llm-ingestion-okf/tests/test_pdf_heading_reserve.py
Kjell Tore Guttormsen b01492b7f5 feat(propose,cli): typography as a reserve, and the two of our own numbers it took to measure it
K3 round 5. Three questions, three answers, and two of them correct a figure
this repository published.

RETRIEVAL FIRST, because it could have reversed a default. hit@8 over the six
questions on BOTH K2 bundles -- Arm B at 629 concepts and the shipped default
at 492 -- is 5 of 6 with ranks 1,1,1,1,1,- on each, so 0 of 6 rows lost. The
order's rule reverses `--unit-fold` at >= 2 of 6; it does not fire, and the
default stands. The gold sets shrink (49->26, 20->17, 43->36, 11->18) while
every rank holds at 1, which is the fold merging concepts rather than removing
a document from the top.

TWO PUBLISHED NUMBERS CORRECTED, both ours. The S7 candidate ranks 96 of 629
and 159 of 492 were measured with the cost vocabulary passed to
`concept_scores` and NOT to `document_scores`, while `build_payload` passes it
to both; scored the way the shipped payload scores it, the same concept is 10
of 629 and 19 of 492. And round 4 attributed its non-delivery to the default
move -- measured here, it is not delivered on the Arm B bundle either, for a
different reason (knapsack eviction at 68 046 bytes of a 120 000 budget, versus
`below_k`). That column had been inherited from round 3's own build, never
re-measured.

`--pdf-headings font-reserve`, OFF, and the hypothesis behind it is falsified
by its own condition rather than by a score: position 7, the one position the
flag exists for, has THREE outline runs, so the reserve is silent there at
every minimum. It changes 0 of 12 cells on the reference and reaches 4 of 39
corpus documents, none of them rated. Built anyway because it was authorised
and because the condition is now measured rather than assumed. The predicate
lives in one place (`propose.heading_reserve_applies`) and the door receives it
as a callable, like `gate`: a plan indexes the exact string it was proposed
against, so a reserve firing on one side only would make every document it
touches a coded rejection.

The `xlsx` re-reading is confirmed on the artifact -- 11 `rule:sheet-section`
units plus 1 `rule:table-block` ingress -- but the number alone makes the cell
worse (distance 1 -> 2), because the criterion counts that ingress as a table
that should have been merged. A hit needs both halves ratified, and the
reference is the operator's.

`--sheet-section-rows` as a default: three cells better and none worse on the
twelve positions, but the K2 control moves -- row 1's gold document splits 1 ->
12 concepts and its best concept ranks 2 instead of 1. Condition not met,
default not moved.

Default build byte-identical before and after (`diff -r`, 30 md files).
Suite 1441 -> 1449; three of the eight were red first.

Report: docs/2026-09-08-k3-runde5-hitat8-og-skriftakse.md

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-09 00:25:51 +02:00

173 lines
7 KiB
Python

"""The font reader as a RESERVE: typography only where the outline gate is empty.
Round 4 measured the font reader two ways and shipped neither as a default. On
TOP of Arm D it took `pdf` from 2 of 8 to 0 of 8, because the documents it broke
already numbered their own chapters and the second heading source could only
ADD a title page and a contents listing. INSTEAD of Arm D it scored 1 of 8 but
landed one position exactly on its reference count, which is the observation
this flag comes from: use typography only where the document does not number
itself.
The mechanism is one condition, and these tests pin the condition rather than
the outcome: the reserve reads a rendering it may not have asked for only when
`outline_runs` admits nothing at the configured minimum. Everything else --
which arm ships, whether the reserve helps -- is a measurement, and the
measurement is in `docs/2026-09-08-k3-runde5-hitat8-og-skriftakse.md`.
**The reserve is OFF by default and the measurement says it must stay off**: on
the twelve-position reference it changes not one cell, because the only
positions where it fires are a PDF whose glyphs carry no ToUnicode mapping and
four office documents the PDF reader never touches.
"""
from __future__ import annotations
import warnings
from pathlib import Path
from llm_ingestion_okf import cli
from llm_ingestion_okf.extract import extract_text
from llm_ingestion_okf.propose import build_plan, heading_reserve_applies
FIXTURES = Path(__file__).parent / "fixtures"
NUMBERED_PDF = FIXTURES / "numbered-font-krav.pdf"
FONT_PDF = FIXTURES / "font-heading-krav.pdf"
def _extract(path: Path, **kwargs: object) -> str:
with warnings.catch_warnings():
warnings.simplefilter("ignore")
return extract_text(path.name, path.read_bytes(), **kwargs) # type: ignore[arg-type]
# --- the condition ----------------------------------------------------------
def test_the_reserve_is_silent_where_the_document_numbers_itself() -> None:
"""The whole point: a document with its own chapter run keeps its own run."""
assert heading_reserve_applies(_extract(NUMBERED_PDF), outline_run=3) is False
def test_the_reserve_fires_where_the_outline_gate_admits_nothing() -> None:
assert heading_reserve_applies(_extract(FONT_PDF), outline_run=3) is True
def test_a_run_shorter_than_the_minimum_is_not_a_run() -> None:
"""The condition reads the gate as CONFIGURED, not a fixed grammar.
The same three-chapter document is a run at 3 and not one at 4, so raising
the arm's minimum widens the reserve rather than leaving it behind.
"""
text = _extract(NUMBERED_PDF)
assert heading_reserve_applies(text, outline_run=4) is True
def test_with_arm_d_off_the_reserve_fires_everywhere() -> None:
"""`--outline-run 0` admits no run at all, so the reserve is unconditional.
Stated rather than left to the reader: that configuration IS round 4's
"font instead of Arm D", which measured 1 of 8. A caller combining the two
flags gets that, and gets it deliberately.
"""
assert heading_reserve_applies(_extract(NUMBERED_PDF), outline_run=0) is True
# --- the effect on a plan ---------------------------------------------------
def _entries(text: str) -> list[str]:
plan = build_plan(
Path("x.pdf"),
text,
b"",
okf_type="reference",
proposed_at="1970-01-01T00:00:00Z",
outline_run=3,
table_grid=True,
unit_fold=True,
)
return [str(entry["title"]) for entry in plan["entries"]]
def test_the_numbered_document_is_segmented_by_its_own_numbering() -> None:
"""The reserve buys nothing here, and that is the guarantee, not a defect.
Measured rather than assumed, and the measurement is stronger than the
guarantee: at the build default this fixture lands on the same three units
from BOTH renderings, because the outline arm's run wins the document and
the fold puts the font-inferred `Forord` inside the first unit rather than
beside it. So the reserve's silence here has two independent reasons, and
only one of them -- the condition -- is what this flag controls. The
condition itself is pinned above, on the extraction, where it is decided.
"""
assert _entries(_extract(NUMBERED_PDF)) == ["Generelle krav", "Merking", "Vedlegg"]
assert _entries(_extract(NUMBERED_PDF, pdf_headings=True)) == [
"Generelle krav",
"Merking",
"Vedlegg",
]
# --- the flag ---------------------------------------------------------------
def test_the_build_default_does_not_reach_the_reserve() -> None:
"""Every byte-pinned golden depends on this, so it is a test and not a note."""
assert cli.DEFAULT_PDF_HEADINGS is False
assert cli.DEFAULT_PDF_HEADINGS_RESERVE is False
def test_the_reserve_is_reachable_from_the_build_command(tmp_path: Path) -> None:
"""`--pdf-headings font-reserve` is a third value on the same axis.
One axis, three values, because `none`, `font` and `font-reserve` are three
answers to one question -- how a PDF's headings are recovered -- and a
separate boolean flag would let a caller ask for two of them at once.
"""
base = [
"build",
str(tmp_path),
"--bundle",
str(tmp_path / "b"),
"--bundle-id",
"x",
"--okf-version",
"0.2",
]
assert cli.parse_args([*base, "--pdf-headings", "font-reserve"]).pdf_headings == "font-reserve"
assert cli.parse_args([*base, "--pdf-headings", "font"]).pdf_headings == "font"
assert cli.parse_args(base).pdf_headings == "none"
def test_the_reserve_reaches_the_proposer_and_the_door_alike(tmp_path: Path) -> None:
"""Propose and apply must choose the SAME rendering, or nothing applies.
A plan indexes the exact string it was proposed against (`text_sha256`), so
a reserve that fired on one side and not the other would turn every
document it touches into a coded rejection. The build is the test: the
unnumbered PDF must come back segmented by its typography, and the numbered
one by its own chapters, in ONE run with ONE flag.
"""
inbox = tmp_path / "in"
inbox.mkdir()
(inbox / "numbered.pdf").write_bytes(NUMBERED_PDF.read_bytes())
(inbox / "unnumbered.pdf").write_bytes(FONT_PDF.read_bytes())
with warnings.catch_warnings():
warnings.simplefilter("ignore")
report = cli.build(
inbox,
tmp_path / "bundle",
bundle_id="reserve-fixture",
okf_version="0.2",
pdf_headings_reserve=True,
)
assert report.rejected == 0
names = sorted(path.name for path in (tmp_path / "bundle").rglob("*.md"))
# The unnumbered PDF: the default leaves it one flat concept (round 4 pins
# that), and the reserve gives it the unit its typography names.
assert "generelle-tekniske-krav.md" in names, names
# The numbered PDF: its own chapters, and no `Forord` the reserve added.
assert "generelle-krav.md" in names, names
assert "merking.md" in names, names
assert "vedlegg.md" in names, names
assert "forord.md" not in names, names