Changelog
Changelog
All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
Unreleased
[0.2.0] - 2026-07-22
Restore correctness release. Several format fixes remove failure modes that could make a printed document unrecoverable, and a measurement campaign settled the long-open question of whether trained OCR models are worth shipping (they are not).
Breaking (pre-1.0, no compatibility shim): the wire format changed. The
identifier stays v1 (magic B1, _VERSION = 1, codec names unchanged), but
documents produced by 0.1.0 do not decode with 0.2.0. Re-create any archive you
intend to keep.
Highlights
- No trained OCR model is needed for any codec.
base16gnever required one;base32gno longer does either, because this release's format fixes moved it from model-required to stock-viable. The publishedglyphive-ocrmodel-*packages were trained on the wrong data and are not recommended for anything. base32gis Courier-only on stock OCR (measured 4–10pt across 3 fonts): it fails on OCR-B and DejaVu Sans Mono at every size. Documented so nobody picks it blind.- A very short final data line no longer decides whether a document restores. This was a lottery every document ran at create time.
Added
glyphive train(experimental) — build an OCR model for a codec/font/size with the data-integrity gates that every previous attempt lacked: it verifies that each row image actually shows its paired text, aborts on any unencodable transcription or non-zero trainer skip ratio, derives a narrowed unicharset from the codec registry instead of inheriting the base model's ~112 characters, and builds its own starter artifacts. Reported CER is labelled a proxy everywhere it appears; a produced model recordsgate_verdict=UNGATEDbecause character error rate has repeatedly failed to predict restore. Not needed to use glyphive — it exists so future experiments are measurable.- Repo-resident E2E benchmark grid harness (
benchmarks/e2e_grid.py): create → rasterize → OCR → restore over a font-size × line-width × codec × line-parity grid. Per-cell status is one ofrestored/not-restored/not-built/error; a configurationcreaterefuses to build is excluded from every restore-rate denominator (reported asn/m testable (k not built), or a loudUNTESTED),--repeat Naggregates across repeat documents, the corpus is a pinned checked-in fixture, and results JSON carries full provenance (commit, engine version, corpus digest). Harness logic is unit-tested with a fake OCR provider, no engine required. - Physical-scan regression fixture + gate
(
tests/fixtures/physical_scan/, wired intotests/test_ocr_transcripts.py): a committed, deterministically degraded page render (Gaussian blur 0.6 + 0.7° rotation, generation script included) restored byte-for-byte throughextract --from-imageswithtesseract-glyphive; skips cleanly without the engine. Guards the real-scan path with zero real user content. - Per-glyph deletion stats in
tools/ocr_font_report.py(align_and_tally): aligned-diff tallies of per-glyph drop rate and insertion adjacency alongside the existing confusion stats — OCR dropping a thin glyph shrinks the line and desynchronizes the fixed-width frame parse, a strictly worse failure than a substitution, so candidate alphabets are now gated on both.
Fixed
- A very short ("runt") final data line no longer decides restore.
Whether a document restored at small font sizes depended on
encoded_length mod line_capacity: a few-character final data line is destroyed by OCR page segmentation (and can corrupt the line above it), so one payload byte more or less flipped restore between OK and FAIL (measured: 13-char final line fails, 153-char passes —benchmarks/results/fourpt-runt-line-20260721.json). The codec now zero-pads the protected stream so the final data line's printed payload is never below half the line width; the header's recorded length stays unpadded, so decode truncates the pad with no decoder change. - Group-packed decode no longer crashes on an out-of-range group. For
codecs where
radix**group_chars > 256**group_bytes(basemaxg, base85, z85) an OCR misread could decode a group past its byte range;int.to_bytesthen raisedOverflowError, which sailed past everyValueErrorerasure handler and aborted the whole restore. Such a group now raisesValueErrorand is absorbed as an ordinary erasure. --descanauto-retry no longer re-OCRs the sharp pass. The failure-path retry ladder re-ran radius 0.0 over every page even though the initial sharp sweep was already computed; the retry now OCRs only the additional radii (0.6, 0.8) and merges onto the existing sharp results as the ordered spine — one full OCR sweep saved per retried document, byte-identical outcome.tools/ocr_font_report.pywhitelist quoting: an alphabet containing a shell-special character (e.g. an embedded") crashed pytesseract's internalshlex.splitbefore OCR ran; the whitelist config is now shell-quoted.
CI
- PyPI publish and GitHub release no longer depend on the docs deploy.
A GitHub Pages/docs failure blocked the v0.1.0 package publish twice;
docs-build/docs-deployare now a parallel, best-effort branch off the release-critical path.
Added (earlier in this cycle)
- OCR per-character confidence drives char-level erasure marking (plan
3). A CRC-failed line used to erase its ENTIRE byte span for the
document-level Reed-Solomon tier, even though the typical cause is one or
two misread characters — the erasure budget was consumed far faster than
the true error mass.
OcrProvider.ocr_imagenow returnsList[OcrLine](text + optional per-character confidence, 0..1 orNone) instead ofList[str]— a breaking change to the provider contract, fine pre-1.0. Tesseract builds confidence fromimage_to_data(word-granularity, broadcast per character — Tesseract's stable API has no finer level); EasyOCR/PaddleOCR broadcast their own per-segment/per-line score.Base16GCodec.decode_spoolaccepts an optionalchar_conf(keyed by physical line order, since a CRC-failed line's own index label may be corrupt): for a failed line with usable confidence, only the byte offsets its low-confidence characters map to are marked as erasures — the line's other bytes enter the RS stream as ordinary (unverified) data. A two-pass, block-local safety valve makes this strictly no-worse than today: if a block still fails RS with the narrower erasure set, it is retried with the touching line(s) promoted to a full-span erasure (today's behaviour) before giving up. This is a hint about erasure position only — acceptance is still CRC/RS/SHA-256, never guessed.char_confabsent (the default) is byte-identical to a build without this feature. Newtools/conf_calibration.pymeasuresP(char wrong | conf < t)and recall of wrong characters against Tesseract to calibrate the default threshold (ships at0.6pending a real calibration run — see the tool's own docstring).
Fixed (earlier in this cycle)
- Default
--line-parity(2) broke restore on the primary OCR path (breaking, pre-release — no compat shim). The constrained Tesseract character whitelist used for scanning strips ALL interior spaces from every printed line, producing the "compact frame" formsplit_frame/split_frame_with_parityalready tolerate. But the reader determined the printed per-line Reed-Solomon field's width by counting whitespace tokens per line (_detect_line_parity_chars), and a one-token compact line cannot vote — every line silently fell back to width 0, folding the line-parity characters into the payload and corrupting the byte count for the whole stream. At 3pt OCR-B end-to-end,--line-parity 0round-tripped fine while--line-parity 2(the default) and4both failed withcannot recover RS parameters: data/parity line counts are inconsistentor a per-line CRC/RS-budget error — the default was worse than off. The protected machine header (layout.py) now carriesnsym_lineas an authoritative field (_machine_header_bytes/_decode_machine_headergain one byte); the restore path (restore/decode.py) reads it from the CRC/RS-protected header — decoded before any payload line is even classified — and passes it straight toRadixCodec.decode_spool, which prefers it over the token-counting heuristic._detect_line_parity_charsremains the fallback for headerless/raw-codec callers (e.g.glyphive inspect's read-onlydescribe_line_stream). Existing documents from before this change do not decode (the machine header envelope grew one byte).
Changed (earlier in this cycle)
-
Page parity lifts the 255-page cap:
--parity-pagesnow supports up to 65,535 total pages (breaking, pre-release — no compat shim). Document-level whole-page recovery (codec/pagers.py) previously required data pages + parity pages <= 255 (one Reed-Solomon symbol per byte, GF(2^8)) — too small for large archives (a 30 MB tree at ~3 KB/page is ~10,000 pages). Page parity now automatically switches to a GF(2^16) field (one symbol per pair of bytes) whenever the total exceeds 255, raising the cap to 65,535; the protected machine header gains apgpar_fieldbyte (8 or 16) recording which field a document uses, andglyphive inspectreports it. GF(2^8) documents (<=255 total pages) are unaffected byte-for-byte. Existing documents from before this change do not decode (the machine header envelope grew one byte to carry the new field). -
Wire format hardened: interleaved parity, kind-covered CRC, optional per-line Reed-Solomon (breaking, pre-release — no compat shim). The
base16g-crc16-rscodec (and every denser radix codec sharing its engine) fixes three defects found while measuring recovery under a substitution-error channel, at zero size cost for the first two: (1) parity byte interleave — the document-level Reed-Solomon parity stream is now written symbol-major (parity byte j of block b atj*nblocks+binstead ofb*nsym+j), so one corrupted parity LINE spreads its damage across every block instead of wiping one block's entire parity budget outright; (2) kind-covered CRC — the per-line check field (and the H/T/Q machine frames inlayout.py) now covers the leadingL/P/H/T/Qkind letter, so a misread that flips one kind into another now fails its own CRC instead of silently producing a CRC-valid phantom line; (3) optional per-line Reed-Solomon — each printed line may now carrynsym_line(0, 2, or 4; default 2, exposed ascreate --line-parity) extra parity bytes over its own index token and payload, self-healing many single/double-character OCR errors in place before they ever touch the document-level RS erasure budget. The group header grew one byte (B1 | version | nsym | nsym_line | orig_len, 8 → 9 bytes) to record the new field; existing documents from before this change do not decode. Measured effect: the recovery cliff (channel_sim, 30 KB docs, 12% document parity) moves from failing at 0.1% substitution error to succeeding at 0.5% at thensym_line=2default.
Fixed (earlier in this cycle)
- Decode hardening (rescues documents that previously hard-failed). A CRC-failed line whose index token was misread used to be trusted for stream geometry, so one bad character could claim an impossible index (~900,000 vs a true max ~1,000) and make decode die with "cannot recover RS parameters" even with the parity budget barely touched — every ~30 KB document failed at 0.1 % character error. Decode now (1) attempts a CRC-guided single-substitution repair of each failed line (accepted only when exactly one candidate reproduces the printed CRC — the CRC is the oracle, never decompressibility), (2) computes stream geometry only from CRC-valid lines and positionally reassigns or drops implausible-index failed lines, and (3) degrades a conflicting-duplicate collision to an erasure (which Reed-Solomon rebuilds) instead of aborting the whole decode. Applies to all existing documents; no bytes on paper change, and the clean-transcript fast path is untouched.
0.1.0 - 2026-07-18
The first public release provides an end-to-end path from a file tree to OCR-friendly printable pages and back to a verified tree.
Changed
- Compact
#!glyphiveheader line: the display-only summary is now#!glyphive v<N> <codec>[,<comp>] files=.. bytes=.. pages=..— a barev<N>token, codec and compression collapsed to one positionalcodec[,comp]token, andsha256/metadropped entirely (they live in the protectedHframes). Fewer characters on page 1 means less to OCR and less display-line overflow. Any line beginning with#!is now treated as a comment on the read path, so documents can carry arbitrary#!notes. (Pre-release format change; no backward compatibility with the old header grammar.) --descan autoretry now sweeps a0.6/0.8blur ladder (was a single0.6pass): the widest glyphs (e.g. Courier 12 pt) can need~0.8to decode from a raw photo, so the auto-retry covers both. The per-line CRC merge keeps this strictly additive — extra blur passes only ever recover more lines.- Default PDF font is now
dejavu-sans-mono(was Courier). DejaVu was one of only two fonts that held up on real photographed scans undertesseract-glyphiveand passes the byte-for-byte restore gate, so it is preferred for recovery robustness. Trade-offs, accepted deliberately: it renders ~26% fewer usable bytes/page than Courier at 8 pt+ (more pages per document) and embeds ~340 KB per PDF (Courier is a zero-embedding core font). Pass--font courierfor the denser, zero-embed alternative. The synthetic density comparison and the override rationale are inbenchmarks/results/FONT_CANDIDATES.md. - Footer-hash mismatches are now advisory, not warnings: an OCR restore
almost always produces a footer-hash mismatch (OCR inserts interior spaces
that change the page text hash while the
L/Plines still decode via CRC/RS), so logging it at WARNING cried wolf on every clean restore. These are now collected separately (meta["_footer_hash_notes"]) and logged at INFO (shown with-v); genuine page-integrity issues (missing/reconstructed pages) stay at WARNING.glyphive inspectreports the advisory count. Correctness is unchanged — it never rested on the page footer hash. - Faster decode on clean input: decode now skips Reed-Solomon entirely on erasure-free blocks (the common case — text transcripts and good scans are mostly clean), since a line whose per-line CRC matched is already trusted and RS has nothing to correct. A partially-damaged stream RS-corrects only the blocks that actually contain a bad/missing line. Correctness is unchanged: the per-line CRC oracle plus the whole-document SHA-256 gate still catch any corruption loudly. (Profiling put RS at ~75% of clean-decode time; the exact speedup is measured in CI, not stated here.)
- Codec identifier renamed
g1->base16c-crc16-rs: exposes the composable parts (16-char OCR-safe alphabet / CRC-16 / Reed-Solomon) instead of an opaque version tag.codec/g1.py->codec/base16c.py;G1Codec->Base16CCodec;--codecdefault andcodec.names()updated. Nothing was published under the old name, so this is an in-place rename with no migration path or dual-format compatibility. - Dense preset documented: the bundled
ocr-bfont (SIL OFL 1.1, OCR-B by Raisty) at 6pt measures 5,050 usable bytes/page — 23% denser than the Courier 8ptsafedefault — and remains safe on both tested engines (Tesseract and PaddleOCR). Select it with--font ocr-b --font-size 6; the shipped default stays Courier 8pt per the project's font-selection design.
Fixed
- PDF frames no longer silently shrink to fit: a machine/data frame
(
H/L/P/T) that overflows the printable width now raises a clear error naming the overflow instead of quietly reducing its font size to fit. Silent shrinking distorted glyphs (hurting OCR) and hid a misconfigured font/size/margin/line-width. The display-only#!glyphivehuman header is exempt and still scales to fit (restore never trusts it). - Overwrite publication now rolls back on partial failure: restoring onto
an existing destination with
--overwritemoves each existing final file aside into a private backup before replacing it. If any staged move fails partway through (disk full, permission error, etc.), every file already replaced is restored from its backup and any newly created file is removed -- the destination is left exactly as it was, never half-migrated. - Machine header now Reed-Solomon protected: the real Tesseract 5.4.0
end-to-end gate (
create-> rasterize 300 DPI ->extract --from-images-> diff) found that--compression zstddeterministically failed restore: a wider header caused Tesseract to misread oneHframe's two duplicate copies identically, so CRC+duplication alone could not recover it.layout.pynow adds one Reed-Solomon parity chunk over the header envelope so restore reconstructs a single damaged chunk instead of only detecting it; corruption spanning more than one distinct chunk still fails loud. Both--compression noneand--compression zstdnow restore byte-for-byte through the real OCR gate. - Large-document parity overhead:
parity_rationow targets aggregate Reed-Solomon parity across all GF(255) blocks. The default is approximately 12% for large streams instead of repeating a capped whole-stream budget per block and inflating parity to roughly 65%.
Added
create --no-header: omit the display-only#!glyphivesummary line from page 1 for the tightest possible page. Restore needs nothing from it (all authoritative metadata comes from the CRC-protectedHframes), so a--no-headerdocument restores byte-for-byte identically.- Bundled
dejavu-sans-monoPDF font: DejaVu Sans Mono 2.37 (permissive DejaVu Fonts License) is now a selectable bundled font (--font dejavu-sans-mono), embedded in PDF output. It was one of only two fonts (with Courier) that held up under thetesseract-glyphiveprofile in real scan/restore testing — offered as a strong option; a byte-for-byte restore gate is pending before it is recommended over Courier. Adds ~335 KB to the wheel. (Also fixes a latent bug where a second bundled font would have collided on the hardcoded"OCR-B"FPDF family name.) glyphive inspect— recovery-headroom report: a read-only subcommand that reports how much damage a document can survive without fully decoding or verifying it (so it works on a partially damaged scan a restore would reject, and writes nothing). Shows data/parity page counts (whole-page recovery budget), the realized per-line Reed-Solomonnsym(scattered-damage budget), and which pages are present/missing/reconstructable.--jsonfor machine output;--strictexits non-zero on an already-unrecoverable document. Backed by a new purecodec.base16c.describe_line_streamshape oracle.--line-width auto|max|<int>spellings oncreate:auto(default) is the OCR-measured-safe capacity (≤60);maxfills the largest row that physically fits the font/size/margins (may exceed 60, not OCR-verified, and an error on formats without font metrics); an integer above the safe cap now requires--force. The renderer interface gains a publicgeometric_payload_capacityhook (uncapped fit) alongside the safety-cappedpayload_capacity.- De-scan blur for photographed input (
--descan, auto by default):extract/listapply a Gaussian blur to image and rasterized-PDF input before OCR. The default--descan autodoes a single sharp pass, then automatically retries once with a light0.6blur if that fails to decode image/PDF input — raw phone photos are often too sharp/noisy for the frame CRC/RS and otherwise fail, and the retry costs an extra OCR pass only on failure.--descan 0disables the auto-retry; an explicit list (--descan 0,0.6,1.0) OCRs each image at every radius and merges the CRC-valid lines across passes — different blurs recover different lines and the per-line CRC makes combining them safe, so a document no single blur can fully read may still restore from the union. Automatic OCR-engine selection also prefers the constrainedtesseract-glyphiveprofile over plaintesseract(measured substantially higher scan-restore success). - Whole-page recovery (
--parity-pages K):createcan emit K extra document-level Reed-Solomon parity pages over the data pages, so a document survives up to K wholly lost / unscannable / destroyed pages — not just the scattered per-line OCR errors the per-line RS already fixes. Restore reconstructs missing data pages from the parity pages and reports which it rebuilt. Default 0 (off, byte-identical to before). Data pages + K must not exceed 255. Independent of--parity-ratio. The header gains apgpartoken (omitted when 0) and the protected machine envelope records K and the page block size. Separately, a missing page no longer hard-fails up front even with K=0: the codec's document-wide RS is given a chance to recover it from the surviving pages when the per-line parity budget allows. - Progress logging for
create/extract: both commands now log intermediate stage events (archived,compressed,encoded,renderedforcreate;staged,publishedforextract) instead of only a final one-line summary, sparsely rate-limited so a large tree doesn't flood the log. The underlyingon_progresscallback is also available to library callers ofrestore_document_spooled/unarchive_spool. - Isolated QR transport primitives: the optional
qrextra provides deterministic, versioned 1000-byte envelopes, level-H Segno PNG generation, and Pillow/ZXing-C++ raw-byte decoding without OpenCV. Extract/list accept explicit--from-qrimage or directory input; ordinary image OCR behavior is unchanged. Explicit--format qrwrites six symbols per Letter page;--format hybridwrites one symbol plus its transcript slice per page. A plain.pdfsuffix continues to select the ordinary PDF renderer. - Constrained Tesseract profile (
tesseract-glyphive): an opt-in OCR provider using PSM 6, Glyphive's exact machine alphabet, and disabled general language dictionaries. The existingtesseractprovider remains unchanged. - Bounded archive and compression primitives: archive records can now be
written and parsed as fixed-size chunks, and the built-in none/gzip/zstd
methods support binary stream adapters. Existing one-shot APIs remain
available while create/restore migrate to disk-backed spools. Create now
streams archive, compression, FEC, pagination, and renderer output through
temporary spools;
--temp-dirand--chunk-sizecontrol spool placement and I/O granularity. Restore stream-decompresses into a size-limited spool, validates the global digest and archive framing, stages files privately, and only then publishes final paths;--max-output-bytescaps expansion. Transcript parsing spools normalized frames in one pass, whilebase16c-crc16-rsretains only compact line offsets and processes Reed-Solomon codewords blockwise. - Explicit plugin discovery (
glyphive.plugins): trusted installed distributions can provide typed codecs, compression methods, render formats, or OCR providers through four documented entry-point groups. Discovery is opt-in through the library API or global--pluginsCLI flag, deterministic, cached, and reports broken candidates without changing normal imports. - Standalone zipapp packaging (
package.py): build a universalglyphive.pyzcontaining the required runtime dependencies and the core text/none/gzip feature set, or explicitly named platform-specific artifacts with optional integrations. Lightweight OCR Python shims are optional; heavyweight OCR engines and models remain external. - Codec registry (
glyphive.codec): typed named lookup with the built-inbase16c-crc16-rscodec. - Compression registry (
glyphive.compression): namednone,gzip, and lazy optionalzstdmethods. - Renderer and OCR registries (
glyphive.render,glyphive.restore.ocr): text, PDF, Word, and optional OCR providers use explicit registries with lazy backend imports. - Codec
base16c-crc16-rs: the measured-safeABCDHKLMPRTVXY34alphabet (16 symbols, 4-bit packing), a full CRC-16 per line, masked five-character indices, and document-wide interleaved Reed-Solomon parity. Scattered OCR errors can self-heal; correctness is judged by CRC and parity rather than speculative character substitution. (Named for its composable parts — 16-symbol OCR-safe alphabet / CRC-16 / Reed-Solomon — instead of an opaqueg1version tag; renamed before the first release, so no migration shim was needed.) - Binary-safe archive stream (
glyphive.archive): length-prefixed records for arbitrary bytes, deterministic ordering, root-level.gitignore/.ignorefiltering, empty-directory records, andnone,gzip, orzstdwhole-stream compression. - Protected layout metadata (
glyphive.layout): safe-alphabetHheader frames andTpage-footer frames carry authoritative document metadata, page identities, and hashes. Unrestricted human-readable summaries are display aids only. - Text, PDF, and Word renderers with selectable font family and size. PDF output uses built-in FPDF font families; Word output accepts installed Word font names.
- Verified restore pipeline (
glyphive.restore): transcript decode, whole-document SHA-256 validation, path-traversal-safe extraction, and an optional multi-provider OCR layer for image input. - Tar-like CLI:
create(c),extract(x), andlist(t) commands with codec, compression, metadata, renderer, and OCR selectors. Leading-c,-x, and-tmode flags work without a positional command. - Automatic document input:
extractandlistclassify each direct-child input independently by magic bytes, extension, then UTF-8 text, so transcript, image, PDF, and DOCX pages can be mixed in deterministic order. A conversion helper renders PDF pages or diagnostic DOCX transcript pages to PNG with configurable DPI and blur; DOCX restore usespython-docx, not an office suite. - Capacity-aware rendering: PDF/DOCX row budgets follow font size and page
margins;
--minimal-marginsuses a compact 12-point profile, and long PDF display headers fit rather than clip. Output format is inferred from-fwhen--formatis omitted. - Bundled OCR-B PDF font:
--font ocr-bembeds a pinned, unmodified SIL-OFL-1.1 font; PDF output also accepts explicit.ttf/.otfpaths. - Documentation and examples: task-focused create, restore, wire-format, OCR, benchmark, and API pages plus a runnable create/restore example.
Changed
- Require
pathlib_next>=0.8.1, including its Python 3.9 import and local path walking fixes. - A directory supplied to CLI
-fexpands its direct child files in stable filename order.
Known limitations
createarchives one directory (or.) at a time; wrap multiple inputs in a directory.- Ignore files are read at the archived-tree root only; nested
.gitignorefiles are not applied. - A fully lost or unscannable page is detected and reported, not reconstructed.
- Protected header/footer metadata detects corruption but is not itself Reed-Solomon-corrected.
- QR-code output is not implemented.