focr
Made by JeffUse when working with focr/franken_ocr: OCR CLI, OcrEngine, robot NDJSON, .focrq artifacts, Model Zoo (GOT, SmolVLM2, OneChart, TrOMR), kernels, quantization, or ~/projects/franken_ocr.
From Jeffrey's Skills.md: professionally crafted skills for any AI agent, installed and kept in sync by the jsm CLI. Subscription only, $20/month.
Works With
Quality and Utility come from canonical skill metadata. Community ratings appear separately as stars when subscriber reviews exist.
Subscribe to download
Public previews are available, but downloads, CLI sync, and premium content require an active subscription.
Sign In to Subscribe- Full skill content with deterministic versioning
- CLI sync + local search index
- Access to Jeffrey’s premium library
Save to your account
Saved skills live in your cloud list. Run jsm sync on each machine to pull them locally.
What's New with this Skill
5Updates, enhancements, and release notes for this skill.
Enhanced: focr
The focr skill's install guidance, release facts, and PDF handling now match focr v0.7.2. Fixes four issues from a subscriber's field report.
- Install section drops the false baked-in-fallback claim and documents the --version pin recovery when release lookup fails
- Release facts refreshed to v0.7.2: 4.16 GB default artifact fetched in 3 parts, versioned unlimited-ocr.v0.7.0.int8.focrq filename
- Born-digital PDF constraint moved into the PDF workflow, with the exact decode error and a pdftoppm -png -r 200 rasterize recipe
- New binary-only Fast Probe (focr --version, models --json, robot triage, robot selftest, doctor), each command verified against a real installed binary
- Warns that GitHub's Latest slot can be a weights release; the fixed upstream installers now select the newest semver tag instead
A subscriber installed focr on a fresh machine by following this skill and hit two failures that traced back to stale skill text. All four reported issues are fixed. The installer section no longer claims a baked-in fallback version (there is none) and documents pinning --version vX.Y.Z as the recovery move when release lookup fails. The Cold-Start Boundary Card carries a dated August 19 update with the current v0.7.2 release facts and a warning that GitHub's Latest release slot can hold a non-code weights release. The PDF workflow now states up front that the native fast path handles scanned
focr v4: franken_ocr v0.6.0 ships certified faster than PyTorch, and the skill teaches the five-ledger discipline behind that claim
franken_ocr published v0.6.0, a real five-platform release named "certified faster than PyTorch, end to end", and the skill now documents both the win and exactly what proves it: an all-green release-readiness scorecard (ship:true, green:13, red:0), a certified gauntlet bundle, and head-to-head PERF_LEDGER rows. New fresh-agent guidance separates five truth surfaces so nobody mistakes a GitHub release for the binary on their PATH.
- franken_ocr v0.6.0 is live: "certified faster than PyTorch, end to end", published 2026-07-08 with binaries plus .sha256 for Apple Silicon, macOS x86-64, Linux x86-64, Linux ARM64, and Windows x86-64.
- The ship gate flipped all-green: release readiness went from ship:false (green:11, red:2) to ship:true (green:13, red:0), with gauntlet convergence at rounds 11/10 and a clean tail.
- The performance claim is backed by evidence: SmolVLM2 untied lm_head int8 top-K-refine certified default-on, plus head-to-head PERF_LEDGER rows against PyTorch, not a one-off timing.
- Five-ledger discipline for fresh agents: release object, source head, installed binary, post-certification corpus/live WIP, and readiness artifact are separate truth surfaces; never let one speak for another.
- Receipt archaeology built in: the bundle certificate reads v0.5.2-8 because it predates the tag; the skill explains the mismatch so agents cite the certification chain correctly.
- Version-string triage: v0.6.0-N-gHASH means main is ahead of the release; -dirty means the tag proves even less; installers still fall back to v0.4.0 when release lookups fail.
About a day and a half after its v0.4.0 release, franken_ocr crossed its ship gate. The v0.6.0 release (published 2026-07-08, five platform binary families plus .sha256 assets) is backed by a receipt chain the skill now walks you through: SmolVLM2's untied lm_head int8 top-K-refine certification became default-on, head-to-head PERF_LEDGER rows landed against PyTorch, the gauntlet converged at rounds 11/10 with a clean tail, and the release-readiness scorecard flipped from ship:false with two red cells to ship:true, green:13, red:0. The skill records the old red state as historical and explains
focr v3: franken_ocr goes public, with v0.5 binary releases, a fully pullable model zoo, and a live doctor
franken_ocr crossed from source project to shipped product, and the skill tracks the new reality. There are now public GitHub releases with five-platform binaries (v0.5.0 brought cross-page parsing, a hardened music lane, and a faster vision tower; v0.5.1 is the vision-pipeline efficiency patch). The whole model zoo is one focr pull away, focr doctor is live with --fix, --dry-run, and undo, and the operator playbook grew to 41 with a Release Ship Gate and a source-vs-binary boundary discipline.
- franken_ocr has public GitHub releases now: v0.5.0 (cross-page parsing, hardened music lane, faster vision tower) and v0.5.1 (vision-pipeline efficiency patch), each with binaries plus .sha256 for five platforms.
- The whole zoo is pullable: focr pull smolvlm2, focr pull onechart, and focr pull tromr install artifacts with their tokenizer sidecars; tromr defaults to int8 with --quant f32 for the bit-exact variant.
- focr doctor is live, not scaffolding: detect-only by default, with --fix, --dry-run, and undo, governed by its own Doctor Repair Contract operator.
- Source-vs-binary discipline: the v0.5.2 tag exists without release assets, installers fall back to v0.4.0 on failed lookups, and the skill teaches you to never let an installed binary speak for current source.
- Seven new operators (41 total): arch-specific prepack, doctor repair, multi-page cross-page parsing, PDF page selection and spread splitting, robot triage ergonomics, release ship gate, and zoo-manifest pullability.
- TrOMR music lane hardened with evidence: per-staff skip resilience, robot staff events, a Spohr real-scan corpus, fit-first geometry, and annotate-only musical-sanity warnings.
- Vision tower speedups with evidence: batched SigLIP frames by default, FOCR_SIGLIP_SEQ=1 as the kill switch, and byte-identical fixture evidence separated from formal perf-ledger claims.
- The skill grew from 500KB to 844KB: SKILL.md more than doubled, TROUBLESHOOTING and VERIFICATION roughly doubled, and the self-test suite doubled.
The biggest change is distribution. In the previous version of this skill, only Unlimited-OCR and GOT-OCR2 were pullable and the skill's job was to stop you from inventing focr pull smolvlm2. Now the manifest lists the whole zoo: focr pull smolvlm2 installs the int8 artifact plus tokenizer.json, focr pull onechart installs the OPT tokenizer sidecars, and focr pull tromr installs tromr.int8.focrq with its four tokenizer tables (add --quant f32 for the bit-exact variant). On top of that sit real GitHub releases: v0.4.0 set the public boundary, v0.5.0 shipped cross-page parsing, the hardened musi
focr v2: a five-model zoo (GOT-OCR2, SmolVLM2, OneChart, TrOMR), a 34-operator playbook, and a full verification-suite tour
The focr skill grew from a single-pipeline OCR guide into the manual for a five-model zoo: GOT-OCR2 task modes, SmolVLM2 image description and VQA, OneChart chart-data extraction, and TrOMR music OCR to MusicXML, next to the original Unlimited-OCR lane. A Model Surface Map states what each model can do and what not to claim, 34 named operators cover the daily work, and a verification tour explains the conformance matrix, differential and metamorphic tests, and the L0-L5 ladder scorecard.
- Model Surface Map: one row per model with the current route, required setup, and an explicit do-not-claim column, so you never invent a focr pull onechart or a focr music subcommand that does not exist.
- GOT-OCR2 is a pullable named model: focr pull got-ocr2, then --task formula|tables|chart|molecular|geometry|music or --format mode (needs qwen.tiktoken beside the artifact).
- SmolVLM2 image description and VQA: focr ocr --task describe --model smolvlm2.int8.focrq [--question ...], with conversion, tokenizer, vision/connector, and quality-guard evidence each mapped to its own operator.
- TrOMR music OCR: --task music --model tromr.focrq turns single-staff and v1 printed/scanned full-page scores into partwise MusicXML, with tokenizer tables supplied beside the artifact.
- OneChart chart-data extraction: --task chart-data with a local or converted onechart.int8.focrq plus the OPT tokenizer files.
- Run state is now a current surface: focr runs plus focr sync export-jsonl / import-jsonl, backed by RunStore and FOCR_RUN_STORE.
- Cooperative cancellation and thread budgets: request_shutdown, cancel_checkpoint, FOCR_THREADS, and page streaming, with guidance for spotting stale binaries that predate them.
- OPERATORS.md grew from 5KB to 110KB: 34 named operators, from OP-LC (Live Contract Probe) to OP-GB (Gauntlet Baseline), including OP-LS (Ladder Scorecard Runner).
- A verification-suite tour: conformance matrix with MUST coverage >= 0.95 and XFAIL discipline, determinism and fixture-provenance gates, differential comparison against a bf16 oracle, metamorphic relations, and golden artifacts.
- Five times the material overall (93KB to 500KB), organized behind a Truth Stack, a First 90 Seconds checklist, and a Claim Taxonomy so fresh agents answer with evidence.
The focr skill's first release documented one pipeline: Baidu's Unlimited-OCR, ported to pure Rust. Since then franken_ocr has grown a real model zoo, and the skill has been rebuilt around it. GOT-OCR2 is now a named, pullable model with task modes for formulas, tables, charts, molecular diagrams, geometry, and sheet music. SmolVLM2 answers questions about images through --task describe. OneChart extracts structured chart data, and TrOMR reads printed or scanned scores into partwise MusicXML. Each lane comes with the exact commands, the artifact and tokenizer setup it needs, and the claims you
focr turns scanned documents into structured markdown
focr drives franken_ocr, a pure-Rust OCR engine that reads images and scanned PDFs into layout-aware markdown plus JSON layout boxes, with a robot NDJSON mode for pipelines. It also keeps you honest about which commands your installed binary actually has, versus what is still phase-gated.
- Load the int8 .focrq model once with focr pull, then OCR fully offline by reusing one long-lived OcrEngine.
- focr ocr invoice.png --json returns schema_version, markdown, and per-region layout boxes; scanned PDFs nest layout per page.
- focr robot run streams versioned NDJSON, one event per line, so pipelines can parse it incrementally.
- Surface discipline is built in: current, scaffolded, and phase-gated (int4) commands stay clearly separated, so you never reach for one your binary cannot run.
franken_ocr is a pure-Rust, CPU-only engine that runs one model, Baidu's Unlimited-OCR. The focr skill teaches both the CLI and the OcrEngine library API: pull the int8 .focrq weights once, then turn invoices, receipts, and scanned PDFs into layout-aware markdown or structured JSON without leaving Rust. For automation, focr robot run emits line-oriented NDJSON you consume one event at a time. The skill also makes you classify each surface as current, scaffolded, planned, or stale before you trust it, so the commands you run match the binary you actually have.