focr

name	focr
description	Use 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.
dependencies	["franken_ocr repo or installed focr binary","Rust nightly when building from source","jq for robot/JSON inspection","Baidu/GOT/SmolVLM2/OneChart/TrOMR source weights or a .focrq model for real inference; TrOMR also needs tokenizer tables beside the artifact"]

focr / franken_ocr

One Rule

Do not invent capability. First classify a surface as current, scaffolded, planned, or stale, then give the user commands and code that match that class.

Truth Stack

Use this order when facts conflict:

1. Current committed franken_ocr source and tests. A dirty live worktree can identify candidate names and code paths, but it is live-WIP until the diff is understood and the matching tests/Beads agree.
2. Live help from the exact binary that will run, after confirming it is not stale.
3. br/bv --robot-* issue evidence for current closure/open status.
4. AGENTS.md, README.md, COMPREHENSIVE_PLAN_FOR_FRANKEN_OCR.md, and docs/ as narrative context that can lag source and Beads.
5. cass session history, as historical context only.

If a command exists in source but not in focr --help, treat the binary as stale and run from source or rebuild before drawing conclusions.

This repository-local skill is the maintained copy for this checkout. If a global installed $focr skill disagrees with .claude/skills/focr, use this project-local skill plus live ~/projects/franken_ocr source/Beads evidence.

If ~/projects/franken_ocr is dirty, run git status --short --branch and git diff --stat before trusting changed prose or comments. Uncommitted docs that advertise a feature, uncommitted source that has not passed the relevant gate, or extreme diffs that look mechanically corrupted are not current capability. Label them live-WIP, then check br show <id> --json, tests, and committed origin/main before updating user-facing claims.

Cold-Start Boundary Card (July 8, 2026)

Dated update (2026-08-18). The card below is preserved verbatim as the July 8 observation; these deltas supersede its point-in-time facts:

• Latest binary release is v0.7.2 (published 2026-07-13), not v0.6.0. Six release binaries now ship, adding Windows ARM64.
• The "Latest" GitHub release slot can be a weights release. Weights ship as their own non-prerelease releases (models-*), so /releases/latest returned models-unlimited-wasm-v1 in August 2026 and broke bare installer runs (GH #12; installers on main now filter for semver v* tags). Never equate "latest release" with "latest binary release" here.
• The default artifact is versioned: focr pull installs unlimited-ocr.v0.7.0.int8.focrq (4.16 GB, 3 verified parts) under the cache root, using the conservative recipe unlimited-ocr-ffn-int8-attn-bf16-lmhead-bf16-v1. The old unlimited-ocr.int8.focrq name and the ~3.9 GB figure are stale.
• All five runtime models (unlimited-ocr, got-ocr2, smolvlm2, onechart, tromr) are pullable; TrOMR publishes int8 (default) and f32 artifacts.

If you are fresh to focr, keep these five ledgers separate before answering:

• GitHub release: latest observed GitHub release is v0.6.0, a real, non-draft, non-prerelease release named "v0.6.0 — certified faster than PyTorch, end to end", published 2026-07-08T14:47:48Z with Apple Silicon, macOS x86-64, Linux x86-64, Linux ARM64, and Windows x86-64 binaries plus .sha256 assets. The release tag is 29516b9 / v0.6.0. This is proof of the release-asset boundary, not proof of the installed binary on the user's PATH, and it does not include source commits after the tag. Historical release name to preserve for archaeology: v0.5.2 — the cold-start release.
• Public source head: origin/main is 592426c, with clean committed source describing as v0.6.0-4-g592426c; the inspected checkout had no tracked source diff and only untracked .claude/worktrees/. The post-v0.5.2 chain is now: 0924479 README source/binary clarification, efd83e8 formal bd-av64.10 G2 re-measurement and closeout, 3f2f97e bd-2mo.26 gauntlet harness fixes, 4291807 SmolVLM2 untied lm_head int8+top-K-refine default-on certification, 91c12fd README selftest evidence refresh, ae7b8f2 bead-scoped perf evidence bundle, c248e6d bd-2mo.26 head-to-head PERF_LEDGER rows, c4c1684 bd-2mo.26 plus rounds 9-10 closeout, c29a78b release certification achievement, 7c7bd00 Beads closeout for bd-wp8.8, bd-wp8.9, and bd-wp8.10, 29516b9 v0.6.0 release version/tag, beaed7c CI/dist supplement notes, db02421 README release-readiness evidence refresh, 5df6395 committed post-certification fuzz corpus growth, and 592426c README release-identity refresh for the now-unified v0.6.0 source tag plus binary release. This is current committed source/tracker evidence, not proof that a user's installed binary came from the same source without an installed-binary check.
• Committed source features: 507cebe mmap-loader half, GOT , / OneChart , SmolVLM2 , and SmolVLM2 untied- int8+top-K-refine default-on certification are current source evidence with scoped proofs. The boundary proves paired describe output byte-identical f32-head vs int8+refine, armed L5 VQA green under the lever, full describe/VQA e2e suite pass, about , and decode ; is the f32-head kill switch. This is not a public VQA benchmark, broad quality proof, or guarantee that the release asset includes the post-tag source. Formal G2/PERF_LEDGER rows now exist for both and : the final state records GOT e2e , OneChart , SmolVLM2 , and decode-per-token ; the later head-to-head rows record focr-int8 vs pinned HF bf16 page_0009 e2e and page_0014 e2e with CER / , best-of-5 warm, N=8 both sides, and cv% under the 5% bar.

Mnemonic: release object, source head/features, installed binary, post-certification corpus/live-WIP boundary, and readiness artifact are five different truth surfaces.

First 90 Seconds

Use this checklist before reading deep references:

1. Name the surface: CLI command, Rust library API, model/artifact, verification gate, performance claim, or source edit.
2. Classify the evidence boundary: installed binary, current source, committed main, live worktree, Beads status, or historical docs. Dirty-worktree claims need an explicit live-WIP/committed-main distinction before they become advice.
3. For install/release/version questions, keep four facts separate: source/Cargo version, pushed tag, GitHub release object/assets, and the actual installed binary help/version. Do not infer one from another. If git describe says v0.5.0-N-gHASH, v0.5.1-N-gHASH, v0.5.2-N-gHASH, or v0.6.0-N-gHASH, current main is ahead of that tagged release. If it says v0.5.2-dirty or v0.6.0-dirty, the tag/release boundary can still be real while the checkout contains uncommitted work. Treat source tag, release object, release assets, and installed binary as four separate facts.
4. Pick one operator/reference from the router below; do not read every file.
5. If the claim is about "done", "green", "conformant", "fast", or "quality", name the exact gate and what it does not prove.
6. End with concrete commands or code and an explicit proof gap when a model, fixture, binary, or tracker close is missing.

Default Answer Shape

When answering from this skill, use this order unless the user asked for a different format:

1. Boundary: "current source", "release binary", "dirty live-WIP", "Beads-open", or "historical note".
2. Action: the exact focr command, Rust API, file path, or verification gate.
3. Exclusions: the adjacent thing this does not prove, for example "not a GitHub release asset", "not full bd-2mo.22 closure", "not SmolVLM2 statics caching", or "not a formal PERF_LEDGER row".
4. Proof gap: one command or artifact the next agent should check if the user needs stronger certainty.

This keeps answers ergonomic: lead with what works, then attach the smallest necessary caveat. Do not bury the usable command under the entire history.

Fresh-Agent Reading Plan

Do not read this skill linearly. Use it like an operator console:

1. Read the Cold-Start Boundary Card and First 90 Seconds checklist.
2. Choose one surface from the Fresh-Agent Router: CLI, library, model/artifact, verification, performance, release, or troubleshooting.
3. Open exactly one matching reference first. Use OPERATORS.md when the question is about proof boundaries, state classification, or avoiding overclaims; use CLI.md, LIBRARY.md, or ARTIFACTS-AND-ENV.md for concrete commands and APIs.
4. Search within the skill for the exact bead, commit, env var, command, model id, or operator tag. If a global/installed copy of the focr skill disagrees with this project-local skill, use this project-local skill plus live ~/projects/franken_ocr source/Beads as the fresher evidence.
5. Stop once you can state the boundary, action, exclusion, and proof gap. More history is not better unless the user asks for archaeology.

Fresh-Agent Router

• Cold start if you know nothing: read this router, classify the exact surface with OP-LC (Live Contract Probe), check the model row in the Model Surface Map, then open only the matching reference. Do not read RESEARCH.md first; it is historical context after current source/help/Beads are settled.
• Asked "can focr do X?" Start with CLI.md, then confirm source/help and Beads before answering yes/no.
• Editing or reviewing franken_ocr source? Use DEVELOPMENT.md and VERIFICATION.md; read repo AGENTS.md and README.md first.
• Handling models, artifacts, env vars, or deployment? Use ARTIFACTS-AND-ENV.md, then FOCRQ.md.
• SmolVLM2 work splits by evidence layer: OP-SC conversion/decoder census, OP-ST tokenizer, OP-SV vision/connector, OP-SP preprocess/prompt/route, OP-VQ Rust VQA guard and C10 CLI e2e script. Sub-epic C is now closed, but each claim still needs the right evidence layer. Source at/after 9b2a03b has smolvlm2::SmolStatics, OcrModel::smol_statics, and smolvlm2.hydrate(cached) as committed pass-8 evidence: SigLIP tower, modality projection, and widened untied text embed hydrate once per model and are shared by sequential describe/VQA and batch paths. Keep it scoped to byte-identical describe source evidence plus lib-green proof. The parent bd-av64.10 now has a formal closeout at efd83e8, but pass 8 by itself is not release readiness, public VQA quality, or a standalone PERF_LEDGER row.
• Model-zoo distribution work is OP-ZM: ece14f9 / closed bd-av64.7 publishes the embedded pull manifest for all five ready models, and bd-av64.8 / bd-av64.9 add the public GitHub release and clean-cache pull verification for models-smolvlm2-v1, models-onechart-v1, and models-tromr-v1. focr pull smolvlm2, focr pull onechart, and focr pull tromr are now current when the running binary embeds or resolves that manifest; clean-cache evidence says the GitHub release assets, hashes, sidecars, idempotent repull, and real inference paths work. Use and , , , , and separate checks for , , , and before making pullability/quant claims. Keep the TrOMR boundary precise after : now publishes both and ; the default resolves the storage artifact, while gets the bit-exact reference artifact. Do not turn that into an int8-compute, speed, or quality-improvement claim. Also keep the mirror boundary precise: production pulls are GitHub-first and verified; HF mirror spot-checks returned 401 and remain a resilience/auth follow-up, and the dedicated one-command pull-e2e script is still deferred.

Claim Taxonomy

User claim	Use	Minimum honest answer
"Can focr do X?"	OP-LC + CLI.md	source/help/Beads classification plus exact command
"Can this model run?"	OP-MD/OP-MA + model operator	registry status, manifest packaging, artifact/tokenizer requirements
"Is output correct?"	OP-DF/OP-MR/OP-VG/model-specific proof	exact oracle, relation, fixture, or corpus boundary
"Is it conformant/release-ready?"	OP-CM plus OP-LS/OP-GA/OP-DG as needed	accounting result and every separate proof family still needed
"Is it deterministic?"	OP-DG, then OP-MR for transform consistency	same-input byte identity vs oracle quality kept separate
"Did the whole ladder pass?"	OP-LS	scorecard schema, all_green, skipped_no_model, and first-failure boundary
"Is it fast?"	OP-BG/OP-GB	architecture/selftest vs PERF_LEDGER/gauntlet rows, benchmark guardrails, and negative evidence kept separate
"Why does my binary differ?"	OP-SQ	exact binary path/version and release-lag/stale-local classification

Model Surface Map

Use this as the first-pass model router. For non-trivial claims, load the matching operator/reference before answering.

• Unlimited-OCR: default focr ocr lane. Setup is focr pull. Do not claim generic arbitrary-checkpoint runtime.
• GOT-OCR2: focr ocr --task formula|tables|chart|molecular|geometry|music or --format. Setup is focr pull got-ocr2; keep qwen.tiktoken beside the artifact. Do not call it a default replacement for plain OCR.
• SmolVLM2: focr ocr --task describe [--question ...]. Setup is focr pull smolvlm2; it installs smolvlm2.int8.focrq plus tokenizer.json. Do not claim public VQA benchmark or human-label quality proof. Source at/after f1ac972 routes vision_rows through vision_siglip::forward_frames_batched by default, with FOCR_SIGLIP_SEQ=1 as the kill switch back to the per-frame loop; the current proof is byte-identical source/fixture evidence plus modest loaded-host timing wins, not a formal quiet-host PERF_LEDGER row.
• OneChart: focr ocr --task chart-data. Setup is focr pull onechart; it installs onechart.int8.focrq plus OPT tokenizer sidecars. Do not invent focr chart or broad chart quality.
• TrOMR: focr ocr --task music --model tromr.int8.focrq after the default pull, or --model tromr.focrq after focr pull tromr --quant f32, to partwise MusicXML. Setup is focr pull tromr; default pull installs tromr.int8.focrq plus four tokenizer tables, while focr pull tromr --quant f32 installs the bit-exact tromr.focrq. Current evidence includes bd-av64.2 per-staff skip resilience, robot staff events, music-run JSON staves, closed bd-av64.6 Spohr real-scan corpus v1 measuring-device evidence, closed bd-av64.14 fit-first geometry/p169 evidence, closed bd-av64.5 annotate-only musical-sanity warnings, closed bd-av64.13 residual-skew plus measured-negative/reverted TTA and one-crop levers, closed bd-av64.12 TrOMR int8 storage publication, and the bd-2sez losing f32 PERF_LEDGER baseline. Do not claim , int8 compute, a perf win, camera dewarp, default/lossless barline quality, , / event names, a schema-version bump for or , every-page PDF music , automatic correction from sanity warnings, or broad real-scan SER completion. is an experimental over-budget-staff recognition-count rescue path only; it is not default behavior or a broad barline-quality proof. The int8 artifact is a published quantized-storage artifact, not a current int8-kernel runtime.

Current Reality Snapshot

Public release/install facts and source facts can still differ, and the July 8, 2026 check now has a published v0.6.0 release plus newer post-release source commits. f0a538b created the v0.4.0 public release boundary. bf28fd7 bumps Cargo.toml / Cargo.lock and the pushed annotated tag to v0.5.0; the v0.5.0 tag points at bf28fd7, and the GitHub release named "cross-page parsing, hardened music lane, faster vision tower" was published 2026-07-07T23:49:41Z with binary plus .sha256 assets for Apple Silicon, macOS x86-64, Linux x86-64, Linux ARM64, and Windows x86-64. 8de3674 tags the source/package as v0.5.1; the GitHub release named "v0.5.1 — the vision-pipeline efficiency patch" was published 2026-07-08T05:02:37Z with the same five platform binary families plus .sha256 assets. 4cedacd tags the source/package as v0.5.2; the GitHub release named "v0.5.2 — the cold-start release" was published 2026-07-08T06:05:23Z with the same five platform binary families plus .sha256 assets. 29516b9 tags the source/package as v0.6.0; the GitHub release named "v0.6.0 — certified faster than PyTorch, end to end" was published 2026-07-08T14:47:48Z with the same five platform binary families plus .sha256 assets. install.sh / install.ps1 still keep v0.4.0 as the fallback constant for failed release lookup, and installed binaries can still lag; normal online installers resolve the latest release object first. The 592426c README has been refreshed for v0.6.0, but the standing support rule remains: when README/manual prose and the release API disagree, verify the actual release object/assets before making support claims. The July 8, 2026 source probe saw public origin/main at 592426c (v0.6.0-4-g592426c for clean source; the inspected checkout had no tracked source diff). The post-v0.5.2 commits are 0924479, a README-only clarification, efd83e8, the formal bd-av64.10 G2 re-measurement / Beads closeout, 3f2f97e, the bd-2mo.26 gauntlet harness bring-up fix, 4291807 SmolVLM2 untied int8+refine certification, CPU selftest README refresh, perf evidence bundle, head-to-head PERF_LEDGER rows, plus rounds 9-10 closeout, release certification achieved, Beads closure for the release certification trio, tag/version bump, CI/dist supplement notes, README release-readiness evidence refresh, committed post-certification fuzz corpus growth, and README release-identity refresh. Treat , , , , , , , , , , , , , , , , , , pass-6 , source version bump , mmap-loader half , Beads note , README/release-readiness commits /, pass-7 OneChart statics /, pass-8 SmolVLM2 statics , README docs , and source/package tag , README clarification , G2 closeout , and harness fixes , SmolVLM2 head certification , selftest README refresh , perf evidence bundle , head-to-head rows , closeout , release certification , Beads closeout , tag , CI/dist supplement , README release-readiness refresh , fuzz corpus commit , and README release-identity refresh as current committed source evidence unless a newer release asset, installed binary, or live checkout proves otherwise. If the checkout is dirty after , classify only those uncommitted diffs as live-WIP; uncommitted docs are not stronger evidence than committed source, tests, and Beads. Always verify the exact installed , help, tag, GitHub release object/assets, and installer path before support claims; stale README/manual prose is not stronger evidence than the release API or the installed binary. The public source/release boundary includes , , JSON layout boxes, , , , int8 , robot commands, scanned PDFs via , and layout APIs. Post- additions covered by this skill include multi-page cross-page parsing ( and PDF ), multi-page streaming events, the 2-page and 10-page L5 multi-oracle rungs plus the frozen 20-page reference-collapse fixture, additive TrOMR robot/JSON observability, fit-first TrOMR staff crop geometry, experimental barline rescue, annotate-only TrOMR / JSON sanity telemetry, per-band TrOMR residual-skew refinement, model-aware rollups, and offline prepacked layouts, default TrOMR storage publication with as the f32 reference artifact, and the closed negative-evidence ledger for TTA/one-crop routing. Treat those as release-or-current-source claims only after the exact binary/help, tag, release assets, and Beads closure agree. Closed-current verification additions after the boundary also include property/fuzz plumbing, the decode decompression-bomb guard, and pass-4 shared SAM/SigLIP/TrOMR/projector pre-transposed-linear evidence ( plus Beads evidence in ), pass-5 SmolVLM2 SigLIP frame batching (), pass-6 GOT statics caching (), pass-7 OneChart statics caching ( / ), pass-8 SmolVLM2 statics caching (), README docs , source/package tag , README split-clarification , formal G2 closeout , gauntlet harness hardening , and SmolVLM2 untied-head certification , head-to-head rows , release certification , Beads closeout, tag , CI/dist supplement , README release-readiness refresh , fuzz corpus commit , README release-identity refresh , and source-current CI/gate closure. / are in the ancestry; post-tag follow-ons now extend through public in this checkout. The important new boundaries are:

• ab6e083 / 50d5dad: SIMD/polynomial-exp softmax in SAM attention measured dead and was reverted. Unit-level exp/softmax accuracy did not preserve token output; use the negative-evidence ledger rather than retrying it as an easy speed lever.
• ab6fa6c / 3f3d9d0: round 8 is CLEAN after 4 fuzz targets x 5 minutes (4.7M total runs, zero crashes), PROPTEST_CASES=2048 8/8 green, and a 6/6 advisory matrix, but this is bounded deep evidence, not exhaustive fuzzing or release readiness.
• 5df6395: the post-certification fuzz corpus is committed, not dirty local WIP. It adds 3,271 seed files across focrq_parse, image_decode, pretok_split, and safetensors_parse. The later convergence sweep records four fuzz targets x four minutes, about 3.65M total zero-crash runs, plus PROPTEST_CASES=2048 8/8 green. Keep the old round-8 4.7M sweep as bounded historical evidence and the round-11 3.65M sweep plus committed corpus as the current convergence/corpus evidence.
• 592426c: README-only current-source refresh for the public v0.6.0 identity. It replaces stale badge/install/manual-download references, documents that source tag and binary assets are both v0.6.0, updates the binary-size claim from the old 5 MB-era language to about 13-17 MB, and clarifies CPU backend wording: Apple Silicon prefers SDOT, non-Apple ARM64 can dispatch SMMLA/i8mm, and Intel/AMD select AVX-512-VNNI, AVX-VNNI, AVX2, or scalar by runtime feature detection. This is README/public-support evidence, not a new runtime feature.
• 8cacf52 / ff22366 / 2665750 / 9bc715e: these are now historical bundle bring-up commits. They made scripts/gauntlet_cert.py --bundle real and converted certification_bundle into a live release_certificate.json cell rather than a hard-coded red/self-referential predicate. The old certified-false state is superseded by the later c29a78b/7c7bd00 certification closure; keep the bring-up details only when debugging how the bundle gate works.
• 48a9896: README and fuzz metadata now name v0.5.0; installer fallback constants and installed binaries are still separate checks.
• f1ac972: SmolVLM2 SigLIP frame batching is current source. It proves byte-identical forward_frames_batched behavior with as a sequential kill switch and loaded-host self-relative wins. The later G2 rows are the formal closeout state; this individual pass remains pass evidence, not a standalone release-readiness or quality claim.

Dirty-worktree-only surfaces to quarantine, not promote, now mean new uncommitted deltas after public 592426c, not the already-landed Linear::from_row_major, --bundle, forward_frames_batched, row-tile-negative-evidence/baseline-restore, GotStatics, or mmap-loader facts or OneChart/SmolVLM2 statics facts, the v0.5.2 source tag, or the README source/binary split, G2 closeout, gauntlet harness fixes, SmolVLM2 head certification, head-to-head perf rows, or release certification described above. In the inspected checkout, committed fuzz corpus growth belongs to 5df6395 and the current README release-identity refresh belongs to 592426c; only new tracked diffs or untracked files after that boundary should be treated as live-WIP. focr pull [MODEL] --manifest <path-or-url> resolves manifest source as explicit --manifest, then FOCR_MANIFEST_URL, then the built-in repo manifest (BUILTIN_MANIFEST_JSON = include_str!("../models/manifest.json")). Closed bd-av64.7 / commit ece14f9 makes that manifest the distribution source for all five ready models: the top-level default unlimited-ocr artifact, and named models.got-ocr2, models.smolvlm2, models.onechart, and models.tromr entries. Non-primary models install under ~/.cache/franken_ocr/models/<model-id>/, with ModelEntry.sidecars carrying tokenizers and auxiliary files. focr models --json exposes each row's pull object with in_manifest and quants; use that instead of guessing from the registry row. Current main also has focr models, .focrq model_id, implemented GOT-OCR2 (focr pull got-ocr2, got-ocr2.int8.focrq), implemented SmolVLM2 (focr pull smolvlm2, smolvlm2.int8.focrq), implemented OneChart (focr pull onechart, onechart.int8.focrq), implemented TrOMR (focr pull tromr, default tromr.int8.focrq plus focr pull tromr --quant f32 for tromr.focrq and tokenizer sidecars), GOT / , and routing for . GOT tasks imply GOT format mode and need ; routes through SmolVLM2 and needs , plus optional ; routes through OneChart and needs plus the OPT tokenizer files beside it; is dual-lane: with it is GOT sheet-music format mode, and with it is native Polyphonic-TrOMR OMR to partwise MusicXML. with no explicit quant now selects the published storage artifact; use or a preverified local when byte-exact reference weights matter, and do not turn either artifact into an int8 compute claim.

Closed bd-av64.8 and bd-av64.9 move the zoo story beyond "manifest prepared": GitHub releases models-smolvlm2-v1, models-onechart-v1, and models-tromr-v1 were published and then clean-cache verified. Evidence covers exact release sizes/hashes, sidecar sets, per-model cache subdirectories, idempotent repull, then-current TrOMR f32 fallback behavior now superseded by the bd-av64.12 default storage-int8 artifact, and real inference for TrOMR MusicXML, SmolVLM2 description, and OneChart chart-data. The known distribution gap is HF mirror auth/resilience: spot-checks returned 401 across the weights mirrors, while GitHub-first production pulls work. Do not describe HF mirroring as verified, and do not claim the deferred one-command pull-e2e script exists until source/Beads show it.

Latest main also wires preprocessing controls into the engine: --base-size, --image-size, and --crop-mode now reach PreprocessOverrides; default crop mode is base. --crop-mode gundam is live reference dynamic tiling, with first e2e evidence in bd-1e9n (rc=0, 7 views, CER 0.0179 / WER 0.0138 on page_0107), but do not round that up to a full parity sweep. SmolVLM2 conversion is now implemented and model-gated-proven on real weights (bd-3jo6.3.2): focr convert --model-id smolvlm2 produced 489 tensors = 224 int8 decoder GEMMs + 265 F32 high-precision tensors with an untied high-precision lm_head; C5 then certified the text-only SmolVLM2 decoder seam (bd-3jo6.3.5: f32 cos 1.000000 + 24-token L4 exact, int8 cos 0.998301 argmax-exact, DISC-002 near-tie flip recorded). C6 tokenizer conformance is also closed (bd-3jo6.3.6): PretokScheme::SmolLm2 is selected from tokenizer JSON, the GPT-2 ByteLevel/Digits pretokenizer path is id-exact, and the real-tokenizer gate is 128/128 token-id and decode exact against pinned HF tokenizers. Current main has C3/C4/A8/A9 closed for SmolVLM2 vision and connector seams: SigLIP post-LN parity on 13 real frames, NaViT bucketized position ids, bit-exact pixel_shuffle, and connector projection within measured tolerance. bd-3jo6.3.7 and bd-3jo6.3.9 are now closed: preprocess_smolvlm2 uses Pillow-bit-exact LANCZOS (resample: 1) with longest-side 2048, row-major 512-frame tiling plus a final global 512 frame; src/native_engine/smolvlm2.rs assembles prompt ids, vision rows, <image> splices, and SmolLM2 KV-cache greedy decode; --question / FOCR_SMOLVLM2_QUESTION set the VQA question; model_arch reports SmolVLM2 implemented=true; and the int8 route was proven live with "What is in the sky..." -> "There is a sun in the sky."

SmolVLM2 sub-epic C is now closed in live Beads/source. bd-3jo6.3.8 closed the C8 parity/e2e quality/perf gate with: L0b preprocess exact, L0c prompt 876/876 id-exact, L2 SigLIP cos 1.0, L3 prefill drift <5e-5 and argmax-exact, L4 opt-in FOCR_SMOLVLM2_CERT_FULL=1 O(n^2) re-prefill greedy 64/64 id-exact, ledgered DISC-003 KV-cache near-tie behavior, and L5 informational VQA 7/7 on both f32 and int8 with int8 answers identical to f32. Measured release-int8 informational times on Apple M4 were describe e2e 51.8s and VQA 13.0s; use A11/PERF_LEDGER rows for fairness-controlled performance claims. bd-3jo6.3.10 closed C10 with per-module tests plus scripts/smolvlm2_describe_e2e.sh; bd-3jo6.3 closed the sub-epic. So support focr ocr --task describe --model smolvlm2.int8.focrq [--question "..."], and support focr pull smolvlm2 after closed bd-av64.7 when the running binary embeds or resolves the updated manifest. Still avoid claiming public benchmark quality or concurrency-safe per-request question handling unless source/evidence specifically says so. Latest C8 VQA work adds an informational L5 guard with exact evidence: scripts/gen_smolvlm2_vqa_fixtures.py writes tests/fixtures/smolvlm2/vqa_fixtures.json from torch-oracle greedy answers on the committed sample photo, and vqa_quality_matches_oracle_l5 scores focr's answers by normalized exact match or symmetric content-word containment >=0.5. The guard fails below 70% f32 or 50% int8 matches when FOCR_SMOLVLM2_DIR has model.safetensors and/or smolvlm2.int8.focrq; live close evidence reports 7/7 on both legs. Treat skips as missing artifacts, not green evidence. Current source also contains scripts/smolvlm2_describe_e2e.sh, a C10 CLI gate that emits smolvlm2_describe_e2e/v1 NDJSON, checks missing-model and wrong-family negative paths, then runs describe and VQA through the real int8 artifact. Treat it as model-gated e2e evidence and cite C10 closure only when live Beads/source agree. If src/cli.rs still has an older doc comment saying OcrTask::Describe is planned, treat it as stale inline prose; the usage guards, forward_smolvlm2 dispatch, Beads, README examples, and DISC-003 are the current route evidence. OneChart is now a ready runtime and pull route for chart-to-data extraction when the running binary embeds or resolves the bd-av64.7 manifest. bd-3jo6.4.2 closed the D2 arch-aware conversion half: focr convert --model-id onechart classifies OPT decoder GEMMs under , verifies the tied against , dedups the tied head, writes plus the Apache-2.0 notice, converts the real 500 MB checkpoint from 384 source records to 383 records, produces 72 int8 GEMMs (12 layers x q/k/v/out/fc1/fc2), keeps vision/projector/number-head/norms high precision, and records a 346 MB artifact. / later publishes that artifact through the committed pull manifest, so is current when the running binary embeds or resolves that manifest. closed D9 tokenizer conformance: (plain GPT-2 regex, no Digits stage), , + + , base control tokens as splittable specials, pinned ids 50265 / 50266 / 50267 / 50268, bos=eos 2, pad 1, and 29/29 token-id-exact