Inkform 0.4.4

Inkform 0.4.4

A naturally short j stem could make the following u fail the new size check in words such as “jump”. This patch compares dotted i/j stems locally with the same letter while keeping their independent source-bank checks. Digits, capitals, repeated letters, scripts, and clear size limits retain their checks.

Sampled vertical graphs also remain stable after layout changes. The renderer handles negligible coordinate roundoff without changing declared anchors, and the PDF checker reconstructs source traces in the same local coordinate frame used for drawing. Tiny closed spans retain their baseline samples.

The source contours and reviewed private handwriting calibration remain separate from the public engine. Existing profiles still need suitable reviewed samples and correct source measurements. No personal writing or assignments are included in these release assets.

Validation: 577 Python tests and 8 launcher tests pass. Packed installation, standalone skill, exported vector PDF, and character-size audit smoke checks pass. Linux/Python 3.11 and macOS/Python 3.13 CI pass. All five release archives pass public-content checks and contain the exact tested engine sources. A private 69-page assignment also passes export and layout checks; all pages and 166 graphs were visually reviewed. Personal input and output files remain excluded from the release.

npm install -g @nazifbd/inkform@0.4.4
inkform setup
inkform skill install --force

GitHub Packages provides the same engine under @user-github-me/inkform. The attached standalone skill includes its runtime; the npm archive supports direct installation.

v0.4.3 — Context-aware lettering and sampled graph strokes

Inkform 0.4.3

Characters with a wrong source scale could pass the previous size fallback, and densely sampled graphs could suppress the writer’s pen deviations. This release checks character size in context and gives computed graph samples an explicit role separate from exact data anchors.

• Select real forms using lowercase body, ascender, capital, digit, word-position, and text/math context. Compare with selected neighboring writing, including repeated letters and multi-digit numbers, and retain legitimate script sizing.
• Check placed ink before export and independently measure the actual PDF paths. Clear size mismatches stop export; uncertain estimates remain visible in the report.
• Use point_mode: "samples" with anchor_indices for computed curves and straight function segments. Reviewed source deviations continue between protected anchors, including horizontal and vertical strokes. Existing exact-point curves keep their behavior.
• Follow the exact source project/profile when refining a PDF, and review local lowercase reference heights before correcting source calibration.

Existing profiles with only incompatible forms need reviewed calibration repairs or additional genuine samples. Installing the engine does not replace private profiles. Source contours remain intact; the geometry checks do not certify handwriting identity or physical print quality.

Validation: all 563 Python tests and eight CLI tests passed. Packed npm/standalone skill PDF smoke, responsive website/clipboard checks, and public archive guards passed. Release assets contain only code, skill resources, and reviewed synthetic examples.

Install:

npm install -g @nazifbd/inkform@0.4.3
inkform setup
inkform skill install --force

The attached npm archive also supports direct installation. GitHub Packages provides the same engine under @user-github-me/inkform.

v0.4.2 — Readable writing and source-drawn graphs

Inkform 0.4.2 improves word grouping, contextual character selection, mathematical corrections, and graphs drawn from reviewed source strokes.

• Place letters by their body shapes while checking complete ink clearance, so a long descender cannot detach the last letter from its word. Learn visible word and letter gaps from reviewed source lines.
• Prefer context-compatible capitals and digits before balancing variants. Reviewed script context and normalization stay separate from estimated size comparisons.
• Connect root hooks and roofs at their actual endpoints. Use reviewed centerline mathematical primitives alongside calibrated outline writing.
• Support partial word/math attempts and explicit mathematical deletions, preserving the intended answer and recording estimated drafts separately from source evidence.
• Transfer reviewed graph axes, curves, ticks, and marker contours while preserving data anchors. Check emitted PDF geometry against the recorded source traces.
• Extend the skill's whole-page calibration and existing-PDF review workflow. Personal samples, profiles, assignment PDFs, and review proofs are excluded from public releases.

A finite source bank still limits variation. Missing capital/digit coverage needs real evidence; source traces are static ink, not recovered pen motion. Graph-element cross-outs and automatic reconstruction of a path-only assignment are not implemented.

The release includes the npm archive, Python wheel/source archive, and a self-contained agent skill. To install the npm archive directly:

npm install -g https://github.com/user-github-me/inkform/releases/download/v0.4.2/nazifbd-inkform-0.4.2.tgz
inkform setup
inkform skill install

Validation: 494 Python tests, 8 Node tests, a packed-package/installed-skill smoke test, and website clipboard/responsive checks passed. Private vector proofs verified source-shaped graph paths, joined root roofs, safe letter placement, and mathematical corrections. No personal proof or source page is included in this release.

Inkform v0.4.1 — Faithful source variants

Inkform now keeps the writer’s observed letter shapes intact. Default rendering selects real reviewed variants, excludes synthesized forms, and removes contour warps and invented letter, arrow, and graph bends. If the source bank runs out, the PDF reports reuse instead of changing the handwriting to force uniqueness.

Training scans the complete preserved corpus and all reviewed templates by default, with an uncapped candidate collection option. Source labels, context, and fitted vector geometry still require review. The skill and website prompt now explain this workflow; public examples use authored synthetic samples.

Validation: 407 Python tests, 8 Node launcher tests, a packed npm installation/render test, independent skill execution, and browser checks at 320–1440 px. PDF handwriting remains vector; physical print resemblance needs actual paper review.

Install this release directly: npm install -g https://github.com/user-github-me/inkform/releases/download/v0.4.1/nazifbd-inkform-0.4.1.tgz, then inkform setup and inkform skill install.

The public npm registry publication is pending an npm verification callback error. The attached archive is the tested npm package and installs without that publication step.

Inkform v0.4.0 — Reviewed styles and personal graphs

Inkform now compares repeated whole-word constructions, proposes extra letter styles from a writer's reviewed forms, and draws graphs with their labels, arrows, and measured axis style.

• Generate private candidates with profile synthesize or train --synthesize; review complete PDF/PNG atlases and approve exact geometry. Generated forms stay separate from observed samples.
• Reject repeated word constructions after removing letter scale and shear. The renderer searches approved alternatives and reports when a bank is insufficient.
• Remember clean or sample-based corrections and reuse the writer's reviewed strike/scribble styles.
• Preserve graph data anchors, ticks, axis direction, and open/filled markers. Actual PDF checks also verify the white interior of open markers.
• Start from the introduction site's copyable agent prompt and synthetic source/output example.

Personal samples, profiles, and generated documents are excluded from these downloads. The bundled examples are authored synthetic handwriting. Visual resemblance and actual-size print calibration remain part of the workflow.

Website: https://user-github-me.github.io/inkform/
Agent skill: https://github.com/user-github-me/inkform/tree/v0.4.0/skills/inkform

Inkform v0.3.0 — Distinct forms, checked on the page

Inkform 0.3.0 addresses letters and arrows that still looked repeated after simple size variation. It now compares actual emitted PDF ink across the same and neighboring pages, including characters shared by prose and math.

• Rotate independent reviewed source forms, blend compatible centerlines, and vary smooth letter geometry and arrow shafts/heads separately.
• Reject exact and near copies of ordinary forms; report a shortage of usable shapes instead of silently repeating them. Small punctuation and straight bars have a separate, explicit limited-range report.
• Add profile diversity, profile mine, and train --mine to find more examples in preserved pages and produce private review queues. Labels and fitted geometry remain subject to review.
• Repeat forms in calibration and guide the skill through collection, review, rendering, actual-PDF checks, and visual refinement.
• Preserve vector print output, semantic content, reviewed correction habits, and private source evidence.

Validation: 311 Python tests and 8 npm launcher tests on Linux/Python 3.11 and macOS/Python 3.13, real packed-package installation and standalone skill rendering, an independent training-to-rendering workflow, and three-page repetition proofs. Release archives contain only code, documentation, and authored synthetic examples.

This remains local exemplar learning. The repetition audit does not certify resemblance or a physical print, and sparse source variety can prevent an export.

Upgrade:

npm install -g @nazifbd/inkform@0.3.0
inkform setup
inkform skill install --force

The attached skill ZIP includes its portable runtime. Contributions and synthetic regression examples are welcome.

Inkform v0.2.0 — natural variation and corrections

Repeated letters and limit arrows could look copied in earlier versions. Inkform 0.2.0 varies each occurrence using reviewed alternative forms and bounded changes to width, height, slant, baseline, shape, and spacing. The PDF stays vector, and a saved seed reproduces the same output.

• Learn width and height spread from distinct reviewed examples, and word-gap spread from reviewed adjacent words. Conservative estimates remain identified when samples are sparse.
• Learn correction frequency from complete, representative sample passages. Reuse the writer's reviewed strike or scribble marks for occasional crossed-out word restarts. Profiles without this evidence receive no automatic cross-outs.
• Preserve explicit deleted text and its replacement with correction runs. Semantic accounting checks every extra mark and deleted word; intended text and mathematical signs remain intact.
• Updated agent training guidance, a runnable examples/corrections.json, and refreshed public website specimens. All public handwriting remains authored synthetic demonstration material.

Upgrade:

npm install -g @nazifbd/inkform@latest
inkform setup
inkform skill install --force

Reload your agent after upgrading the skill. Existing writer profiles remain compatible; resume training with reviewed repeated forms and correction evidence to learn those personal habits. Appearance and actual-size print calibration are still required.

Validation: 241 Python tests, 8 Node tests, a real packed npm installation producing vector PDFs, independent installed-skill execution, inspected PDF/site previews, and archive checks excluding private samples and profiles.

MIT licensed and open for contributions. See the introduction and contributing guide.

Inkform v0.1.2 — scoped npm release

The npm distribution is now named @nazifbd/inkform. The project remains Inkform and its command is still inkform. npm rejected the unscoped name as too similar to an existing package.

npm install -g @nazifbd/inkform
inkform setup
inkform skill install

This patch updates the website and installation instructions, verifies scoped-package installation, and supports upgrading a verified earlier standalone skill installation. Handwriting behavior is unchanged.

Release assets include the Python wheel/source package, scoped npm tarball, standalone skill ZIP, and SHA-256 checksums. The ZIP contains its own runtime: after extracting it, run node runtime/bin/inkform.mjs setup from the skill directory.

Validated with 165 Python tests, 8 Node tests, an actual packed-package setup/render, standalone skill reuse, and archive privacy checks. Public examples contain only synthetic forms.

Introduction · Contributing

Inkform v0.1.1 — first public release

npm installation update: use npm install -g @nazifbd/inkform. The matching scoped distribution is available in v0.1.2. npm rejected the original unscoped name; the archives below remain usable.

Inkform learns a reusable private handwriting profile from reviewed samples, creates a calibration proof, and uses that profile to produce printable vector PDFs. This is the first public release, under the MIT license.

• Local Python engine and agent skill, with reviewed source provenance and immutable profile versions.
• Sample-first training, smooth outline refinement, semantic text/math layout, and actual PDF checks.
• npm launcher with explicit environment setup and a standalone skill installer.
• Introduction and examples and contribution guidance.

Release assets include a Python wheel/source package, the npm tarball, a self-contained skill ZIP, and SHA-256 checksums. The skill ZIP includes its runtime; after extracting it, run node runtime/bin/inkform.mjs setup from the skill directory. The npm tarball can be installed with npm install -g ./inkform-0.1.1.tgz, followed by inkform setup and inkform skill install.

Validated with 165 Python tests, 7 Node tests, a clean packed-package installation, vector PDF rendering, and standalone skill reuse. GitHub CI passed on Linux and macOS. All release archives passed the public-data check.

Training uses reviewed exemplar geometry. Segmentation and geometry review still need an agent or person; connected cursive, learned word spacing, and physical resemblance need further work and print calibration. Public examples are synthetic and contain no personal handwriting.