now each get their own person, location, and employer. (#519)
• The postcode in a generated address now matches the row's postcode column. (#522)
• get_data_path() no longer leaves temporary files behind. (#520)

Documentation

• The User Guide sections are now ordered to follow how people might use Pointblank: the core validation workflow first, then data tools, production use, and other interfaces. (#509)
• Retired OpenAI and Claude models have been replaced in the AI examples. Thanks @pcbeingused333! (#472, #493)
• Several small documentation fixes, including clearer report terminology and customization guidance. (#512)

Maintenance

• The duckdb extra now requires ibis-framework>=11.0.0 and narwhals>=2.8.0, which fixes queries on ibis 10.x with DuckDB 1.4+. DuckDB is temporarily capped at <2 until ibis and Narwhals support it. (#476)
• The minimum versions are now narwhals>=2.8.0 and great-tables>=1.1.0, and the generate extra now requires chatlas>=0.10.0. (#496, #513)