Praecepta 0.2.0

Boolean, time and date-time columns; the DMN conformance kit (TCK) in every build; fast builds for
large FIRST and PRIORITY tables; and strict FEEL when reading DMN files. Read Changed before
upgrading: DMN files that relied on unquoted strings or bare dates no longer load.

Added

• Boolean, time and date-time columns (#12):
  ColumnType.BOOLEAN, TIME and DATE_TIME, with cells true / false, time("08:30:00") and
  date and time("2025-01-31T08:30:00") (or the bare ISO text), comparisons and intervals on times
  and date-times, and the bindings booleanInput, timeInput and dateTimeInput. DmnReader
  reads boolean, time and dateTime columns, DmnWriter writes them, and the Maven plugin
  generates Boolean, LocalTime and LocalDateTime components. Times are local: values with a
  zone or offset are rejected. The DMN TCK now passes 51 of 86 checks (was 16).
• Input names with spaces (#47):
  DmnReader accepts FEEL names such as Credit Score or Applicant data.Monthly income.
• DMN TCK results (#22): a new
  praecepta-tck module runs the decision-table test cases of the DMN Technology Compatibility Kit
  in every build, and keeps the published results (praecepta-tck/RESULTS.md) current: what passes,
  and what each unsupported check needs. Every supported check passes.
• Broader benchmarks, in which every engine, Praecepta included, reads the same DMN document:
  • tables from 10 to 10,000 rules, and with 5 to 20 columns;
  • every hit policy, including UNIQUE and PRIORITY;
  • requests that match no rule;
  • latency percentiles and memory allocated per decision;
  • throughput with an engine shared by 1, 4 and 8 threads;
  • time to build an engine from DMN, warm and in a fresh JVM;
  • run-benchmarks.sh runs them all and records the machine and versions with the results.
• Fuzz tests (Jazzer) for the cell parser and the DMN reader: they replay saved inputs in every
  build, fuzz weekly in CI, and fuzz every pull request that touches the engine (ClusterFuzzLite).
• Releases attach their build provenance (praecepta-X.Y.Z.intoto.jsonl) for offline
  verification.
• Project documents: governance (roles, decisions, continuity), architecture, a security
  assurance case and a one-year roadmap. The website shows the roadmap, a C4 system context diagram,
  and what users can expect on security, and links them all.
• A project website at https://bruno.pacheco.lu/praecepta/, with the API documentation of the
  latest release, and a Classic Rome identity: a bronze tabula ansata logo, Cinzel and EB Garamond.
• An enforced code style: Checkstyle checks every build (checkstyle.xml, described in
  CONTRIBUTING.md), and compiler warnings fail the build (-Xlint:all -Werror).

Changed

• DmnReader reads cells as strict FEEL: an unquoted word in a string cell is a reference to
  another variable, so it is rejected as unsupported instead of being read as a string (the DMN TCK
  caught Complex.aString being compared as text); and dates and times need their functions
  (date("2025-01-31")), since in FEEL a bare 2025-01-31 is a subtraction. Excel, CSV and code
  keep the shorter PT,ES and bare ISO forms; CellParser.parseFeel is the strict parser.
• ColumnType has three new constants (BOOLEAN, TIME, DATE_TIME): an exhaustive switch
  over it needs new cases.
• CsvLoader and ExcelLoader are final, and BaseLoader's helpers are final too: the
  loaders read their file in the constructor, which a subclass could otherwise observe half-built.
  Extend BaseLoader to write a loader of your own.
• Exceptions declare a serialVersionUID.

Fixed

• Large FIRST and PRIORITY tables build quickly (#29):
  the unreachable-rule check, still exact, now remembers the regions it has proven covered, tests
  whether one rule covers the rest with a single mask intersection, and looks for an uncovered input
  before searching deeper. A 1,000-rule, 9-column table now builds in 0.5 s under FIRST (was 3.2 s)
  and 0.7 s under PRIORITY (was 65 s), against 0.3 s under COLLECT; the reports are unchanged.

Praecepta 0.1.0

The first release since the 2022–2023 prototype: the engine, loader and Maven plugin are rewritten.
Code written against the prototype needs the changes listed under Changed.

Changed

• Licence: Praecepta is now licensed under the Apache License 2.0 (previously MIT). No version
  was ever released under MIT.
• Java 21 baseline (tested on 21, 25 and 27). The engine is a JPMS module (io.github.brunopacheco1.praecepta.engine,
  requires only java.xml). The loader and plugin declare stable automatic module names.
• Engine API:
  • engines are built once with PraeceptaEngine.builder(), and are immutable and thread-safe;
    register() is gone;
  • inputs bind per column (input, numberInput, dateInput, enumInput) or as a map with
    inputs(...);
  • evaluate returns an empty list when no rule matches.
• Strict DMN 1.3 hit policies (UNIQUE, ANY, PRIORITY, FIRST, OUTPUT_ORDER, RULE_ORDER, COLLECT
  with SUM/MIN/MAX):
  • rule order is the order of the rule list; Praeceptum.priority is only an identifier now;
  • UNIQUE overlaps and ANY disagreements throw HitPolicyViolationException;
  • PRIORITY and OUTPUT_ORDER take output value lists (outputPriority).
• Matching uses a bitset index: memory grows linearly with the table (10,000 rules in under
  1 MB), and engines build in milliseconds.
• The Maven plugin is rewritten as a real plugin (goal praecepta:generate):
  • it reads .xlsx, .csv or .dmn tables;
  • validates them at build time (<failOn>);
  • writes them as DMN resources;
  • generates typed input/output records (enums for string value lists) and a decision class
    that needs only the engine at runtime;
  • it runs on Maven 3.9 and later (tested on 3.9, 3.10 and 4.0);
  • the old COMPILATION and in-memory generation modes are gone.

Added

• Cells in a DMN FEEL subset on string, number and date columns: lists, not(...),
  comparisons, intervals, null. Unquoted PT,ES cells from the prototype keep working.
• Typed columns: number values compare as exact decimals, dates are LocalDate, and map-based
  inputs declare types with columnType.
• Input value lists (allowedValues), enforced at evaluation: other values throw
  InvalidInputException.
• Validation reports (validationReport(), failOn(...)): overlapping rules, conflicting
  outputs, unreachable rules, and gaps (declared inputs no rule matches, with an example).
• DMN 1.1–1.5 import (DmnReader), including Camunda Modeler files, and DMN 1.3 export
  (DmnWriter). Builder.table(...) applies a whole table.
• Enum input bindings (enumInput): lookups by ordinal instead of hashing (104 ns vs 233 ns on
  the bundled 119-rule table).
• Maven plugin <fileHeader>: your own licence header on generated sources.
• Loaders: the Excel and CSV loaders also read from an InputStream.
• Maven Central publishing (io.github.brunopacheco1): praecepta-engine,
  praecepta-loader, praecepta-maven-plugin. Each release is GPG-signed, with a CycloneDX SBOM
  per module, build provenance and third-party notices.
• Project tooling:
  • REUSE-compliant licensing (SPDX headers) and a dependency licence check;
  • reproducible builds;
  • a mandatory SonarCloud quality gate;
  • CodeQL, the OpenSSF Scorecard and ORT licence policy;
  • Dependabot with full dependency-graph submission, and dependency review;
  • pinned CI actions.

Fixed

• Excel loader: numeric cells keep their exact value (they were rounded); date cells become ISO
  dates (they were serial numbers); formula cells use their cached result.

Security

• DmnReader refuses DTDs and external entities (XXE), and parses with the JDK's secure
  processing enabled.
• The Maven plugin rejects <fileHeader> text with backslashes, which could otherwise inject
  Java unicode escapes into generated sources.

Removed

• The App Engine benchmark services and Gatling stress tests (replaced by in-process JMH
  benchmarks against Camunda 8 and Drools).
• MissingOutputException (no match now yields an empty result), InvalidPraeceptaException
  and the trie visitor API.