Chosen: a stdlib BMP reader, because `read_image` is on the CORE path and an
asset's name is its content digest. Measured first, as the order requires:
Pillow 12.3.0 IS in this tree (transitively under `pdfplumber`) and it DOES
decode RLE8 correctly -- a hand-written stdlib decoder and Pillow agree on
19 of 19 of R761's real files, RGB per pixel. So the choice does not rest on
capability. It rests on two properties of this package: `.html` and `.xml`
carry images with no `[extract]` extra installed, so a Pillow converter
either makes a core path depend on an optional binary wheel or buys the
second runtime dependency; and encoding through an installed library would
make a bundle's identity move with that library's version, which is the
property 0.10.0 felled page rasterisation over and `encode_png`'s docstring
already defends. Pillow keeps the job it is good for: the INDEPENDENT decoder
in the tests, on neither side of the conversion.
The defect, measured over the frozen R761 delivery's `assets/`, denominator
50: 29 JPEG, 2 PNG and 19 RLE8 BMP. The 19 are byte-correct files nothing
reads, so 19 figures were present and invisible while `images: N` reported
that they had arrived.
- `VIEWABLE_MEDIA_TYPES` is tested against every asset's SNIFFED type, so it
is a property and not a list of formats we met. WebP is on it and `sniff`
does not recognise one; the limit is stated, not implied.
- `bmp_to_png`: 8-bit uncompressed, 8-bit RLE8, 24-bit uncompressed. All five
RLE8 opcodes. 19 of 19 real files convert with RGB identical to Pillow's
decoding of the source, 2 366 365 pixels compared.
- `asset_not_viewable` and `asset_bmp_unsupported`, both published, both
leaving the concept's "not carried" line.
- Traceability on the pointer's second line, where the rest of the asset
metadata already lives: original media type, original sha256 in full, new
sha256 in full. A converted asset is ONE asset.
- The ceiling is paid on the DECLARATION before a row is allocated, and an
RLE run is one clipped slice -- painting pixel by pixel leaves the memory
bounded and the CPU unbounded.
Two repairs the change forced, each measured rather than assumed:
- `tests/test_assets.py`'s "dimensions absent is absent" used a TIFF, which
is now refused before `read_image` returns. The property still has a
reachable case -- a JPEG whose frame header never arrives -- and uses it.
- `asset_holds` in the accounting gate proved a carry by hashing the SOURCE
file, which a converted image's bundle cannot satisfy. It now also reads
the two digests the bundle states and HASHES THE ASSET ITSELF, so a bundle
claiming a conversion it did not perform still fails.
`tools/okf_asset_census.py` is the committed instrument for the
known-positive: one row per image, from two pinned trees. It was caught by
the rule it serves -- its first version handed `_pdf_images` the wrong page
object and reported 0 images over 67 PDFs with exit 0. The attribute is
asserted now and a known-positive runs before the sweep.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Round 3 of the 0.10.1 review, and the finding is the pattern the three rounds
share: each bound an OUTPUT, and the bomb stepped one link along. The
declared size, then the first `FlateDecode`, then every `FlateDecode` -- and
then a link this package had documented as safe.
`ASCII85Decode` was classed as bounded "by its own input because it shrinks".
It quadruples: `z` is the shorthand for four zero bytes. And the output was
never the cost -- `base64.a85decode` appends one 4-byte object per group to a
list, about a hundred bytes of memory per byte of INPUT (101.4x at 1 MiB,
96.1x at 4 MiB, 94.5x at 16 MiB on CPython 3.14).
Paired subprocesses, idle machine, both sides from PINNED trees, the document
built once by a third process and read from a file because `ru_maxrss` never
falls and `b"z" * 64 MiB` alone costs 171 MB:
[/Fl /A85] z x 32 Mi 33 475 B CARRIED 3 261 599 744 -> too_large 42 070 016
[/Fl /A85] z x 64 Mi 66 090 B CARRIED 6 461 558 784 -> too_large 40 280 064
[/A85] z x 8 Mi 8.4 MB CARRIED 933 085 184 -> too_large 62 484 480
[/Fl /A85 /Fl] z x 32 Mi 33 488 B samples_invalid 3 519 180 800 -> too_large 43 438 080
The picture was CARRIED in three of the four: not a bound that fired late, no
bound at all. Doubling the `z` run trebles the old cost and leaves the new one
where it was.
WHY THIS FORM. `assets.MAX_FILTER_DECODE_BYTES` (512 MiB) is what decoding ONE
link may cost -- a separate number from `MAX_IMAGE_BYTES`, because that one
bounds the picture and this one bounds producing it. `FlateDecode` is measured
as it is paid; every other permitted filter carries a MEASURED cost ratio
(`assets.PDF_FILTER_COST_RATIO`) checked against its input BEFORE its decoder
is called, since those decoders take a whole string and return a whole string.
A filter with no ratio is refused unread. The budget TRAVELS: a deflate link
is inflated under the smaller of the picture's bound and what the next link's
decoder may be handed, or `[/Fl /A85]` pays 256 MiB for a refusal.
A chunked ASCII85 decoder written here was the alternative and was FELLED: it
would bound `_check_stream_cost` and not the run, because `stream.get_data()`
decodes the whole chain again with pdfminer's own decoder, and it would make
this package rather than pdfminer the authority on an image's bytes. The cap
is the only number that bounds that. `resource.setrlimit(RLIMIT_AS)` was
MEASURED before anything was built on it, as the order required, and is not
usable: Darwin 26.6.2 raises `ValueError: current limit exceeds maximum limit`
and does not enforce it. No child-process cap exists.
THE CAP IS READ OFF THE CORPORA, the posture `MAX_IMAGE_PIXELS` has: over the
9 668 image objects of the 77 PDFs on this machine, 16 decode through an
ASCII85 link and the largest input to one is 450 739 bytes, against a cap of
about 5.0 MB.
A PROPERTY TEST REPLACES THE LIST OF KNOWN SHAPES: every chain of length 1-3
over the ten filters pdfminer decodes, 1 110 of 1 110, both payload fills,
each delivered under the bound or refused with a published code and never paid
for on the way (`tracemalloc`, which counts allocations and is not disturbed
by load). Known-positive beside it: 258 of 258 chains over the permitted
filters still carry a small image.
MAJOR -- the backstop had no test. `check_payload` at the end of
`_check_stream_cost` could be deleted with the whole suite green, because the
second one after `get_data()` gives the same code one step later. The two
differ in whether the payment was made, so the test asserts `get_data` was
never called.
10 OF 10 MUTANTS KILLED, control green, each killer named in the report. Four
survived a first pass and two tests exist because of it.
NOT ONE PICTURE CHANGES HANDS, MEASURED BY NAME: `_pdf_images` over every PDF
on this machine from both pinned trees -- 9 306 -> 9 306 carried over 77
files, 50 -> 50 on R761, 0 of 78 files moving a count and 0 moving a code.
R761 also settles a question raised while this order was open: 50 objects, 29
[/DCTDecode], 21 [/FlateDecode], 0 ASCII85 links -- so round 2's count of 580
`[/FlateDecode /ASCII85Decode]` objects is reproducible from nothing on this
machine. It changes no decision; a bomb shape does not need a corpus.
Version stays 0.10.1, no tag. README, CHANGELOG, CLAUDE.md and errors.py
corrected TO what the code does; the round-2 report carries a correction block
rather than a rewrite. Report:
docs/2026-09-18-utgangsbudsjett-per-ledd.md
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The two findings of the 18.09 PM checkpoint of `0f308c1`. Red tests landed
first in `3b587ea`; this is what turns them green.
BLOCKER -- `_check_inflated` read `filters[0]`, measured that one link and
returned, which is not a bound: a PDF decodes a stream through a LIST of
filters. Measured in paired subprocesses from two pinned trees, idle machine:
[/FlateDecode] 400 MB 408 516 B 59 232 256 -> 62 017 536 B
[/FlateDecode x2] 400 MB 1 636 B 886 554 624 -> 52 367 360 B
[/FlateDecode x3] 400 MB 1 070 B 889 393 152 -> 61 390 848 B
[/FlateDecode x2] 1,2 GB 2 927 B 2 567 204 864 -> 60 403 712 B
542 000x the file at two links, and the picture WAS refused at the end -- by
`check_payload` after `get_data()`, once the memory was spent. The single-link
row is the control and does not move. It also left the 16 corpus objects behind
an `[/ASCII85Decode /FlateDecode]` chain unmeasured, since `filters[0]` is not
`FlateDecode` there.
`_check_stream_cost` walks every link. THREE CLASSES and no fourth
(`extract.bounded_pdf_filters`, pinned by a test): `FlateDecode` MEASURED, a
link with another expanding link behind it inflated under the same bound and
handed on; `ASCII85Decode`/`ASCIIHexDecode` bounded by their own input because
they SHRINK; `DCTDecode`/`JPXDecode`/`JBIG2Decode` PASS THROUGH. Everything
else -- `LZWDecode`, `RunLengthDecode`, `CCITTFaxDecode`, `/Crypt`, anything
written later -- is refused UNREAD with a new code `asset_pdf_unbounded`,
decided before the FIRST link is decoded so a document cannot make the run pay
for the links in front of the one we cannot bound. An encrypted stream is
deciphered and then measured, where `stream.decipher is not None` used to
return unmeasured; 0 of 5 142 objects here are in an encrypted document, which
is why nothing caught it.
NOT ONE PICTURE CHANGES HANDS, AND IT IS MEASURED BY NAME. Every PDF on this
machine -- 78 documents, K2 in both trinn1 and trinn2, the shipped fixtures and
R761 -- run through `_pdf_images` page by page from both pinned trees:
images carried 9 356 -> 9 356
documents losing one 0 of 78
documents gaining one 0 of 78
asset_pdf_unsupported 322 -> 314
asset_pdf_unbounded 0 -> 8
The 8 are the 4 `CCITTFaxDecode` stencil masks (`/ImageMask true`,
`/BitsPerComponent 1`), counted twice because trinn1 and trinn2 hold the same
document. They were refused before and are refused now, one step earlier and
under a code that says why.
MAJOR -- `check_payload(len(data))` after `get_data()` is the counted refusal
four documentation surfaces point at, and deleting exactly that line passed all
2 132 tests. It is reachable through a stream pdfminer has ALREADY decoded
(`decode()` sets `rawdata` to `None`), which is now the ONLY case outside the
bound and has a test.
Eight mutations, one line each, every one DEAD, with the unmutated tree run
first as the control: first-link-only, loop dropped, inequality reversed,
encrypted skipped, backstop deleted, unknown filter passed through,
intermediate link not carried forward, whole check removed.
`tools/okf_accounting_gate.py` gains one line, the new code in
`REJECTION_CODES` -- what a rejection code requires and nothing more. Gate
unchanged: exit 1, GATE RED rows 2, 3, 6. Version stays 0.10.1, untagged.
Suite after `git add` against a clean tree: `uv run pytest -q` ->
2152 passed, 1 skipped (226 s). ruff, ruff format --check, mypy --strict clean.
Report: docs/2026-09-18-filterkjeden-og-backstoppen.md
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A second independent review read `230d1cb` -- the commit that closed the
`v0.10.0` review's two MAJOR findings -- and found one of them open. The
bound read `/Width` and `/Height`, which an untrusted document writes, while
`get_data()` pays for the stream beside them; `/Length` is the COMPRESSED
length and the two numbers are independent.
Re-measured here on `ed8d9d7` before anything changed, in its own
interpreter: a 408 516-byte PDF declaring 1x1 and carrying 400 MB of deflated
zeros was CARRIED, no rejection, 891 904 000 B peak RSS. After: 0 carried,
`asset_too_large`, 57 065 472 B. At 1,2 GB of zeros, 2 436 MB -> 64 569 344 B
-- the cost no longer scales with the bomb. End to end through the CLI with
the shipped defaults: 838 000 640 B and an asset written -> exit 0,
79 650 816 B, `0 carried of 1 found`, no `assets/`.
Three numbers are bounded now, not one: what a container DECLARES, what a
carried FILE measures (`read_image`, so a 49 MP PNG of 47 705 bytes is not
passed on to a consumer), and what a PDF stream DECOMPRESSES to
(`assets.inflated_size`, a chunk at a time, output discarded, before
`get_data()`). The limit is stated rather than implied: the stream
measurement runs where `FlateDecode` is the first filter and the document is
not encrypted; every other chain is a check on the decoded length AFTER the
decode, a counted refusal and not a bounded one.
A non-positive declared dimension is `asset_size_invalid`, its own code,
raised before the stream is read. `-1 x 40000000000` is a NEGATIVE pixel
count, under which every `>` bound read as satisfied, so the check returned
silently and the refusal arrived from `encode_png` as
`asset_samples_invalid`. Its own code because a publisher shipping a picture
bigger than this package carries and a dictionary written to be read wrong
are different facts about a document.
Two smaller findings in the line that says what is missing, both introduced
by the first fix: the address was written twice, once bare, and a linkifying
renderer autolinks a bare URL -- written once now, in one code span; and
`label` became a dead parameter, so the figure's caption was dropped, a
regression against 0.10.0. It is written again in the `-- <label>` form a
carried pointer uses.
Version bumped to 0.10.1 across all ten places. Nine were unbound and stale:
four README install lines naming the previous release, two prose lines, the
"current tag" entry, `uv.lock`, and a CHANGELOG whose 0.10.1 content sat
under `[Unreleased]`. Two new packaging tests bind them to `__version__`, and
the README's guard tag to `[tool.uv.sources]`.
Every test was red first. The fate of every image is identical with and
without the new bound on three K2 PDFs carrying 800 images (464/464, 311/311
with the same 12 rejections, 25/25), and the second inflate is below the
noise floor there. 0 shipped artifacts move: no bundle under `examples/`,
`skills/` or `tests/fixtures/` carries an image pointer at all, measured
against a known-positive control.
`asset_too_large` was undocumented in the error registry; both codes are
there now. `tools/okf_accounting_gate.py` gains the new code in its closed
list -- one string, no behaviour change, stated because that file belongs to
another order.
Suite 2141 passed / 1 skipped, ruff + format + mypy --strict clean.
Report: docs/2026-09-18-bildestien-holder-0-10-1.md
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Until now no reader in this package fetched, named, described or copied a
single image. `<img>`'s attributes were never read, a NISO-STS `<graphic>`
was walked past, a PDF was opened for its text alone, the converter's
markdown writer dropped every picture, and the only writer into a bundle
took `content: str`. The two lossiness warnings said so on every run, which
made the loss honest and did not make it smaller.
Measured on R761 Prosesskoden:2025, published as a 701-page PDF and as a
NISO-STS delivery: the process text is carried in full while 12 `Tabell N-N`
and 9 `Figur N-N` captions stand over nothing, because that publisher ships
those tables as raster pictures in both. Process 84's "toleranseklasse ...
er gitt i tabell 84-2" points at empty space.
THE GATE WAS WRITTEN FIRST AND RED. `tests/test_asset_gate.py` reads its
denominator out of the source (`page.images`, `word/media/`, `ppt/media/`,
`<img`, `<graphic`), never from a constant here. Measured at 332961a, built
from `git archive` and not from the editable tree: carried 0 of 8 local
images across 5 documents (9 declared), and no `assets/` at all. After: 8 of
8, with the ninth a remote source carried as a pointer without a file.
FIVE READERS PLACE, ONE MODULE DECIDES. `assets.py` owns what an image is
(sniffed from the bytes, never from the claimed extension), what it is
called (`<sha256[:12]>-<the source's own basename>`) and how it is pointed
at (one two-line block, one regex). `.xlsx` is deliberately not a row: a
block inside its pipe tables would break the `source_rows` locator, and 0 of
4 K2 workbooks hold media.
A PDF stream that is already a file is carried VERBATIM (29 of R761's 50
objects are DCTDecode); raw samples are encoded to PNG with stdlib zlib, so
no new dependency. Rendering the page region was the alternative and was
felled on determinism: a rasterised crop's bytes, and therefore the asset's
content-addressed name and the bundle's digest, would depend on the
installed rasteriser. What the encoder cannot express exactly is refused
with a code and counted, never approximated.
NO SIZE FLOOR, and that is a measurement: over the 4 828 image objects of
the K2 corpus the size distribution is a broad spread with no gap, unlike
OCR_CID_SHARE's bimodal one, so a threshold would be a number we chose.
ON BY DEFAULT, AND THE CONTROL IS TWO WHOLE BUILDS. The 43-document
reference corpus at 332961a versus rebuilt at HEAD with `--no-assets`:
865 files on both sides, `diff -rq` reports ONE difference, the added
`Images: NOT CARRIED` line in log.md. Every concept byte-identical.
Against the default: 453 -> 454 concepts, 865 -> 867 md, 0 -> 2 964 assets
(2 964 carried of 3 145 found, 4 622 pointers), 4.7 MB -> 115 MB, 2 414 s ->
3 088 s, peak RSS 6.26 -> 8.74 GB, 422 of 865 md files differ. The one new
concept has a measured cause: the pointers are body text, so a section
holding 146 of that document's images grew from 19.0 % to 30.6 % of the
extracted text and crossed `--outline-gate`'s 0.20 share clause.
THE IMAGE BYTES ARE NOT SCREENED. The guard is text-only, the pointer block
passes the gate as body text, the picture beside it passes nothing, and
log.md says so on every run.
Also fixed, both found by measuring rather than by reading:
- a markdown image is no longer read as a cross-reference. `structure._LINK`
never looked at the character in front of the bracket, so every pointer
would have arrived in the index as an edge to a concept that cannot exist.
- Door C carries the assets its merged concepts point at. Before this,
importing a bundle built with `--assets` merged 6 of 6 concepts and wrote
no `assets/` at all, so every pointer named a missing file.
Report: docs/2026-09-17-bilder-i-bundlen-trinn1.md
Spec proposal: docs/plan/okf-assets-section-6-4.md
Suite 1 955 passed / 1 skipped (from 1 896), ruff and mypy --strict clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A PDF carries no notion of a heading -- a heading in a PDF is a typographic
fact -- so the text stream `pdfplumber` hands the segment proposer has already
thrown away the only evidence there was. The `docx` path never had that problem:
the converter emits ATX headings and `_ATX` cuts on them. Two readers close the
gap, and both are OFF.
`--pdf-headings font` infers a heading from the conjunction this repository
already measured (size above the document's character-weighted body median AND
a bold font name, recall 1.000 / precision 0.846) and emits it as ATX in the
SAME markdown the office path produces, so `_ATX` applies unchanged and no
PDF-only heading grammar exists.
It stays off BY MEASUREMENT, and the measurement is the point of the round:
against the operator's unit worksheet it takes `pdf` from 2 of 8 to 0 of 8,
losing two exact matches. The mechanism of the loss is stated rather than
guessed -- on those documents the outline rule already recovers the document's
own numbered chapters, so a second heading source can only add. Whole-corpus
screen: 25 of 32 `pdf` change, 0 of 5 `docx`, 0 of 2 `xlsx`. The default bundle
is byte-identical before and after this commit (`diff -r`, exit 0).
`--ocr` reads a page as an image when its own text never arrived: empty, or
`(cid:N)` placeholder codes at or above a threshold READ OFF a measured
distribution -- 834 pages over 32 files, 818 at exactly 0.0 and 16 at 0.93 or
above, nothing in between. On the one corpus document with the failure: 95.07 %
cid to 0 %, 44 to 2561 words of four or more letters, 17 to 18 pages with text.
Its engine is an optional dependency group and never a runtime dependency; a
packaging test pins both halves, and without the group every affected file is a
coded rejection (`extractor_ocr_group_missing`) rather than a crash.
Also corrects two stale published facts found while measuring: the README still
said two segmentation rules were on by default after `f6fea13` made it three,
and CLAUDE.md's K2 digest named the round-3 default. The current default is
492 concepts / 944 files, `bdefa679...`.
Report: docs/2026-09-08-k3-runde4-pdf-skrift-og-ocr.md
Co-Authored-By: Claude <claude-opus-5>
PM decision B6 asked for a list-taking _render_sources so a concept can record
more than one source, and prescribed the block list as the emitted form. The
list is delivered; the block form is not.
Three measurements, not an argument. Our own parse_frontmatter skips indented
lines, so a block list round-trips to an empty value with every entry silently
gone -- and _is_ingest_owned reads through that same parser. The consumer B6
was written for accepts the multi-entry flow sequence and classifies a block
sequence as unreadable provenance, so block would hand it exactly the state it
cannot read. And B6's own acceptance test asks for a round trip through this
parser, which no block form can pass.
A single source renders byte-identically, so all six goldens are unmoved. The
unquotable-value gate now runs on every entry, not just the first. New code
sources_empty refuses an empty list.
1023 -> 1034 tests, including the negative control that pins the block form's
silent data loss.
Arm E, capability only. A profile MAY name a renderer per suffix; no
domain-aware renderer is written here, that stays a Non-Goal, and `_RENDERERS`
is empty on purpose so the emptiness reads as a decision rather than an
omission.
THE LAYERING IS THE DESIGN, not an implementation detail. `extract.py` is the
extraction registry and must not import the contract layer, or the dependency
runs backwards and the registry stops standing on its own. So `extract_text`
gains a keyword-only `renderer: Callable[[str], str] | None`, knowing nothing
about profiles, and `inbox.py` -- which already holds the profile at that call
site -- resolves a NAME to a function. A test asserts extract.py still contains
no reference to the profile layer, because that constraint is the whole reason
the parameter is shaped this way.
The renderer runs AFTER extraction, never instead of it, so it never has to
re-implement a reader and the two cannot drift. The default is identity, which
is what keeps the five byte-pinned goldens byte-pinned -- asserted per suffix
rather than once.
An unknown renderer NAME is refused rather than falling back to identity: a
silent fallback would produce a bundle that looks rendered and is not, which is
the failure mode this arm exists to make visible. That needed a registered code
(`unknown_renderer`) and its test -- slightly beyond the step's named files,
but the capability cannot ship without defining what an unknown name does.
`tests/test_profile.py`'s exact-field-set assertion went red, as the plan's risk
table predicted. Updated deliberately with the reason recorded: that assertion
exists so a field cannot arrive without someone deciding it should, and its red
run is the mechanism working.
Suite 917 -> 926. All five goldens byte-identical.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Registered before any step raises them, so no later step invents a spelling:
- `extractor_binary_missing` -- the converter binary is absent (distinct
from the extra not being installed)
- `extractor_binary_version` -- present, but not the pinned version
- `extractor_convert_error` -- the converter failed on this file
- `extractor_empty_conversion` -- the converter returned no text
Also widened the `extractor_extra_missing` type list, which still read
"pdf/docx/xlsx".
Denominators recounted after the change rather than carried from a note --
the stale 49/48 figure is what made the recount a step requirement:
code bullets in errors.py 50 -> 54
distinct codes 49 -> 53 (one code documented twice)
test definitions in this suite 56 -> 57 (one parametrized definition,
four cases -- one per code)
The four tests assert only what is true at this step: the code is documented
and an ExtractionError carries it. The resolver and the seam replace each with
a behavioural raise-site test; a code that never gains one stays visible here
as a test that still only reads a docstring.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A plan is selected by content hash, so a mistyped source_sha256 matched
nothing, every dropped file fell through to the one-concept rule, and
process_inbox returned an ordinary success over a flat bundle. The
operator asked for segmentation, got none, and had no error to read --
the silent skip this library refuses everywhere else. vegnormal-okf is
about to run an N500 corpus through this path, where a silent zero
would read as "the corpus has no concepts".
The refusal asks whether a covering plan was FOUND, not whether every
file was examined, so an unreadable drop cannot mask it; and coverage
is recorded at selection, not after path validation, so a matched plan
with a refused entry path still reports its own per-file code. The
first cut got that second question wrong and an existing collision test
caught it; the case is now pinned by its own test, verified red against
the earlier form.
New code segmentation_plan_unmatched, registered in the SegmentationError
docstring register in the same commit. Fail-fast before any disk
mutation. The four existing goldens are byte-identical to baseline.
Co-Authored-By: Claude <claude-opus-5>
Order G2a. Populates the optional `[extract]` extra for the first time with
one parser, `pdfplumber>=0.11.10,<0.12` (MIT), and wires `pdf` through it.
The default install is untouched: exactly one runtime dependency, stdlib
otherwise, enforced by test_packaging.py.
The gate for `pdf` becomes an import probe rather than a frozenset membership
test, exactly as extract.py's docstring had promised. The rejection does not
change: without the extra, `pdf` still raises `extractor_extra_missing` with
the same message. That behaviour is asserted UNCONDITIONALLY via a sys.modules
monkeypatch, so it holds on machines where the parser is installed too — a
skip would have preserved nothing there. Verified in a clean venv without the
extra: 589 passed, 7 skipped; with it, 596 passed.
`docx`/`xlsx` are unchanged and still fail fast — the extra names exactly what
it ships.
The parser choice was forced by measurement, not preference (b73dd9d,
docs/2026-08-21-g2-pdf-extraction-measurement.md): on a real requirement table
pdfplumber keeps 4 of 4 rows with label and value on one line, where pypdf,
pdfminer.six and pymupdf each keep 0 of 4. pymupdf is additionally out on
licence (AGPL-3.0), which an MIT package must not push onto a consumer.
Three facts from that measurement are now carried in code rather than in a
report:
- Extracted text is pinned to an exact transitive parser version
(pdfplumber pins pdfminer.six==20260107; date-stamped, no stability
contract). tests/test_extract.py freezes the expected text of a committed
hand-written fixture so a parser upgrade breaks something visible instead of
drifting silently. Reasoning at the declaration site and in
tests/fixtures/README.md.
- Determinism within a version is now held by a test, not only measured once.
- Drawn content does not survive extraction. Every pdf extraction emits the
new `ExtractionWarning`: figures have no text to recover, so a bundle built
from drawn documents is incomplete by construction. Stated categorically
rather than detected — deciding "is there a figure here" is the layout
heuristic G2b declined.
Two new error codes, both mirroring existing patterns: `extractor_empty_pdf`
(a scanned/image-only PDF, refused rather than persisted as an empty concept)
and `extractor_pdf_error` (parser failure wrapped, never leaked).
Structured table recovery (G2b) is NOT implemented and is documented as out of
scope: two independent parsers return the same wrong shape, so the breakage is
document geometry, not a library choice. PDFs enter as prose.
Also corrects an install promise this change would otherwise have published:
the README no longer presents a bare `pip install 'llm-ingestion-okf[extract]'`
as working, because the package is not on an index.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HtNhsdHnMGtMi7U2mvMU8z
The v0.2 profile lands additively: DEFAULT keeps stating commons' §5 layer
byte for byte (the golden suite is the proof, not the claim), and OKF_V0_2
adds the families v0.2 introduces on top of DEFAULT's key order.
Two questions the plan left open were the operator's, and both were decided
before code rather than discovered during it:
- `generated.by` is `process:llm-ingestion-okf` — plan V1's option (d), the
§7 process actor form. The value carries no version, which is what keeps a
byte-compared fixture stable across releases and leaves a shared
cross-implementation fixture possible. A-E3 was written against the
`<producer>/<version>` form and is now STALE in the pilot specification; the
correction is owed to portfolio-optimiser-claude before they run.
- `sources[].resource` is the manifest source's locator verbatim: the file
root, the sql connection_ref (an env-var NAME, never its value), or the http
base_url. `credential_ref` is not a locator and is never emitted.
Ownership becomes a policy on the profile rather than a literal in the gate.
The emitter and `_is_ingest_owned` are coupled through the stamp value, so
OwnershipPolicy is where they meet and can only change together. Recognition
is one-way by decision: OKF_V0_2 owns the v0.1 stamp too, so a DEFAULT-written
bundle re-runs IN PLACE (operator, 2026-07-27), while DEFAULT still refuses a
v0.2 file rather than replacing it — V-A3's fail-safe is preserved.
An unquotable locator is refused rather than emitted. Measured with PyYAML
rather than reasoned: `[{ id: a, resource: data, backup }]` raises nothing and
parses to a mapping with a `backup` key nobody wrote, so the failure mode is a
silently wrong provenance record. Validation, not repair.
Deliberately NOT here: the `okf_version` declaration. §12 makes it a MAY and
none of upstream's four reference bundles exercises it; WHERE it goes is open
between upstream's root-index frontmatter and catalog's body-line convention,
and catalog verifies against upstream first. It is declared once, at D5.
V-A5 is extended to the new profile — the one whose NAME is the place a
version literal would look natural, and it carries none.
542 tests, mypy --strict clean.
Reads an external OKF bundle as {bundle-relative path -> document text},
hands it WHOLE to an injected gate over the guard's okf.import_bundle (a
bundle-level call: it resolves the cross-link graph across concepts), and
merges only concepts clearing the non-blocking floor. Same injection pattern
as Door B, so the core stays dependency-free while the CI channel for the
real guard is settled.
Two constraints shaped the design and are pinned by tests:
- A merged concept is written VERBATIM. Stamping provenance into it would
require round-tripping its frontmatter through this library's line-oriented
parser, which cannot represent the block lists the guard's parser accepts --
silent data loss -- and would persist bytes the guard never screened.
- Ownership is therefore proven by content identity: identical bytes at the
target name are a no-op re-merge (re-import of an unchanged bundle is
idempotent), and anything else at the name is refused. Curated content and
an updated concept are refused alike; refusing is what never destroys.
The floor is fail-closed beyond the plan's "no error" wording: an error, an
unrecognised disposition, and a concept the gate returned no verdict for are
all refusals. quarantine_review stays its own bucket, as at Door B.
origin/channel are validated against the guard's pinned vocabulary -- it
derives trust from origin by enum identity, so an unrecognised string would be
silently downgraded rather than caught.
Three primitives promoted for reuse rather than duplicated:
reduce_to_id_grammar and check_filename_length to materialize.py, and
extract.decode_text. Door C slugs the WHOLE concept path, so tables/users.md
and views/users.md stay distinct. Concept discovery folds case explicitly
rather than globbing *.md, whose case-sensitivity follows the filesystem and
would import the same bundle differently on APFS and ext4.
README's "what is gated today" section corrected: it claimed nothing is gated,
which is no longer true, but the honest statement is narrower than "the doors
are gated" -- the library cannot verify that an injected adapter is a real
guard, and a permissive stub is believed.
405 tests green; ruff, ruff format and mypy --strict clean.
`process_inbox(inbox_dir, bundle_dir, ingested_at, *, okf_type, gate)`:
per dropped file, bytes -> extract_text -> guard gate -> render -> collision
gate -> write -> index link. Returns an InboxResult whose four buckets
(persisted / quarantined / rejected / failed) are disjoint and complete, so a
file that vanished shows up as a missing entry rather than as nothing.
The guard is INJECTED rather than imported. The library calls no guard
function and makes no security decision: the Gate adapter returns a
GateDecision carrying the guard's own disposition value and the sanitized
text, and the flow only branches on it. That keeps the core dependency-free
while B2 (the guard's CI channel) is still open, and lets the persist and
refuse branches run deterministically against a test double.
Decisions taken with the operator this session:
- What is persisted is the gate's SANITIZED text, not the extracted text.
Screening one string and writing another would make the verdict a statement
about bytes nobody kept. `sanitize` is exported by the guard precisely for
callers composing the checklist themselves, so this is sanctioned API, not
a reimplementation. Door B has no model call, so the fenced text the
bookends produce for a transform is never persisted.
- Quarantine is reported apart from rejection. QUARANTINE_REVIEW means "hold
for human review" — an operator queue — where FAIL_SECURE is a decision.
No quarantine directory in v1; that stays an extension point.
Fails closed by construction: only the guard's non-blocking floor (`warn`)
persists, so a renamed member, a future disposition or an adapter typo lands
in `rejected` rather than being guessed safe.
One bad file never aborts the run. Only three conditions fail the whole run,
and each is wrong for every file at once: an invalid `ingested_at`, a reserved
`okf_type`, and a missing inbox directory.
New code `inbox_slug_collision`: two dropped files reducing to one generated
name are BOTH refused. Persisting one would let iteration order decide the
winner, and overwriting would lose the other's content.
Phase 1 primitives are reused, never duplicated, so four helpers become
package-internal names (write_bytes, link_in_index, parse_frontmatter,
INDEX_NAME) and link_in_index learns to append to an empty index — Door A
always seeds its index with bundle_summary, but Door B has no summary to
invent and must not open with a blank line.
The slug inherited the whole filename stem with no length bound, so a long
dropped name produced a filename the filesystem cannot hold — surfacing as an
untyped OSError at the write, with an errno that differs per platform (63 on
macOS, 36 on Linux). The inbox contract promises a typed per-file outcome, so
the gate belongs in the pure function that forms the name.
Refusal, not truncation: truncating is lossy AND collision-prone — two long
names sharing a prefix would reduce to one filename and the second write would
silently claim the first file. Same posture as `inbox_slug_empty`: the library
never invents a filename the operator did not give it. The message carries the
actual size and the limit, since the fix is to rename the dropped file.
Limit verified empirically on APFS 2026-07-25: a 255-byte component writes, a
258-byte one raises errno 63. ext4 and NTFS cap at the same 255.
New MaterializationError code `inbox_slug_too_long`, registered in the errors
docstring and in the error-code conformance suite.
Second guard-independent step of Phase 2: the pure functions that turn a
dropped file's name and extracted text into an OKF concept file, with the
§7-analogous honesty marker. No runtime dependency, no guard call, no model
call — the persist gate is still steps 4–5.
- `inbox_slug(filename)` reduces a name to the Phase 1 id grammar
(`[a-z0-9][a-z0-9-]*`): drop the final extension, lowercase, collapse every
run of non-grammar characters to one `-`, strip the ends. A name that
reduces to nothing fails fast rather than inventing an `untitled` fallback
that would silently collide across unrelated files.
- `inbox_filename(slug)` namespaces it `inbox-{slug}.md`, disjoint from
`index.md`, Door A's `ingest-*`, and `promoted-verdict-*` for every slug the
grammar admits (asserted over a hostile-name table, not one example).
- `render_inbox_concept(...)` emits the six-key layer in fixed order — `type`,
`title`, `source_file`, `source_sha256`, `ingested_at`, `generated: true` —
with the digest taken over the ORIGINAL dropped bytes, never the extracted
text, so provenance stays re-verifiable against the operator's file. Body is
normalised to LF-only with exactly one trailing newline (dropped files
legitimately arrive with CRLF; Door A can validate instead because it renders
its own bodies).
Slugging normalises to NFC first. macOS (APFS) hands filenames over
decomposed, so `é` arrives as `e` + combining acute: without normalising, the
combining mark alone becomes a separator and the base letter survives
(`cafe`), where the composed form is one non-grammar character (`caf`) — one
visual filename, two slugs, depending on where the string came from. The test
pins both forms with explicit escapes so it cannot depend on the test file's
own encoding. No transliteration, deliberately: mapping non-ASCII to ASCII is
a semantic claim the slugger cannot make (Norwegian `møte`/meeting would
become `mote`/fashion). The readable name survives verbatim in `title`.
Fail-fast gates, all typed: reserved `verdict` layer refused at this door too
(the promotion gate stays the only path in); a title that is multi-line or
contains `[`/`]` refused, matching Door A's manifest rule, because it renders
verbatim into `- [title](target)`; a multi-line `source_file` refused because
line-oriented frontmatter would take the injected lines. Four new stable
MaterializationError codes, each with its entry in the test_error_codes.py
registry: `inbox_slug_empty`, `inbox_title_invalid`,
`inbox_source_file_invalid`, `okf_type_reserved`.
`ingested_at` validation moves to a shared `validate_ingested_at` in
materialize.py and is called by both doors — one rule in one place is what
keeps the determinism contract from drifting apart per door.
TDD: tests/test_inbox.py precedes the implementation. 325 tests green;
mypy --strict, ruff check/format, and the sanitize|quarantine|lexicon
boundary grep-gate clean; runtime dependencies still exactly none; the Phase 1
golden suite still passes byte-for-byte.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A2aKJxLejT9S8jYwoZ9fut
First, guard-independent step of Phase 2: a stdlib-only registry mapping a
dropped file's extension to its text extractor, with the fail-fast gates that
keep binary parsing out of core. `extract_text(filename, data)` dispatches
(case-insensitively) to:
- `md`/`txt` — utf-8-sig passthrough (BOM never leaks, baseline parity with
Door A's read_csv);
- `csv` — the Phase 1 `render_table` (renderer reused, not duplicated);
- `json` — verbatim inside `render_fenced_block`;
- `html`/`htm` — text via `html.parser`, `script`/`style` stripped, tags as
word boundaries (spec B3: adequate for v1, richer is out of scope).
`pdf`/`docx`/`xlsx` are `[extract]`-gated; until that extra ships a parser they
fail fast with a typed error naming the extra — never a silent skip, never a
bundled parser in core. New `ExtractionError(IngestError)` carries four stable
codes (`extractor_unknown`, `extractor_extra_missing`, `extractor_decode_error`,
`extractor_empty_csv`); a non-UTF-8 file is a typed corrupt-input failure, never
a leaked UnicodeDecodeError. `extract_text` returns text content only — LF
framing and concept frontmatter are the materializer's job (step 2).
No runtime dependency and no guard call yet (the guard pin and 0.4.0 land with
the persist gate in steps 4–5). TDD: test_extract.py + the four codes in the
test_error_codes.py registry precede the implementation; mypy --strict, ruff,
and the `sanitize|quarantine|lexicon` boundary grep-gate all clean; the Phase 1
golden suite still passes byte-for-byte.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HBbjgS5A55RVavoyjJC4FX
safe_resolve let an untyped ValueError escape when a target carried an
embedded NUL: os.path.realpath rejects the string before the boundary
check runs. That broke SourceError's documented contract ("always typed,
never a leaked OSError") and the v0.3.0 promise of stable machine-readable
codes on every failure.
A target that cannot be resolved at all now fails closed under the
existing `path_escape` code, with its own message. No new code is added —
`path_escape` already means "the boundary refused this target"; its
docstring is widened to say so. Consumers asserting on the code are
unaffected.
Found by consumer review: portfolio-optimiser's navigate_bundle leaks the
same ValueError from the same class of input, which makes an entire bundle
unreadable instead of skipping one link. That defect is theirs to fix; this
commit closes the library-side half.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WdVgowYC4LARgvNdNMiuvz
Consumer tests were binding message wording (pytest.raises(match=...)) to
distinguish sub-causes within one error type. Every raise site now carries
a documented, stable code attribute; the registry lives in the errors.py
class docstrings. Codes are stable API; message text is explicitly not.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
TDD step 5: materialize_bundle takes three explicit inputs (manifest,
bundle dir, regex-validated ingested_at — no wall-clock default) and
stages every extraction in memory before the first disk mutation. The
§3 collision gate refuses to overwrite any file without the ingest
stamp, before any mutation; replacement removes exactly the stamped
set. Frontmatter is the seven §5 keys in spec order with whitespace
collapse; output is LF-only raw bytes with one trailing newline; the
provenance stamp hashes the same bytes that are parsed. Fresh index.md
gets bundle_summary plus idempotent links in extraction order (§6
creation path; preserve/removal rules come next). The §8 network gate
refuses http fail-fast without the per-run opt-in; the http transport
itself lands in step 7.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QeqhJpYQyghASjiJo5EhGg
TDD step 3: read_csv executes a file-source extraction — boundary-checked
path resolution via commonpath on canonical paths (rejects .. traversal,
absolute paths, symlink escapes, and prefix-collision siblings), utf-8-sig
decoding, streaming max_rows cap as a typed error (never silent
truncation), ragged-row rejection, and verbatim cells (escaping stays in
the §5 renderers). New SourceError in the typed hierarchy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QeqhJpYQyghASjiJo5EhGg
TDD step 2: markdown-table rendering with the load-bearing escape order
(backslash, then pipe, then newlines to spaces with CRLF as one unit),
SQL value-to-text conversion (NULL empty, plain-decimal integers,
shortest round-trip floats, explicit bool/BLOB rejection — never silent
coercion), and the verbatim fenced block for http bodies. Table shape
and fenced-block form match the reference implementation byte for byte
ahead of the shared §11 golden fixtures.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QeqhJpYQyghASjiJo5EhGg
TDD step 1 of the phase 1 plan: typed error hierarchy rooted in
IngestError, manifest parsing with schema validation before any source
call — id grammar, polymorphic source types, extraction rules, verdict
reservation (okf_type and filename namespace), and the no-embedded-
credentials rule for base_url. Dev tooling (pytest/mypy/ruff) added as
a PEP 735 dependency group; runtime dependencies stay empty.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QeqhJpYQyghASjiJo5EhGg
