PDF2Data v0.2.0-rc.20

Changelog

Unreleased

0.2.0-rc.19 — 2026-11-03

• Retain complete fact and structure candidates in processed exports, with exact
  artifact hashes and page witnesses. The additive inventory describes the new
  export snapshot; it never reseals an older package.
• Show package source-quality outcomes separately from extraction selection in
  readable documents, page HTML/Markdown and overlay data. Repair holds remain.
• Retain the exact rc18 custody decoder for existing supported operations.
  Existing extraction jobs and consumer pins require explicit adoption.

0.2.0-rc.18 — 2026-11-03

• Route exact KrutiDev010 source fonts to Hindi glyph-map review and retain
  known font/inventory holds through capture, assessment and machine decisions.
  Preserve source text; native agreement does not decode or approve language.
• Retain the exact released rc17 decoder for registered custody operations,
  with changed-code refusals and an independently replayed frozen table proof.
  Existing jobs, quality policies and consumer pins require explicit adoption.

0.2.0-rc.17 — 2026-11-02

• Give custody, targeted proof and native table review workers isolated Python
  processes with fresh private cache prefixes and bounded cleanup. Keep cache
  writes disabled; environment-only settings do not establish this boundary.
• Retain the exact released rc16 decoder as a historical proof epoch, alongside
  rc12/rc13/rc14 and current code. Changed code and unsupported operations remain
  refused under the existing source, parent and dependency checks.
• Preserve originals, frozen quality proofs, failed receipts and consumer pins.
  Proof reconstruction does not grant semantic, legal or publication acceptance.

0.2.0-rc.16 — 2026-11-02

• Bind native grouping strokes and adjacent text blocks to exact independent
  paths, transforms, paint, source rectangles and layout witnesses. Keep group
  and header relationships as candidates; preserve blank cells without filling.
• Account for literal frames, separators and footnote rules using isolated
  original-object raster pixels. Bounding bands grant no ink ownership;
  foreign adjacent paint and genuine glyph intersections remain held.
• Give font traces sharing a paint operation distinct glyph keys and require
  complete paint cardinality before claiming individual-character ink bounds.
• Retain component diagnostics, original packages and failed receipts. Source
  quality repairs create separate immutable successors with fresh registered
  decoder checks; they do not repin jobs or grant legal/publication acceptance.

0.2.0-rc.15 — 2026-11-02

• Add bounded named local image-reading candidates with exact original/prompt/profile receipts. Preserve older readings, retain incomplete output and all source-quality holds, and leave selected L1s unchanged.
• Add scoped native source-object/raster accounting without transcription,
  language, layout or page-quality clearance.
• Compare configured multiword operators across explicit layout spacing while
  preserving raw spans, separate numeric checks and frozen quality outcomes.
• Retain the exact released rc14 proof decoder alongside rc12/rc13; require all
  new command modules in distributable-wheel checks.

0.2.0-rc.14 — 2026-11-02

• Add explicit bounded English cardinal and printed-percent quantity modes.
  Preserve literal/source spans, comparison context and quality holds; keep
  configured unit hints separate from anchored printed units.
• Add exact rational/null quantity assertions to field checks and enforce the
  quantity-only normalization boundary in the public extraction-rule schema.
• Retain exact released rc12 and rc13 custody decoder compatibility alongside
  current code, with the existing operation, source, dependency and hash gates.
• Qualify seven retained real-source word quantities and separate printed-unit
  witnesses. This does not establish legal, calculation or pipeline acceptance.

0.2.0-rc.13 — 2026-11-02

• Replay frozen native proofs with explicit hash-checked custody bindings and an exact captured decoder closure; preserve captured provenance and reject substituted parents or external review inputs.

• Support native table source proofs from installed wheels by requiring all
  proof modules to share the actual current package directory. Mixed modules
  remain refused; captured extraction origins stay data rather than import paths.

• Refuse proof outputs overlapping source packages, extraction runtimes or input
  receipts. Keep incomplete cell review as an explicit terminal hold without
  creating a quality version.

• Qualify the installed route on bounded source cases with existing independent
  cell/region decoders. Preserve raw grids, source text, numeric/operator holds
  and all frozen consumer artifacts; no source acceptance rules are relaxed.

0.2.0-rc.12 — 2026-11-02

• Expose native-whitefill-prose-reproof over the existing registered source
  proof route. Extend its closed predicates to account for thin, long footer
  rules disjoint from glyphs, alongside exact white paragraph backgrounds.
• Require exact single-column candidate/cell/rectangle ownership, fresh paint
  visibility, independent source lines and all eight ordinary source gates.
  Native table previews include the underlying original-object diagnostics.
• Keep every raw table, selected text, frozen quality ledger and consumer pin.
  Passing proof produces a separate immutable L1 version and re-runs its
  registered decoder. Real grids and ambiguous candidates remain held.
• Bounded source cases and mechanical refusal controls do not establish general
  table accuracy, semantic interpretation or legal/publication acceptance.

0.2.0-rc.11 — 2026-11-02

• Add admit-cached-l1 to verify and copy a pinned legacy pdf-extract reading
  into an immutable native package without PDF text extraction, OCR or models.
• Preserve original and cached file hashes, page text/status, table geometry and
  historical receipts; expose cached page dimensions and explicit missing facets.
• Re-derive admission artifacts during ordinary native verification. Configured
  fields can read exact cached spans while every page retains a current source
  proof hold. Migration integrity grants no source or legal acceptance.

0.2.0-rc.10 — 2026-11-01

• Add semantic-fields prepare/run/validate over verified native multipart
  fields using AskMill's opt-in request/result v2 contract. Models select IDs and
  configured roles; PDF2Data projects exact source parts and retains quality holds.
• Require complete schema-valid responses, selected-atom accounting and pinned
  source/profile/request re-derivation. Runs verify AskMill receipts and settlement;
  failed raw output stays available without repair or provider fallback.
• Include the CERC Sharing supplementary-draft field rules and JSON checks.
  The real-source development run covers seven fields and eight source parts;
  role membership/coverage checks do not establish general semantic accuracy,
  legal acceptance, eligibility or financial correctness.

0.2.0-rc.9 — 2026-11-01

• Opt-in fields v3 matches within retained native source units and preserves
  ordered value/capture parts across physical pages. Parent identity, full quality
  observations and page seams remain explicit; separate parents are never joined.
• Native re-derivation rejects modified multipart evidence. Exact source labels
  and JSON check rules evaluate v3; MCP page filtering retains complete records.
• Three development cases cover October COD/GNA, July deviation wording and July
  procedure continuation. These do not establish corpus, legal or formula accuracy.
• V1/v2 interpretation stays page-local. Existing bundles keep captured runtimes;
  consumers explicitly adopt the new wire. Source-image witnesses remain v2-only.

0.2.0-rc.8 — 2026-11-01

• Add fields check with versioned JSON assertions for field counts, literals,
  source pages, capture bindings and forbidden fragments. New cases require no
  Python changes; source re-derivation remains mandatory.

• Freeze check configuration and report per-assertion expected/actual results,
  uncovered fields and source/rule/evaluator identities. Passing checks does not
  grant source, formula, legal or calculation approval.

• Include the source-main field captures and source-image witness work developed
  after rc.1. Older bundles retain their captured runtimes; adoption requires
  explicit compatibility validation.

• Versioned JSON extraction rules: initialize, catalog, extract, search and evaluate
  new fields across document categories without per-field parser changes.

• Exact source/ruleset provenance, typed date/quantity candidates and bounded regex
  workers; reviewed-reference evaluation reports precision, recall and errors.

• Optional --rules / PDF2DATA_RULES integration in readable exports, workspace
  processing and MCP fields reads. No additional runtime dependency.

0.2.0-rc.1 — 2026-10-31

First packaged release candidate of PDF2Data. Existing local_vector_batch imports
and the legacy CLI entry point remain compatible.

• Full-document Markdown, HTML, JSON, optional Word and source-linked layout views.
• Compact local upload/workspace with live stages and artifact downloads.
• Source-linked document roles, headers/footers, bilingual metadata and worked-formula candidates.
• Complete physical source-location export, including empty and whitespace readings.
• Native raster evidence and nested-table geometry safeguards; prior outputs remain snapshots.
• Optional MCP server: parse, status and bounded source reads over the same CLI pipeline.
• An agent skill and a small npm launcher, with no automatic AI-model downloads.
• Optional DuckDB/NetworkX profiling example and Cytoscape-compatible graph export.

This is a prerelease. Parsing quality and legal/factual review remain separate.
See the benchmark receipts for the scope of measured re...

PDF2Data v0.2.0-rc.19

Preserve complete fact and structure candidates in processed/readable exports, with source spans, artifact hashes and an explicitly labelled export snapshot. Page HTML, Markdown and overlays now distinguish selection status from package quality; repair holds remain.

Validation: 1,664 source tests passed (8 skipped) before release preparation; 49 focused export/proof controls and five npm tests passed after preparation. Independent review passed. A clean installed wheel reproduced the real two-page export, retaining all prior fields, five fact candidates, 71 structure candidates and both page-repair holds.

The exact rc18 custody decoder remains registered for the existing supported operations. Existing packages, jobs and consumer pins are unchanged; consumers must explicitly qualify and adopt this release. No new OCR, model calls or source/legal acceptance.

GitHub artifacts are published independently of npm registry publication. SHA256 checksums cover all downloadable assets.

PDF2Data v0.2.0-rc.18

Preserve legacy-font review across PDF capture, quality assessment and machine decisions. Exact KrutiDev010 source fonts now require Hindi glyph-map proof even when native readers agree on Latin codes; source text is unchanged.

Known font/inventory holds cannot be silently cleared. Exact rc17 decoder custody remains supported for registered operations, with changed-code refusals and independently replayed source/table evidence.

Validation:1656 Python tests passed,8 skipped;5 npm tests;41 independent font controls;35 installed decoder controls;two frozen real-source custody replays. Installed CLI identified83 actual legacy glyphs and retained the page hold.

This release does not decode Hindi, raise a global benchmark score or grant legal/client/current-policy acceptance. Existing jobs and source proofs require explicit consumer adoption. Use the SHA256-pinned GitHub assets; npm registry publication is separate.

PDF2Data v0.2.0-rc.17

Native custody, targeted proof and table review workers now select fresh private Python bytecode-cache prefixes explicitly in isolated mode. Cleanup and existing timeouts remain bounded. Exact released rc16 decoders remain supported under the existing source, parent, operation and dependency checks.

Local validation: 1,629 tests passed and 16 optional/configuration cases skipped; 5 npm launcher tests passed. Independent benign child-process probes checked isolated mode, fresh prefixes, disabled cache writes and cleanup. Existing originals, failed receipts and consumer pins remain unchanged. Source-quality proof does not grant legal, semantic, publication or Query acceptance.

Assets include the Python wheel/source distribution, npm launcher tarball, agent skill and SHA-256 checksums. Verify asset hashes before installation. npm registry publication is tracked separately; this release does not claim registry availability.

PDF2Data v0.2.0-rc.16

Native grouping and footnote source proofs now bind exact independent paths, transforms, paint and isolated-object raster pixels. Foreign halo ink and real glyph intersections remain held; group/header meaning stays candidate and blank cells stay blank.

Local validation: 1,620 tests passed, 16 optional/configuration cases skipped; 5 npm launcher tests passed. The independently reviewed source change preserves original packages, failed receipts and consumer pins. Source-quality proof does not grant legal, numeric, publication or Query acceptance.

Assets include the Python wheel/source distribution, npm launcher tarball, agent skill and SHA-256 checksums. npm registry publication is pending local publisher authentication; this release does not claim registry availability.

PDF2Data v0.2.0-rc.15

Changelog

Unreleased

0.2.0-rc.15 — 2026-11-02

• Add bounded named local image-reading candidates with exact original/prompt/profile receipts. Preserve older readings, retain incomplete output and all source-quality holds, and leave selected L1s unchanged.
• Add scoped native source-object/raster accounting without transcription,
  language, layout or page-quality clearance.
• Compare configured multiword operators across explicit layout spacing while
  preserving raw spans, separate numeric checks and frozen quality outcomes.
• Retain the exact released rc14 proof decoder alongside rc12/rc13; require all
  new command modules in distributable-wheel checks.

0.2.0-rc.14 — 2026-11-02

• Add explicit bounded English cardinal and printed-percent quantity modes.
  Preserve literal/source spans, comparison context and quality holds; keep
  configured unit hints separate from anchored printed units.
• Add exact rational/null quantity assertions to field checks and enforce the
  quantity-only normalization boundary in the public extraction-rule schema.
• Retain exact released rc12 and rc13 custody decoder compatibility alongside
  current code, with the existing operation, source, dependency and hash gates.
• Qualify seven retained real-source word quantities and separate printed-unit
  witnesses. This does not establish legal, calculation or pipeline acceptance.

0.2.0-rc.13 — 2026-11-02

• Replay frozen native proofs with explicit hash-checked custody bindings and an exact captured decoder closure; preserve captured provenance and reject substituted parents or external review inputs.

• Support native table source proofs from installed wheels by requiring all
  proof modules to share the actual current package directory. Mixed modules
  remain refused; captured extraction origins stay data rather than import paths.

• Refuse proof outputs overlapping source packages, extraction runtimes or input
  receipts. Keep incomplete cell review as an explicit terminal hold without
  creating a quality version.

• Qualify the installed route on bounded source cases with existing independent
  cell/region decoders. Preserve raw grids, source text, numeric/operator holds
  and all frozen consumer artifacts; no source acceptance rules are relaxed.

0.2.0-rc.12 — 2026-11-02

• Expose native-whitefill-prose-reproof over the existing registered source
  proof route. Extend its closed predicates to account for thin, long footer
  rules disjoint from glyphs, alongside exact white paragraph backgrounds.
• Require exact single-column candidate/cell/rectangle ownership, fresh paint
  visibility, independent source lines and all eight ordinary source gates.
  Native table previews include the underlying original-object diagnostics.
• Keep every raw table, selected text, frozen quality ledger and consumer pin.
  Passing proof produces a separate immutable L1 version and re-runs its
  registered decoder. Real grids and ambiguous candidates remain held.
• Bounded source cases and mechanical refusal controls do not establish general
  table accuracy, semantic interpretation or legal/publication acceptance.

0.2.0-rc.11 — 2026-11-02

• Add admit-cached-l1 to verify and copy a pinned legacy pdf-extract reading
  into an immutable native package without PDF text extraction, OCR or models.
• Preserve original and cached file hashes, page text/status, table geometry and
  historical receipts; expose cached page dimensions and explicit missing facets.
• Re-derive admission artifacts during ordinary native verification. Configured
  fields can read exact cached spans while every page retains a current source
  proof hold. Migration integrity grants no source or legal acceptance.

0.2.0-rc.10 — 2026-11-01

• Add semantic-fields prepare/run/validate over verified native multipart
  fields using AskMill's opt-in request/result v2 contract. Models select IDs and
  configured roles; PDF2Data projects exact source parts and retains quality holds.
• Require complete schema-valid responses, selected-atom accounting and pinned
  source/profile/request re-derivation. Runs verify AskMill receipts and settlement;
  failed raw output stays available without repair or provider fallback.
• Include the CERC Sharing supplementary-draft field rules and JSON checks.
  The real-source development run covers seven fields and eight source parts;
  role membership/coverage checks do not establish general semantic accuracy,
  legal acceptance, eligibility or financial correctness.

0.2.0-rc.9 — 2026-11-01

• Opt-in fields v3 matches within retained native source units and preserves
  ordered value/capture parts across physical pages. Parent identity, full quality
  observations and page seams remain explicit; separate parents are never joined.
• Native re-derivation rejects modified multipart evidence. Exact source labels
  and JSON check rules evaluate v3; MCP page filtering retains complete records.
• Three development cases cover October COD/GNA, July deviation wording and July
  procedure continuation. These do not establish corpus, legal or formula accuracy.
• V1/v2 interpretation stays page-local. Existing bundles keep captured runtimes;
  consumers explicitly adopt the new wire. Source-image witnesses remain v2-only.

0.2.0-rc.8 — 2026-11-01

• Add fields check with versioned JSON assertions for field counts, literals,
  source pages, capture bindings and forbidden fragments. New cases require no
  Python changes; source re-derivation remains mandatory.

• Freeze check configuration and report per-assertion expected/actual results,
  uncovered fields and source/rule/evaluator identities. Passing checks does not
  grant source, formula, legal or calculation approval.

• Include the source-main field captures and source-image witness work developed
  after rc.1. Older bundles retain their captured runtimes; adoption requires
  explicit compatibility validation.

• Versioned JSON extraction rules: initialize, catalog, extract, search and evaluate
  new fields across document categories without per-field parser changes.

• Exact source/ruleset provenance, typed date/quantity candidates and bounded regex
  workers; reviewed-reference evaluation reports precision, recall and errors.

• Optional --rules / PDF2DATA_RULES integration in readable exports, workspace
  processing and MCP fields reads. No additional runtime dependency.

0.2.0-rc.1 — 2026-10-31

First packaged release candidate of PDF2Data. Existing local_vector_batch imports
and the legacy CLI entry point remain compatible.

• Full-document Markdown, HTML, JSON, optional Word and source-linked layout views.
• Compact local upload/workspace with live stages and artifact downloads.
• Source-linked document roles, headers/footers, bilingual metadata and worked-formula candidates.
• Complete physical source-location export, including empty and whitespace readings.
• Native raster evidence and nested-table geometry safeguards; prior outputs remain snapshots.
• Optional MCP server: parse, status and bounded source reads over the same CLI pipeline.
• An agent skill and a small npm launcher, with no automatic AI-model downloads.
• Optional DuckDB/NetworkX profiling example and Cytoscape-compatible graph export.

This is a prerelease. Parsing quality and legal/factual review remain separate.
See the benchmark receipts for the scope of measured results; this release does
not claim a new OCR accuracy score.

Earlier development notes

• Added resumable source-bound page repair queues, explicit worker-crash retry
  states, model/policy-version attempt keys, and current-page status separate
  from historical attempt counts.
• Replaced fixed OCR worker groups with configurable small page jobs and
  persistent recognizers per process; completion counts unique source/page/text
  receipts and validates full requested coverage.
• Added atomic multi-page repair commits with fresh source quality checks,
  original-version checks, source-review precedence and untouched-page checks.
  Committing local evidence does not declare a remote search index ready.
• Added optional local PDF extraction and a loopback-only drop/review/configure/
  embed/export webpage, including every-page review and bounded advanced
  encoder controls.
• Added local macOS Vision / installed Tesseract OCR, cached Vision compilation,
  garbled native-text detection, retained native/OCR evidence, and explicit
  unreadable/low-confidence blockers.
• Preserved source/page/text hashes, table bounding boxes and blank cells, and
  exact spans for unverified value candidates.
• Added exact pre-windowed passages input with owning-chunk offsets and a new
  vector-ID namespace.
• Replaced silent tokenizer truncation with overflow errors in both encoders;
  added effective tokenizer receipts and input-record coverage audits.
• Added local typed-candidate evidence validation / CLI for page hashes, exact
  or explicit whitespace-normalized quote spans, units and role enums. Table
  relationships and semantic roles require review; no facts are auto-verified.
• Added optional tokenizer-offset overflow splitting (default in PDF web jobs),
  bounded token overlap, tokenizer hashes and stable child IDs; safe windows and
  legacy error-policy fingerprints remain unchanged.
• Fixed the module entry point for spawned CPU workers; validated real two-worker
  CPU and Mac GPU runs. Added tests for offset fidelity, coverage gaps, garbled
  text recovery, low-confidence OCR, token overflow and invalid web controls.

Remaining: scanned table row/column reconstruction, institution/date role
validation, clause hierarchy, Windows/Tesseract runtime validation, and broader
extraction/retrieval quality benchmarks. OCR confidence and file integrity do
not establish factual accuracy.

PDF2Data v0.2.0-rc.9

Changelog

Unreleased

0.2.0-rc.9 — 2026-11-01

• Opt-in fields v3 matches within retained native source units and preserves
  ordered value/capture parts across physical pages. Parent identity, full quality
  observations and page seams remain explicit; separate parents are never joined.
• Native re-derivation rejects modified multipart evidence. Exact source labels
  and JSON check rules evaluate v3; MCP page filtering retains complete records.
• Three development cases cover October COD/GNA, July deviation wording and July
  procedure continuation. These do not establish corpus, legal or formula accuracy.
• V1/v2 interpretation stays page-local. Existing bundles keep captured runtimes;
  consumers explicitly adopt the new wire. Source-image witnesses remain v2-only.

0.2.0-rc.8 — 2026-11-01

• Add fields check with versioned JSON assertions for field counts, literals,
  source pages, capture bindings and forbidden fragments. New cases require no
  Python changes; source re-derivation remains mandatory.

• Freeze check configuration and report per-assertion expected/actual results,
  uncovered fields and source/rule/evaluator identities. Passing checks does not
  grant source, formula, legal or calculation approval.

• Include the source-main field captures and source-image witness work developed
  after rc.1. Older bundles retain their captured runtimes; adoption requires
  explicit compatibility validation.

• Versioned JSON extraction rules: initialize, catalog, extract, search and evaluate
  new fields across document categories without per-field parser changes.

• Exact source/ruleset provenance, typed date/quantity candidates and bounded regex
  workers; reviewed-reference evaluation reports precision, recall and errors.

• Optional --rules / PDF2DATA_RULES integration in readable exports, workspace
  processing and MCP fields reads. No additional runtime dependency.

0.2.0-rc.1 — 2026-10-31

First packaged release candidate of PDF2Data. Existing local_vector_batch imports
and the legacy CLI entry point remain compatible.

• Full-document Markdown, HTML, JSON, optional Word and source-linked layout views.
• Compact local upload/workspace with live stages and artifact downloads.
• Source-linked document roles, headers/footers, bilingual metadata and worked-formula candidates.
• Complete physical source-location export, including empty and whitespace readings.
• Native raster evidence and nested-table geometry safeguards; prior outputs remain snapshots.
• Optional MCP server: parse, status and bounded source reads over the same CLI pipeline.
• An agent skill and a small npm launcher, with no automatic AI-model downloads.
• Optional DuckDB/NetworkX profiling example and Cytoscape-compatible graph export.

This is a prerelease. Parsing quality and legal/factual review remain separate.
See the benchmark receipts for the scope of measured results; this release does
not claim a new OCR accuracy score.

Earlier development notes

• Added resumable source-bound page repair queues, explicit worker-crash retry
  states, model/policy-version attempt keys, and current-page status separate
  from historical attempt counts.
• Replaced fixed OCR worker groups with configurable small page jobs and
  persistent recognizers per process; completion counts unique source/page/text
  receipts and validates full requested coverage.
• Added atomic multi-page repair commits with fresh source quality checks,
  original-version checks, source-review precedence and untouched-page checks.
  Committing local evidence does not declare a remote search index ready.
• Added optional local PDF extraction and a loopback-only drop/review/configure/
  embed/export webpage, including every-page review and bounded advanced
  encoder controls.
• Added local macOS Vision / installed Tesseract OCR, cached Vision compilation,
  garbled native-text detection, retained native/OCR evidence, and explicit
  unreadable/low-confidence blockers.
• Preserved source/page/text hashes, table bounding boxes and blank cells, and
  exact spans for unverified value candidates.
• Added exact pre-windowed passages input with owning-chunk offsets and a new
  vector-ID namespace.
• Replaced silent tokenizer truncation with overflow errors in both encoders;
  added effective tokenizer receipts and input-record coverage audits.
• Added local typed-candidate evidence validation / CLI for page hashes, exact
  or explicit whitespace-normalized quote spans, units and role enums. Table
  relationships and semantic roles require review; no facts are auto-verified.
• Added optional tokenizer-offset overflow splitting (default in PDF web jobs),
  bounded token overlap, tokenizer hashes and stable child IDs; safe windows and
  legacy error-policy fingerprints remain unchanged.
• Fixed the module entry point for spawned CPU workers; validated real two-worker
  CPU and Mac GPU runs. Added tests for offset fidelity, coverage gaps, garbled
  text recovery, low-confidence OCR, token overflow and invalid web controls.

Remaining: scanned table row/column reconstruction, institution/date role
validation, clause hierarchy, Windows/Tesseract runtime validation, and broader
extraction/retrieval quality benchmarks. OCR confidence and file integrity do
not establish factual accuracy.

PDF2Data v0.2.0-rc.8

Changelog

Unreleased

0.2.0-rc.8 — 2026-11-01

• Add fields check with versioned JSON assertions for field counts, literals,
  source pages, capture bindings and forbidden fragments. New cases require no
  Python changes; source re-derivation remains mandatory.

• Freeze check configuration and report per-assertion expected/actual results,
  uncovered fields and source/rule/evaluator identities. Passing checks does not
  grant source, formula, legal or calculation approval.

• Include the source-main field captures and source-image witness work developed
  after rc.1. Older bundles retain their captured runtimes; adoption requires
  explicit compatibility validation.

• Versioned JSON extraction rules: initialize, catalog, extract, search and evaluate
  new fields across document categories without per-field parser changes.

• Exact source/ruleset provenance, typed date/quantity candidates and bounded regex
  workers; reviewed-reference evaluation reports precision, recall and errors.

• Optional --rules / PDF2DATA_RULES integration in readable exports, workspace
  processing and MCP fields reads. No additional runtime dependency.

0.2.0-rc.1 — 2026-10-31

First packaged release candidate of PDF2Data. Existing local_vector_batch imports
and the legacy CLI entry point remain compatible.

• Full-document Markdown, HTML, JSON, optional Word and source-linked layout views.
• Compact local upload/workspace with live stages and artifact downloads.
• Source-linked document roles, headers/footers, bilingual metadata and worked-formula candidates.
• Complete physical source-location export, including empty and whitespace readings.
• Native raster evidence and nested-table geometry safeguards; prior outputs remain snapshots.
• Optional MCP server: parse, status and bounded source reads over the same CLI pipeline.
• An agent skill and a small npm launcher, with no automatic AI-model downloads.
• Optional DuckDB/NetworkX profiling example and Cytoscape-compatible graph export.

This is a prerelease. Parsing quality and legal/factual review remain separate.
See the benchmark receipts for the scope of measured results; this release does
not claim a new OCR accuracy score.

Earlier development notes

• Added resumable source-bound page repair queues, explicit worker-crash retry
  states, model/policy-version attempt keys, and current-page status separate
  from historical attempt counts.
• Replaced fixed OCR worker groups with configurable small page jobs and
  persistent recognizers per process; completion counts unique source/page/text
  receipts and validates full requested coverage.
• Added atomic multi-page repair commits with fresh source quality checks,
  original-version checks, source-review precedence and untouched-page checks.
  Committing local evidence does not declare a remote search index ready.
• Added optional local PDF extraction and a loopback-only drop/review/configure/
  embed/export webpage, including every-page review and bounded advanced
  encoder controls.
• Added local macOS Vision / installed Tesseract OCR, cached Vision compilation,
  garbled native-text detection, retained native/OCR evidence, and explicit
  unreadable/low-confidence blockers.
• Preserved source/page/text hashes, table bounding boxes and blank cells, and
  exact spans for unverified value candidates.
• Added exact pre-windowed passages input with owning-chunk offsets and a new
  vector-ID namespace.
• Replaced silent tokenizer truncation with overflow errors in both encoders;
  added effective tokenizer receipts and input-record coverage audits.
• Added local typed-candidate evidence validation / CLI for page hashes, exact
  or explicit whitespace-normalized quote spans, units and role enums. Table
  relationships and semantic roles require review; no facts are auto-verified.
• Added optional tokenizer-offset overflow splitting (default in PDF web jobs),
  bounded token overlap, tokenizer hashes and stable child IDs; safe windows and
  legacy error-policy fingerprints remain unchanged.
• Fixed the module entry point for spawned CPU workers; validated real two-worker
  CPU and Mac GPU runs. Added tests for offset fidelity, coverage gaps, garbled
  text recovery, low-confidence OCR, token overflow and invalid web controls.

Remaining: scanned table row/column reconstruction, institution/date role
validation, clause hierarchy, Windows/Tesseract runtime validation, and broader
extraction/retrieval quality benchmarks. OCR confidence and file integrity do
not establish factual accuracy.

PDF2Data v0.2.0-rc.14

Changelog

Unreleased

0.2.0-rc.14 — 2026-11-02

• Add explicit bounded English cardinal and printed-percent quantity modes.
  Preserve literal/source spans, comparison context and quality holds; keep
  configured unit hints separate from anchored printed units.
• Add exact rational/null quantity assertions to field checks and enforce the
  quantity-only normalization boundary in the public extraction-rule schema.
• Retain exact released rc12 and rc13 custody decoder compatibility alongside
  current code, with the existing operation, source, dependency and hash gates.
• Qualify seven retained real-source word quantities and separate printed-unit
  witnesses. This does not establish legal, calculation or pipeline acceptance.

0.2.0-rc.13 — 2026-11-02

• Replay frozen native proofs with explicit hash-checked custody bindings and an exact captured decoder closure; preserve captured provenance and reject substituted parents or external review inputs.

• Support native table source proofs from installed wheels by requiring all
  proof modules to share the actual current package directory. Mixed modules
  remain refused; captured extraction origins stay data rather than import paths.

• Refuse proof outputs overlapping source packages, extraction runtimes or input
  receipts. Keep incomplete cell review as an explicit terminal hold without
  creating a quality version.

• Qualify the installed route on bounded source cases with existing independent
  cell/region decoders. Preserve raw grids, source text, numeric/operator holds
  and all frozen consumer artifacts; no source acceptance rules are relaxed.

0.2.0-rc.12 — 2026-11-02

• Expose native-whitefill-prose-reproof over the existing registered source
  proof route. Extend its closed predicates to account for thin, long footer
  rules disjoint from glyphs, alongside exact white paragraph backgrounds.
• Require exact single-column candidate/cell/rectangle ownership, fresh paint
  visibility, independent source lines and all eight ordinary source gates.
  Native table previews include the underlying original-object diagnostics.
• Keep every raw table, selected text, frozen quality ledger and consumer pin.
  Passing proof produces a separate immutable L1 version and re-runs its
  registered decoder. Real grids and ambiguous candidates remain held.
• Bounded source cases and mechanical refusal controls do not establish general
  table accuracy, semantic interpretation or legal/publication acceptance.

0.2.0-rc.11 — 2026-11-02

• Add admit-cached-l1 to verify and copy a pinned legacy pdf-extract reading
  into an immutable native package without PDF text extraction, OCR or models.
• Preserve original and cached file hashes, page text/status, table geometry and
  historical receipts; expose cached page dimensions and explicit missing facets.
• Re-derive admission artifacts during ordinary native verification. Configured
  fields can read exact cached spans while every page retains a current source
  proof hold. Migration integrity grants no source or legal acceptance.

0.2.0-rc.10 — 2026-11-01

• Add semantic-fields prepare/run/validate over verified native multipart
  fields using AskMill's opt-in request/result v2 contract. Models select IDs and
  configured roles; PDF2Data projects exact source parts and retains quality holds.
• Require complete schema-valid responses, selected-atom accounting and pinned
  source/profile/request re-derivation. Runs verify AskMill receipts and settlement;
  failed raw output stays available without repair or provider fallback.
• Include the CERC Sharing supplementary-draft field rules and JSON checks.
  The real-source development run covers seven fields and eight source parts;
  role membership/coverage checks do not establish general semantic accuracy,
  legal acceptance, eligibility or financial correctness.

0.2.0-rc.9 — 2026-11-01

• Opt-in fields v3 matches within retained native source units and preserves
  ordered value/capture parts across physical pages. Parent identity, full quality
  observations and page seams remain explicit; separate parents are never joined.
• Native re-derivation rejects modified multipart evidence. Exact source labels
  and JSON check rules evaluate v3; MCP page filtering retains complete records.
• Three development cases cover October COD/GNA, July deviation wording and July
  procedure continuation. These do not establish corpus, legal or formula accuracy.
• V1/v2 interpretation stays page-local. Existing bundles keep captured runtimes;
  consumers explicitly adopt the new wire. Source-image witnesses remain v2-only.

0.2.0-rc.8 — 2026-11-01

• Add fields check with versioned JSON assertions for field counts, literals,
  source pages, capture bindings and forbidden fragments. New cases require no
  Python changes; source re-derivation remains mandatory.

• Freeze check configuration and report per-assertion expected/actual results,
  uncovered fields and source/rule/evaluator identities. Passing checks does not
  grant source, formula, legal or calculation approval.

• Include the source-main field captures and source-image witness work developed
  after rc.1. Older bundles retain their captured runtimes; adoption requires
  explicit compatibility validation.

• Versioned JSON extraction rules: initialize, catalog, extract, search and evaluate
  new fields across document categories without per-field parser changes.

• Exact source/ruleset provenance, typed date/quantity candidates and bounded regex
  workers; reviewed-reference evaluation reports precision, recall and errors.

• Optional --rules / PDF2DATA_RULES integration in readable exports, workspace
  processing and MCP fields reads. No additional runtime dependency.

0.2.0-rc.1 — 2026-10-31

First packaged release candidate of PDF2Data. Existing local_vector_batch imports
and the legacy CLI entry point remain compatible.

• Full-document Markdown, HTML, JSON, optional Word and source-linked layout views.
• Compact local upload/workspace with live stages and artifact downloads.
• Source-linked document roles, headers/footers, bilingual metadata and worked-formula candidates.
• Complete physical source-location export, including empty and whitespace readings.
• Native raster evidence and nested-table geometry safeguards; prior outputs remain snapshots.
• Optional MCP server: parse, status and bounded source reads over the same CLI pipeline.
• An agent skill and a small npm launcher, with no automatic AI-model downloads.
• Optional DuckDB/NetworkX profiling example and Cytoscape-compatible graph export.

This is a prerelease. Parsing quality and legal/factual review remain separate.
See the benchmark receipts for the scope of measured results; this release does
not claim a new OCR accuracy score.

Earlier development notes

• Added resumable source-bound page repair queues, explicit worker-crash retry
  states, model/policy-version attempt keys, and current-page status separate
  from historical attempt counts.
• Replaced fixed OCR worker groups with configurable small page jobs and
  persistent recognizers per process; completion counts unique source/page/text
  receipts and validates full requested coverage.
• Added atomic multi-page repair commits with fresh source quality checks,
  original-version checks, source-review precedence and untouched-page checks.
  Committing local evidence does not declare a remote search index ready.
• Added optional local PDF extraction and a loopback-only drop/review/configure/
  embed/export webpage, including every-page review and bounded advanced
  encoder controls.
• Added local macOS Vision / installed Tesseract OCR, cached Vision compilation,
  garbled native-text detection, retained native/OCR evidence, and explicit
  unreadable/low-confidence blockers.
• Preserved source/page/text hashes, table bounding boxes and blank cells, and
  exact spans for unverified value candidates.
• Added exact pre-windowed passages input with owning-chunk offsets and a new
  vector-ID namespace.
• Replaced silent tokenizer truncation with overflow errors in both encoders;
  added effective tokenizer receipts and input-record coverage audits.
• Added local typed-candidate evidence validation / CLI for page hashes, exact
  or explicit whitespace-normalized quote spans, units and role enums. Table
  relationships and semantic roles require review; no facts are auto-verified.
• Added optional tokenizer-offset overflow splitting (default in PDF web jobs),
  bounded token overlap, tokenizer hashes and stable child IDs; safe windows and
  legacy error-policy fingerprints remain unchanged.
• Fixed the module entry point for spawned CPU workers; validated real two-worker
  CPU and Mac GPU runs. Added tests for offset fidelity, coverage gaps, garbled
  text recovery, low-confidence OCR, token overflow and invalid web controls.

Remaining: scanned table row/column reconstruction, institution/date role
validation, clause hierarchy, Windows/Tesseract runtime validation, and broader
extraction/retrieval quality benchmarks. OCR confidence and file integrity do
not establish factual accuracy.

PDF2Data v0.2.0-rc.13

Changelog

Unreleased

0.2.0-rc.13 — 2026-11-02

• Replay frozen native proofs with explicit hash-checked custody bindings and an exact captured decoder closure; preserve captured provenance and reject substituted parents or external review inputs.

• Support native table source proofs from installed wheels by requiring all
  proof modules to share the actual current package directory. Mixed modules
  remain refused; captured extraction origins stay data rather than import paths.

• Refuse proof outputs overlapping source packages, extraction runtimes or input
  receipts. Keep incomplete cell review as an explicit terminal hold without
  creating a quality version.

• Qualify the installed route on bounded source cases with existing independent
  cell/region decoders. Preserve raw grids, source text, numeric/operator holds
  and all frozen consumer artifacts; no source acceptance rules are relaxed.

0.2.0-rc.12 — 2026-11-02

• Expose native-whitefill-prose-reproof over the existing registered source
  proof route. Extend its closed predicates to account for thin, long footer
  rules disjoint from glyphs, alongside exact white paragraph backgrounds.
• Require exact single-column candidate/cell/rectangle ownership, fresh paint
  visibility, independent source lines and all eight ordinary source gates.
  Native table previews include the underlying original-object diagnostics.
• Keep every raw table, selected text, frozen quality ledger and consumer pin.
  Passing proof produces a separate immutable L1 version and re-runs its
  registered decoder. Real grids and ambiguous candidates remain held.
• Bounded source cases and mechanical refusal controls do not establish general
  table accuracy, semantic interpretation or legal/publication acceptance.

0.2.0-rc.11 — 2026-11-02

• Add admit-cached-l1 to verify and copy a pinned legacy pdf-extract reading
  into an immutable native package without PDF text extraction, OCR or models.
• Preserve original and cached file hashes, page text/status, table geometry and
  historical receipts; expose cached page dimensions and explicit missing facets.
• Re-derive admission artifacts during ordinary native verification. Configured
  fields can read exact cached spans while every page retains a current source
  proof hold. Migration integrity grants no source or legal acceptance.

0.2.0-rc.10 — 2026-11-01

• Add semantic-fields prepare/run/validate over verified native multipart
  fields using AskMill's opt-in request/result v2 contract. Models select IDs and
  configured roles; PDF2Data projects exact source parts and retains quality holds.
• Require complete schema-valid responses, selected-atom accounting and pinned
  source/profile/request re-derivation. Runs verify AskMill receipts and settlement;
  failed raw output stays available without repair or provider fallback.
• Include the CERC Sharing supplementary-draft field rules and JSON checks.
  The real-source development run covers seven fields and eight source parts;
  role membership/coverage checks do not establish general semantic accuracy,
  legal acceptance, eligibility or financial correctness.

0.2.0-rc.9 — 2026-11-01

• Opt-in fields v3 matches within retained native source units and preserves
  ordered value/capture parts across physical pages. Parent identity, full quality
  observations and page seams remain explicit; separate parents are never joined.
• Native re-derivation rejects modified multipart evidence. Exact source labels
  and JSON check rules evaluate v3; MCP page filtering retains complete records.
• Three development cases cover October COD/GNA, July deviation wording and July
  procedure continuation. These do not establish corpus, legal or formula accuracy.
• V1/v2 interpretation stays page-local. Existing bundles keep captured runtimes;
  consumers explicitly adopt the new wire. Source-image witnesses remain v2-only.

0.2.0-rc.8 — 2026-11-01

• Add fields check with versioned JSON assertions for field counts, literals,
  source pages, capture bindings and forbidden fragments. New cases require no
  Python changes; source re-derivation remains mandatory.

• Freeze check configuration and report per-assertion expected/actual results,
  uncovered fields and source/rule/evaluator identities. Passing checks does not
  grant source, formula, legal or calculation approval.

• Include the source-main field captures and source-image witness work developed
  after rc.1. Older bundles retain their captured runtimes; adoption requires
  explicit compatibility validation.

• Versioned JSON extraction rules: initialize, catalog, extract, search and evaluate
  new fields across document categories without per-field parser changes.

• Exact source/ruleset provenance, typed date/quantity candidates and bounded regex
  workers; reviewed-reference evaluation reports precision, recall and errors.

• Optional --rules / PDF2DATA_RULES integration in readable exports, workspace
  processing and MCP fields reads. No additional runtime dependency.

0.2.0-rc.1 — 2026-10-31

First packaged release candidate of PDF2Data. Existing local_vector_batch imports
and the legacy CLI entry point remain compatible.

• Full-document Markdown, HTML, JSON, optional Word and source-linked layout views.
• Compact local upload/workspace with live stages and artifact downloads.
• Source-linked document roles, headers/footers, bilingual metadata and worked-formula candidates.
• Complete physical source-location export, including empty and whitespace readings.
• Native raster evidence and nested-table geometry safeguards; prior outputs remain snapshots.
• Optional MCP server: parse, status and bounded source reads over the same CLI pipeline.
• An agent skill and a small npm launcher, with no automatic AI-model downloads.
• Optional DuckDB/NetworkX profiling example and Cytoscape-compatible graph export.

This is a prerelease. Parsing quality and legal/factual review remain separate.
See the benchmark receipts for the scope of measured results; this release does
not claim a new OCR accuracy score.

Earlier development notes

• Added resumable source-bound page repair queues, explicit worker-crash retry
  states, model/policy-version attempt keys, and current-page status separate
  from historical attempt counts.
• Replaced fixed OCR worker groups with configurable small page jobs and
  persistent recognizers per process; completion counts unique source/page/text
  receipts and validates full requested coverage.
• Added atomic multi-page repair commits with fresh source quality checks,
  original-version checks, source-review precedence and untouched-page checks.
  Committing local evidence does not declare a remote search index ready.
• Added optional local PDF extraction and a loopback-only drop/review/configure/
  embed/export webpage, including every-page review and bounded advanced
  encoder controls.
• Added local macOS Vision / installed Tesseract OCR, cached Vision compilation,
  garbled native-text detection, retained native/OCR evidence, and explicit
  unreadable/low-confidence blockers.
• Preserved source/page/text hashes, table bounding boxes and blank cells, and
  exact spans for unverified value candidates.
• Added exact pre-windowed passages input with owning-chunk offsets and a new
  vector-ID namespace.
• Replaced silent tokenizer truncation with overflow errors in both encoders;
  added effective tokenizer receipts and input-record coverage audits.
• Added local typed-candidate evidence validation / CLI for page hashes, exact
  or explicit whitespace-normalized quote spans, units and role enums. Table
  relationships and semantic roles require review; no facts are auto-verified.
• Added optional tokenizer-offset overflow splitting (default in PDF web jobs),
  bounded token overlap, tokenizer hashes and stable child IDs; safe windows and
  legacy error-policy fingerprints remain unchanged.
• Fixed the module entry point for spawned CPU workers; validated real two-worker
  CPU and Mac GPU runs. Added tests for offset fidelity, coverage gaps, garbled
  text recovery, low-confidence OCR, token overflow and invalid web controls.

Remaining: scanned table row/column reconstruction, institution/date role
validation, clause hierarchy, Windows/Tesseract runtime validation, and broader
extraction/retrieval quality benchmarks. OCR confidence and file integrity do
not establish factual accuracy.