Round 6 left two things behind: a measured coverage hole with the repair built
but not the number to move it, and a retrieval regression that kept two
spreadsheet rules off the default. Both are measured; four defaults move.
THE HOLE IS BIGGER THAN THE RULE BUILT FOR IT. Measured over the 39-document
corpus, the pre-move default left 207 435 characters -- 11.92 % -- in no
segment: 163 804 above the first entry, 26 041 between entries, 17 590 after
the last. `--first-span-from-zero` closes the first part entirely, 79 % of the
whole, leaving 43 631 (2.51 %) over 8 of 32 documents. It adds no boundary and
the K2 concept count is identical with and without it (425 = 425); hit@8 holds
[1,1,1,1,1,-] under both tie-breaks and the 12-position reference does not move
one cell.
THE RANKER WAS PUNISHING FINE-GRAINING FOR BEING FINE-GRAINED. Round 6 held
`--sheet-section-rows --keep-table-heading` back because on a bundle built with
them row 1 fell rank 1 -> 2, and ruled out the document prior because its rank
over 39 DOCUMENTS was 1 on both bundles. RRF reads a rank over CONCEPTS:
splitting the gold document 1 -> 12 puts its own twelve concepts in that
signal's whole top tie group, so the one leading the body signal takes position
11, contributing 1/71 where the undivided concept contributed 1/61. A signal
that scored them all EQUALLY still emitted twelve ranks, ordered by concept_id,
and the fusion read alphabetical order as a measurement.
AND THE REPAIR WAS ALREADY IN THE TREE, OFF FOR A COST THAT HAD BEEN REMOVED.
`--tie-shared-rank` shipped off 2026-09-08 because hit@8 fell 5 of 6 to 4 of 6.
Swept over 2 prior exponents x 3 bundles x 6 rows: that fall exists only at
DOCUMENT_PRIOR_EXPONENT 1.0. Round 6 moved the exponent to 0.5 for an unrelated
reason and correctly reported it moved no hit@8 row -- nobody measured the
pair. A flag's "off by measurement" is a measurement of a CONFIGURATION, not a
property of the flag.
Shipped, each with an explicit opt-out and both directions measured rather than
asserted: --no-first-span-from-zero --no-sheet-section-rows
--no-keep-table-heading reproduces the previous bundle byte for byte (diff -rq,
0 differences), and the no-flag build equals the explicit-flag build.
--no-tie-shared-rank reproduces the previous excerpt order -- the one change
here that alters a payload with no bundle changing.
K2 default: 436 concepts / 832 files, digest 8dff8a8e6c15d2f7..., hit@8 5 of 6
ranks [1,1,1,1,1,-]. The pin now holds its own known-negative on those bytes:
read with --no-tie-shared-rank the same bundle gives [2,1,1,1,1,-], so the
green assertion names its cause. Consumer cost is a re-run: 425/810 -> 436/832
on the reference corpus, 15/30 -> 26/52 on a five-document folder.
pptx and md measured end to end for the first time, on two hand-built
documents: md recovers 3 of 4 declared headings, pptx segments per slide only
where the converter recognised a title placeholder. A converter attribute leaks
into titles ({#slide-N}, {#sheet-1}) on 2 of 810 K2 files and 1 of 30 on a
five-document folder -- not fixed, because a filename is reduced from its title
and the fix renames concept ids a consumer has cited.
Suite 1486 (1478 before), ruff and mypy --strict clean over 21 files.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
One rule explains every remaining `pdf` miss on the twelve-position reference:
where a document DECLARES headings, Arm D's RECOVERED headings are the whole of
the excess, and every declared one is a unit the reference wants. `--outline-gate`
admits recovery only where the document declares none of its own, plus any one
recovered heading covering OUTLINE_SHARE (0.20) of the text. It is `fold_units`
clause 2's own principle moved from voting to admission, and it filters at
ADMISSION so the text a removed mark opened is carried by the mark above it --
the post-filter form scores identically on all twelve positions and loses that
text, which is why only one of them shipped.
`--outline-gate` and `--drop-wrapped-outline` become the package default, one
decision because neither carries the reference alone: `pdf` 2 of 8 -> 5 of 8
alone, 7 of 8 together; the sheet 5 of 12 -> 10 of 12; `docx` unchanged at 3 of
3. Each keeps an explicit opt-out. The bar the move had to clear was not the
reference: hit@8 on a K2 bundle built with it holds 5 of 6 at ranks 1,1,1,1,1,-,
no row losing rank 1. `--sheet-section-rows --keep-table-heading` reaches 11 of
12 and does NOT ship, because on a bundle built with it row 1 falls rank 1 -> 2.
Cost to a consumer is a re-run: 492 concepts / 944 files -> 425 / 810.
DOCUMENT_PRIOR_EXPONENT makes the document prior sublinear (total/n**0.5). A sum
measures size and a density is diluted by every unit carrying none of the
question, so a document split 1 -> 12 lost its prior by 12. Swept over five
values on 18 rows it is at least as good as the delivered density everywhere and
strictly better on three. Stated plainly: end to end it moved NOT ONE hit@8 row
on any of four bundles, so it did not solve the knot it was adopted for -- what
did is that the `pdf` gain never needed `--sheet-section-rows`.
`--first-span-from-zero` is off and repairs a measured loss found while chasing
one position's 940 characters: 32 of the 32 documents that get a plan leave the
text above their first concept in no segment -- 159 704 characters, 9.18 % of
the corpus, 45 841 from one document. It changes nothing on the reference. Off
because it moves the first span of essentially every bundle with no hit@8 number
behind it yet.
vegnormal-okf FUNN 2: SPEC section 8's own star row parsed as prose, so every
concept behind one was unreachable to the section 9.2 walk. `IndexPolicy.also_reads`
carries it for the SEGMENTED profiles, read-only, after the emitted pattern
misses -- the asymmetry `sources` already has. DEFAULT and STRICT_V1 untouched (O2).
vegnormal-okf FUNN 1: Door C's own outcome was refused at exit 1,
`bundle_id_missing`. `import_bundle` now takes `root_frontmatter_values`,
keyword-only, rendered before any disk mutation, written only when the index is
created -- Door B's mechanism and ordering.
Report: docs/2026-09-09-k3-runde6-outline-gaten-og-prioren.md.
Suite 1478 passed (1449 before), ruff and mypy clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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>
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>
The reading direction existed only for someone standing in a clone. `consume`,
`contract_check` and `skill` moved from `tools/` into the package and are
reachable as `okf consume`, `okf check` and `okf skill`; `okf project` is new
and does the whole thing in one command.
The red measurement: a consumption skill generated from a checkout carried 4
lines naming that checkout by absolute path, 2 of them the commands the skill
tells a reader to run. It now names `okf consume` and `okf check`, and a test
asserts this repository appears in it nowhere, with a known-positive so the
zero is a measurement rather than a search that could not find.
The `tools/` files stay as ALIASES, not re-exports: a re-export binds copies of
the names into a second module object, so a caller patching one patches a
binding the implementation never reads. Two tests that monkeypatch okf_consume
went green again only under the alias. Every published reproduction block runs
unchanged.
The template and docs/consumption-contract.md (the section 7.4 known-positive)
are force-included into the wheel from the file they are authored in, so both
travel with the commands that cannot run without them and there is still one
authored copy of each.
Step 0, before any of it: okf build's default gained Arm E (--table-grid),
with --no-table-grid as its opt-out. The default moved to D plus F earlier the
same day on Arm F's published 5 of 12 -- a figure measured with Arm E ON.
Without it the fold has no joined table to fold, and the shipped default scored
2 of 12 with docx 0 of 3. Measured on the operator's folder: 30 md / 15
concepts on the new default against 43 / 28 without Arm E.
Install measurement from a fresh uv tool install, empty folder, this repository
nowhere on PYTHONPATH: 5 documents in, 15 concepts out, 0 references to tools/
in the generated skill, okf check conformant (15 rules, 0 findings).
Deviation stated rather than hidden: the order asked that
tests/test_okf_consume.py be left untouched. Two assertions in it read a PATH,
which is the one thing this work changes. Both were moved and the second made
stronger -- it now asserts every command the README recipe names is a
subcommand the CLI registers, which a file existing on disk never proved.
Suite 1414 -> 1427. ruff clean, mypy --strict clean over 21 files.
Record: docs/2026-09-08-o5-okf-project.md
Co-Authored-By: Claude <claude-opus-5>
The default move landed in the previous commit; this is the number it costs a
consumer, measured rather than predicted. `okf build` with no flag over the
43-document reference corpus now yields 517 concepts in 969 files, content
digest `42c8646d...`, `merged + coded rejections = 43 = N`, 817.6 s. The
delivered 2026-09-03 bundle (629 concepts, 1108 files, `9cd74519...`) is
unchanged on disk and reproducing it needs `--outline-run 0 --no-unit-fold`
AND the stamp it was built with -- "the opt-out flags" alone is true and
incomplete, so the report says both.
README and CLAUDE.md now state which two rules are on, what each opt-out is,
and that the proposer's own defaults did not move. One CLAUDE.md sentence was
false as of the move ("a no-flag build stays byte-identical") and is corrected
in place rather than deleted: adding a flag still leaves the default
byte-identical; moving it does not, which is why it took an operator decision.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
K3 round 3, order 20260908T170037Z-3622420612-from-.claude, plus the PM
injection of 17:55Z carrying the operator's D3 answer.
TWO NEW RULES, both OFF by default.
`--sheet-section-rows` cuts an open table block at the rows that label its
sections: a RUN of at least three rows whose first cell is a bare numeric
label. It is the first rule here that opens a span INSIDE a table, and the
opposite direction from Arm E -- that arm decides how far a block extends,
this decides where it is cut inside. The spreadsheet is the one file type no
arm had ever moved: 1 concept -> 12 on the priced sheet, of which 11 are the
reference's 11 cost groups. The run guard buys exactly three candidates on the
corpus (the rows stating a computation basis), and each section carries its
own `source_sheet` + `source_rows`, verified on the artifact.
`--drop-wrapped-outline` declines an outline candidate whose line continues
onto the next one. Round 2 falsified the LENGTH axis on this case; this one is
not about size. Over the 12-document sample it separates 8 of 34 candidates --
the quoted regulation paragraphs and the risk-table rows -- and none of the 26
the operator kept. Position 4: 4 concepts -> 1, the reference. The cell is
`treffer` by count and lossy by bytes, and the report says so.
Whole corpus, one cached extraction: 1 of 39 documents changes under the first
rule (0 of 32 pdf, 0 of 5 docx), 5 of 39 under the second (all pdf). Arm B and
F2 are byte-identical against a frozen c3b645b, both halves by `diff -r`.
THE `okf build` DEFAULT MOVED, and this is the operator's decision executed,
not a side effect: no flag now means `--outline-run 3 --unit-fold`. Each arm
keeps an explicit opt-out (`--outline-run 0`, `--no-unit-fold`) and the pair
reproduces the pre-move bytes exactly. The proposer's own defaults do NOT
move: they are what the goldens and every published reproduction block are
pinned to, so the two layers now disagree on purpose and `cli.py` says where.
Two shipped expectations moved with it and both are stated in place. MEASURED
AND REPORTED BACK: the configuration made default scores 2 of 12 on the unit
worksheet, `docx` 0 of 3 -- Arm F's published 5 of 12 was measured with
`--table-grid` ON, and without it the fold has no joined table to fold.
Position 1 is declined for the second round on the axis the order named: 3 of
3 recovered chapters have body under them, so "heading-on-heading" separates 0
of 3.
The K2 ranking control round 2 could not measure: two bundles from frozen
c3b645b differing only in `--keep-table-heading`, 2 of 1108 files apart. The
priced concept ranks 96 of 629 in BOTH and is delivered at rank 10 in both, so
the flag buys 35 payload bytes and zero rank positions. Round 2's prediction
is falsified: the concept's NAME was already restored by the orphan-name
inheritance; the flag adds the heading LINE.
The stale S7 literal is restated where it stood (`c759a657...`, 171 614 B,
re-measured on this HEAD), so the next order cannot inherit the superseded one.
1397 -> 1414 tests. Report: docs/2026-09-08-k3-runde3-per-filtype.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The consumer's question about `vann- og frostsikring` in a subsea tunnel
delivered 0 of the 16 concepts covering it, best of them at fused rank 14.
Reproduced with the denominator, then decomposed per signal before anything
was built.
It is not a matcher miss. `normalise("vann- og frostsikring")` already returns
`('vann', 'frostsikring')` on HEAD, the prefix rule already bridges the
inflections, and the best covering concept already answers 7 of 7 question
tokens -- more than any delivered one. A tokeniser rule had nothing to widen.
It is the fusion, but not a weight. RRF ranks every concept in every signal,
including a signal that scored them all the same, and the declared
`(-score, concept_id)` tie-break then orders that group by id. On N500 the
document prior has TWO distinct values over 270 concepts, so the third signal
contributed alphabetical UUID order spread from 1/61 to 1/329 -- enough to put
a concept leading the body signal behind concepts sharing only `tunnel` and
`vann`.
`--tie-shared-rank` lets concepts a signal scores equally share that group's
first rank. The miss closes: best covering 14 -> 3, 2 of 16 delivered. OFF BY
DEFAULT, by the order's own rule: the three requirement lookups hold at rank 1
and the K2 digest holds, but hit@8 over the six published questions falls 5 of
6 to 4 of 6. Decomposed rather than guessed -- K2's prior is coarse (6 values
over 39 documents) rather than degenerate, and one gold sat early in its tie
group. That benefit was never a measurement, but it is a published row.
`--withheld-titles` gives each withheld entry the concept's title, so a reader
can see WHAT was withheld without reading the bundle. 11 lines of code; the
bytes are why it is off. It grows an N500 payload 37.9 % and takes the
629-concept K2 bundle's BOOKKEEPING to 122 704 B -- past the 120 000-byte limit
itself -- which would falsify the breaking point published in the tracked
`skills/okf-consume/SKILL.md` on the day it shipped.
Defaults measured, not asserted: six payload digests built from a frozen
`ff79cfa` (`git archive`, `__file__` checked) and from this tree with both
flags omitted are 6 of 6 identical, and `okf_skill.py` output is identical
apart from the paths each copy writes about itself. Contract checker exit 0 on
eight payloads, both values.
One known-positive did not reproduce and is reported rather than matched: the
order's S7 literal `2ae46f68`/169 573 B is stale by three excerpt-form commits;
HEAD measures `c759a657`/171 614 B.
Suite 1388 -> 1397. Report: docs/2026-09-08-rangeringsbom-sammensatte-ord.md
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
K3 round 2, per file type. Order 20260908T143513Z-6327528123-from-.claude, carrying two operator decisions taken beforehand: D1 the orphan-gate variant goes behind a flag, D2 the arms become reachable from `okf build`. No default moved. Report: docs/2026-09-08-k3-runde2-per-filtype.md.
THE REPRODUCTION HELD, all three numbers, before any edit: `okf build` on the five-document tender folder gives 31 markdown files with both PDFs flat and 5/5 merged; the tender PDF gives no boundary without a flag and 9 with `--outline-run 3` (reference 9); the price sheet gives 1 on HEAD. Both proposer runs had to go through `bash -c` -- zsh does not word-split an unquoted `$flags`, so a sweep hands `--outline-run 3` to argparse as one token and every row comes back exit 2.
D2 -- `cli.py:_propose_plans` called the proposer with no arm argument, so the build path ran Arm B while `tools/okf_propose_segments.py` could run D, E and F. It now passes `--outline-run`, `--table-grid`, `--unit-fold` and `--keep-table-heading` through unchanged. THE DEFAULT DOES NOT MOVE and that is measured, not asserted: same folder, no flags, before and after the change, digest 3af10770...8fbbe2 both times and `diff -rq` clean. The "before" bundle was built before the first edit, because the editable install reads src/ live. Red test on the PLANS and on titles rather than a count, with the same fixture and no flags as its control. Per-document table for B/D/E/F/F2 is in the README and the report; the tender PDF is 1 under the default and 9 under every arm above it, and the reference is 9.
D1 -- a sheet heading with a table opening under it has an empty body, so the orphan check drops it: the NAME survives (carried onto the table block), the LINE does not. `--keep-table-heading` lets the heading survive and absorb the table instead. Price sheet 1 -> 1 concepts, `source_offset` [34, 11048] -> [0, 11048], body now starting at the heading. ELEVEN IS NOT REACHABLE THIS WAY and the number says why: the sheet is one heading and one continuous pipe-table block, and the eleven cost groups are eleven ROWS inside it (lines 10-20 of 103). What is missing is a section-row rule inside a sheet -- the opposite of `--table-grid`. Corpus: the flag changes 2 of 39 documents, both `.xlsx`, under arms B, E and F alike; known-negative 0 of 32 `pdf` and 0 of 5 `docx`. With it off, Arm E over all 43 is byte-identical to session 109's tree (33 plans, 43 `.err`, 4 FAILED, diff exit 0, counts asserted first).
THE PDF REMAINDER, one at a time. Position 9: clause 1 read the list AFTER the orphan check, and a contents list without dot leaders is a run of bodiless headings, so all but the last entry were already gone and the run was one. The run is now measured on the pre-orphan list, predicate written once and read in both places. 11 -> 10. Position 7: the same clause required siblings, and a numbered report's contents list interleaves 1.1/1.1.1/2.1 -- its 34 entries are one block that the level condition cut into runs of 9, 1, 1, 1, 5, 2, 10, 2 and 3, so the short runs survived. The level condition is dropped; the run LENGTH, which is what the CONTENTS_RUN sweep bought, is unchanged. Measured outward: the relaxation changes 1 document of 39 and removes exactly the leftover line. 34 -> 33.
TWO REMAINDERS ARE DECLINED WITH NUMBERS RATHER THAN FIXED. Position 1: the three level-1 candidates are 3 of 3 `rule:outline`, same level, same grammar, and the operator keeps one of them by prose alone -- there is no property to read. Position 4: a title-length rule was measured on paper and falsified -- a real chapter is 56 characters and a real heading in a document the arms already score correct is 88, sitting between position 4's 86 and 91, so no threshold separates the classes. Position 0 stays an extraction failure.
ONE SHIPPED EXPECTATION MOVED and is stated rather than quietly updated: `Innhold 1` is now discarded with the contents list it heads. Its body is in no segment afterwards, which is a real cost on a fixture where that heading has one.
Nine new tests: five red before the implementation, four green by construction and named as such. Three mutations, three red, unmutated control green each time -- restoring the level condition, computing the run post-orphan, absorbing a table unconditionally. 1379 -> 1388 tests. ruff clean, mypy --strict clean on 17 files. K2 bundle untouched (1108 files, 9cd74519...). The K2 ranking control is NOT measured: no bundle was rebuilt with the flag, so the rank is a prediction.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Order 20260908T133512Z-139864689-from-.claude. First iteration of the
per-file-type directive (operator 2026-09-08 13:05Z), not the last. No
threshold is set: ratifying a bar is the operator's, and setting one inside the
work that produces the measurement would be fitting the bar to the number.
[skip-docs] covers README.md only, and it follows a precedent re-measured this
round rather than quoted: `grep -c` for outline-run, table-grid, Arm C, Arm D
and Arm E returns 0 in README.md and CHANGELOG.md, while --path-prefix, a real
interface change, has a CHANGELOG entry. The rule is "interface and behaviour
changes yes, arm flags no", and --unit-fold is an arm flag that defaults off.
CLAUDE.md IS updated, because its `okf build` bullet enumerates which arms are
off there and would otherwise become false.
FUNN 1, and step 1 asked for it: the reproduction broke. Arm E on HEAD is
byte-identical to the archive on 31 of 33 plans; the two that differ are 2 of 2
spreadsheets in the corpus. The cause is EXTRACTION, not segmentation --
56ae274 writes a workbook as pipe tables, and the sample's price sheet extracts
to 11 048 characters where the worksheet records 100 694, which is the figure
that commit's own message predicts. The consequence is a segmentation
regression against the reference: K3 position 3 went 3 concepts -> 1 under both
Arm D and Arm E, where the operator wants eleven. The mechanism is the orphan
check dropping the sheet heading once a table opens below it (propose.py:461),
already reported there as a ranking regression. Doors unchanged: 43 .err, 4
FAILED, extractable 39/43.
THE MATCH CRITERION WAS WRITTEN DOWN BEFORE ANY CELL WAS SCORED, and it stalls
at 7/12 on the literal calibration gate after three rounds, each revision
recorded. The five failures are not the criterion's: at every one it agrees
with the operator's own (a), (b) or free text and disagrees only with (c).
Column (c) is a RELATIVE judgement ("closest today"); the four K3 categories
are absolute. The only way to reach 12/12 is to define "correct" as "the
closest arm", which reads (c) back out of itself. The dominance gate, declared
in advance as the second reading, holds at 11/12.
ARM F is one rule with three clauses derived from the operator's three, not
twelve special cases, and it only MERGES or DISCARDS: a run of at least
CONTENTS_RUN same-level page-numbered headings is a contents list and goes; a
heading deeper than the unit level folds into its parent, extending the
parent's span; a table folds back into the shorter heading that introduces it,
keeping the HEADING's name. K3 first rater, n=12: 2 coarse / 5 fine / 0
duplicate / 5 correct -- best of four arms, ceiling was 4, two moved, nothing
regressed anywhere.
THE PAPER MEASUREMENT CAME FIRST AND FALSIFIED THE FIRST VERSION. Clause 2 was
letting rule:outline -- Arm D's RECOVERY of an integer numbering run -- vote on
the unit level, which took K3 positions 1, 7 and 9 to 3, 4 and 7 concepts
instead of 17, 34 and 11. A recovered numbering is a heuristic, not a level a
document declares, and the unit worksheet showed the operator ATX and dotted
headings only. Fixed with its own red test; 11 of 12 predictions correct after.
PER FILE TYPE, which is the directive: docx 3 of 3 (solved on this sample), pdf
2 of 8 (lags, unchanged by Arm F, and the remainder is decomposed per position
rather than left as one number), xlsx 0 of 1 (regressed, see FUNN 1). Outside
the corpus, n=1 each: pptx and odt byte-identical, rtf proposes nothing either
way, txt differs and exposes clause 2's fallback.
CONTENTS_RUN swept 1..5 and off. Distance prefers 1; three ships anyway,
because at 1 the body chapter "... i henhold til TEK 17" is deleted for ending
in a number, and no K3 cell differs between 1 and 4 -- the metric prefers a
value that provably deletes a chapter and cannot see the cost.
Whole corpus, all 43 through arm_run in ascending foreground chunks: 32 plans,
491 entries against Arm E's 679, 14 documents changed, 1 plan disappeared
entirely (three drawing-schedule numbers that clause 1 correctly reads as a
contents run) and that is reported rather than special-cased.
THE okf build MECHANISM IS REPRODUCED AND IT IS NOT DOOR B: cli.py calls the
proposer with no arm flag at all, so the shipped build path is Arm B. On a
tender PDF that means no boundary where Arm D finds nine and the reference says
nine. Largest per-file-type gap this round found; it is a default change and
therefore the operator's.
5 tests red first, 1373 -> 1379. ruff clean, mypy --strict clean on 17 files.
K2 consumer bundle unchanged: 1108 files, digest 9cd74519... with the flag off.
No bundle built, no version bump, no tag, no push.
Report: docs/2026-09-08-k3-arm-f-mot-enhetsarket.md
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Four questions, two bundles, one run each, in a scratch project outside this
repository with a generated skill per bundle. All four passed, and zero numbers
or identifiers appeared in any answer that were not in the delivered set or in
the payload's own identities (62, 45 and 35 unique numeric tokens checked).
The skill triggered WITHOUT being named in the prompt and selected the right one
of two installed skills from the question alone, so no special invocation syntax
is needed: the generated `description`, which carries the bundle id, the concept
count and the ref, is enough to route on.
One defect the runs found, and it was in the prose rather than the payload. The
citation guidance listed the four locator keys this library writes, so on the
270-concept third-party bundle the model reported "no page locator, the address
is at document level" while the excerpt in front of it carried
`source_element_id` - that bundle's own locator, correctly delivered by the
prefix rule. The guidance now tells the reader to cite whichever `source_*` keys
are present. On the re-run the same question returned the element id. Two runs
of one question, the second measuring a changed artefact and not retrying the
first.
One finding that is not a defect in this chain: the first attempt at a
known-negative was not one. The bundle covers water and frost protection on 17
of its 270 concepts and the ranker put none of them in the cut. The consumer
behaved exactly as the contract asks - refused, named its denominator, reported
its own zero as unmeasured because `withheld` entries carry no titles, and did
not go around the cut. Recorded as a retrieval miss rather than replaced, and
it is the same shape as the open fusion finding.
A correction to this session's own measurement is in the record too: a first
sweep used `grep -rhoE "^source_[a-z_]+:"`, whose character class excludes
digits, and so missed `source_sha256` on 270 of 270 concepts. A pattern that
cannot match what it is looking for returns a zero that reads like a fact.
README gains "Consume in Claude Code": folder to answer in three commands, every
one of them run in this session. A test holds that the recipe invokes only
scripts this repository ships, at the paths it names.
Suite 1373 (1339 at the session baseline), ruff clean, mypy src clean. No
version bump, no tag, no push.
Co-Authored-By: Claude <claude-opus-5>
Two changes, one theme: what a reader needs in order to cite is a property of
the PRODUCER, so neither the excerpt nor the skill may hard-code a list of the
producers someone thought of.
The pass-through rule is now the `source_` PREFIX, not the five keys this
library writes. Measured on the N500 bundle currently on disk: 269 of 274
concepts carry `source_element_id`, a locator that repository chose under this
chain's own rule ("the key says what it indexes") and that this library never
writes. The allowlist dropped it, and an excerpt that names a document without
naming the place in it is the defect this work exists to close. A prefix and
never a substring - `resource_owner` contains the literal and is not a locator,
and promoting it would be fabricated provenance produced by a matching bug. The
known-negative is tested: `bundle_id`, `type` and `ingested_at` do not travel.
Contract 8.5 states the rule as a prefix rather than a list.
K2 control, re-measured against the frozen tool at b6a8c8b, same question and
same k: the RANKING is untouched - same 8 ids in the same order, identical
`text_sha256`, identical `withheld`, denominators 629 = 621 + 8. The FIELD moved:
payload 108 877 -> 113 143 B (+3.92 %), spent 18 606 -> 22 210 (+450.5 B per
excerpt), excerpt members 9 -> 17, 99 changed lines. Known-positive follows the
contract document's bytes again: 12 049 -> 12 563 measured, 11 719 -> 12 227
raw, delta 330 -> 336.
`tools/okf_skill.py` instantiates the template for one bundle: id, ref, concept
count, the conditional-field table with a denominator per field (the `source_*`
rows DISCOVERED from the bundle, not listed), the whole-bundle cost by the gate's
own instrument, the share one measured answer spent, the concept count at which
the withheld bookkeeping alone reaches the limit, and the index-walk-against-
directory control - run once at generation time, never on the question path.
The form was chosen on a measurement that came out against the obvious gate:
the contract checker passes the UNFILLED template against a real payload, and
passes a skill built for a different bundle against this one's. It cannot tell
the two forms apart, so conformance could not decide it. What decides it is that
5's denominators, 6.4's conditional fields and 7.6's breaking point are
per-bundle numbers - a generic skill either leaves them as holes (the template's
own definition of unfinished) or states another corpus's numbers, which is worse
than a gap. Every gate the checker lacks is therefore a test here: no placeholder
survives, the skill names its own bundle's id and ref and not another's, its
commands are absolute and point at files that exist, and it refuses a directory
with no index (exit 1, `bundle_unreadable`), an index with no `bundle_id`
(`bundle_id_missing`), an empty bundle, and an occupied target without --force.
It lives in `tools/` for the reason `okf_consume.py` and `okf_contract_check.py`
state for themselves - outside `src/`, so no consumer's install surface changes -
and because a wheel-installed `okf skill` would emit a command pointing at
`tools/okf_consume.py`, which the wheel does not contain.
Suite 1372 (1347 before), ruff clean, mypy src clean.
Co-Authored-By: Claude <claude-opus-5>
The pre-pass delivered the right concept and the answer could not name it.
Measured by portfolio-optimiser 2026-09-08 over three paid arms: the gold
concept came back at rank 1 of 8 on 3 of 3 bundles, and the model answered
correctly on 1 of 3, because a delivered excerpt carried `concept_id`, body
text and nothing the document is known by. The previous session measured the
same gap from the other side: the provenance it had just written into every
concept did not reach the payload at all.
`excerpt_for` now carries `title` unconditionally, and `req_number`, the SPEC
5.1 address `sources` and each locator key (`source_pages`, `source_sheet`,
`source_rows`, `source_lines`, `source_offset`) when the concept has them. A key
the producer did not write stays absent: an empty value would assert that they
wrote an empty one, which is the contract's 6.4 failure.
`sources` is read in BOTH YAML forms, on a measurement rather than a taste. K2
writes the flow form on 629 of 629 concepts; the largest N-bundle writes the
block form on 270 of 270 and carries no locator key at all, so a flow-only
reader delivers that bundle with no address whatsoever. Reading the block form
is not a licence to write it - the emission rule is untouched, because the
line-oriented parser still cannot round-trip a block list. A `sources` value
this reader cannot decode is named (`sources_unreadable`), never dropped into
the same silence as an absent one.
Contract 8 gains the requirement and the checker gains its code
(`excerpt_unnamed`, 15 rules now, was 14): an excerpt a reader cannot name is
one an answer cannot cite, whatever its rank. `req_number`, `sources` and the
locators are SHOULD, not MUST - they are conditional on the producer, and a
bundle whose concepts carry no identifier cannot deliver one.
K2 controls, same question and same k, before against a frozen copy of the tool
at b6a8c8b: the RANKING does not move - the same 8 concept ids in the same
order, identical `text_sha256`, identical `withheld`, identical denominators
(629 = 621 + 8). The FIELD is what moved: payload 108 877 -> 111 744 B
(+2.63 %), budget spent 18 606 -> 20 907 (+287.6 B per excerpt), excerpt
members 9 -> 15, 83 changed lines. The contract document's own bytes moved with
8, so the budget instrument's known-positive moves with it: 10 349 -> 12 049
measured, 10 060 -> 11 719 raw, delta 289 -> 330.
New fixture `tests/fixtures/consume-provenance`: the two address forms and a
concept carrying neither address nor identifier. Purpose-built, because the two
real bundles are complementary and neither exercises both forms.
Suite 1347 (1339 before), ruff clean, mypy src clean.
Co-Authored-By: Claude <claude-opus-5>
A concept named its source file by basename and, when segmented, carried a
`source_offset` into the text THIS LIBRARY extracted. Following that pointer
needed the corpus directory, the extractor and its exact transitive version --
none of which the bundle carries. Hand-walked on a real K2 concept: six steps,
four of them requiring knowledge from outside the bundle, to learn that a
requirement sits on pages 12-13 of a 20-page document.
The address is spec's: `sources: [{ resource, title }]`, where `resource` is
the dropped file's inbox-relative path (SPEC v0.2 5.1:303-306 -- "an absolute
URL, a bundle-relative path, or a path into a `references/` subdirectory").
The locator is ours, and it has to be: 5.1 has no field for a place within a
resource, and the pinned guard (1.3.0) rejects every route to putting one
inside a `sources` entry -- a non-allowlisted key by name, a nested flow list
as "scalar leaves only", and quoting as an unsupported form. So the locator is
top-level keys shaped like `source_offset`, and a path carrying a flow
terminator is refused fail-fast rather than mangled.
The unit table is built AT EXTRACTION, where the extracted text and the
original's structure are known to agree: pdf -> `source_pages` from
pdfplumber's own page numbers (a page that yielded no text does not renumber
the ones after it), xlsx -> `source_sheet` + `source_rows`, everything else ->
`source_lines`. `source_offset` stays.
Two measurements changed the design before it shipped. A `paragraphs` key for
docx would name a number the document does not have: `<w:p>` counts of
108/27/65/176/57 against converted-markdown lines of 75/33/67/144/63, not one
pair agreeing -- so the key is `source_lines` and says what it indexes. And an
empty spreadsheet row renders exactly like a table separator: the content-based
rule ate 8 empty rows on the K2 price sheet and reported its last row as 92
against a workbook that says 100. The separator is now found by position, and
`tomrad.xlsx` keeps that red.
One profile moves. `provenance` is a policy object, `None` everywhere but
`SEGMENTED_OKF_V0_2`; the other five shipped profiles are byte-identical.
K2 rebuilt from a frozen src copy: 629 concepts, 1108 files, name set identical,
0 ids moved, 479 files byte-identical, 629 changed and 0 lines removed anywhere.
629/629 now carry an address and a locator. New ref
`sha256-tree:665563a2f74423fcbcc8e4f0b0954ee73b73985ac0418de4f6987bd162a1f7c8`;
`2f82fcfe...` is stale. The pre-pass payload does not grow by one byte
(209 092 B before and after, 18 changed lines: the ref and eight per-concept
digests) -- because an excerpt carries the body, not the frontmatter, which is
also why the consumer still cannot cite "file X page 12" from a payload alone.
Report: docs/2026-09-08-proveniens-k2.md. 1339 tests, ruff and mypy clean.
Co-Authored-By: Claude <claude-opus-5>
A question naming a requirement number now delivers that requirement at rank 1
on all three vegnormal bundles (was 96, 9, 35 of 446, 1 133, 270). Two
mechanisms, both measured, both default because no published figure moves.
The matcher: `tokens_match` compared four leading characters, so the unique
identifier `3.3.1-13` read as 135 of 446 common and the rarity weight ranked a
common adjective above the number naming the document. An identifier now
matches by equality alone; df falls to 1/1/1. Words keep the prefix rule, which
was measured for Norwegian compounds. Equality has no floor either, so a
three-character identifier stops matching nothing at all -- measured, `9.2`
reached 0 concepts while sitting verbatim in one title.
The lookup: a question carrying an identifier that sits verbatim in a concept's
title or id is answered by a partition over the fusion's output, not by a
fourth signal. The form was chosen by measurement -- a fourth RRF signal was
simulated first and put the gold at rank 26 / 15 / 19, none of them delivered,
because RRF consumes ranks only and one signal contributes at most 1/(RRF_K+1).
No frontmatter key list is declared: of 1 846 concepts carrying `req_number`,
1 846 also carry that identifier in the title.
The matcher alone is NOT a monotone win (N200 9 -> 26, because that gold's body
cross-references a neighbouring number that the prefix rule counted as a hit on
the question's). Only the partition delivers; the table is in the record.
Consumer corpus: every named control byte-identical against a frozen
`git archive` copy of 116d3e1 -- four payload digests, eight candidate ranks,
six hit@8 rows, both known-negatives. One document that was withheld at
position 621 of 621 is now delivered at rank 1, on a corpus with no
requirement-number grammar at all.
13 tests (12 red before the rules existed), 7 mutations, 7 red. 1 320 passed.
Consumption-side only; no bundle ref moves.
Record: docs/2026-09-08-eksakt-oppslag.md
Co-Authored-By: Claude <claude-opus-5>
O2b asked whether a requirement number can be made worth more than a common
word by weighting each hit with the token's rarity in the bundle, with no
hand-set constant and no declared token class. It can, on one of the three
bundles, and the two it cannot are decomposed rather than guessed.
The rule is log(N/df) over the concepts' own tokens, counted with the same
four-character prefix rule a hit is scored with. It enters the RANKING and
never the GATE: `lexical` stays a count, because `krav` weighs exactly 0 on
all three bundles and a weighted gate would drop every concept matching only
that word -- which is the gate 54a0bc2 falsified for other reasons. One df
table per bundle reaches every stage that scores the question against text,
including the document prior. One pass, 0.241 s over 1 133 concepts.
Measured on four corpora, before and after, with every published figure
reproduced first: gold fused rank 96 -> 103, 9 -> 8 (withheld -> DELIVERED at
rank 8) and 35 -> 35; K2's priced sheet candidate rank 10 -> 2 with the cost
vocabulary and 251 -> 78 without; Q-good unmoved at rank 1; hit@8 5 of 6 with
every rank identical; the S7 control payload byte-identical on the default
command.
DEFAULT OFF, decided by the number and not by taste: it does not win on all
four, because N100's gold loses seven rank positions. Off means the bytes that
were already published, and that is measured -- 8 of 8 payload digests
identical against a frozen copy of 56c1205 built with git archive.
Two limits, both someone else's mechanism and both named: MIN_SHARED_PREFIX=4
makes a unique identifier read as 135-of-446 common on N100, so the weight
correctly ranks a common adjective above the exact requirement number; and RRF
consumes RANKS only, so on N500 -- where the gold already leads the one signal
that can see the identifier, and the other two cannot see it at all -- no
weighting inside a signal can move anything.
Consumption-side only, so no rebuild: the K2 bundle ref 2f82fcfe... stands.
Report: docs/2026-09-08-sjeldenhetsvekt.md. 13 new tests, red first; 8
mutations, 8 red, two of them only after the survivors were read as code -- one
exposed a fixture that put the identifier where the real corpus does not, and
the corrected fixture is what found the RRF limit. Suite 1295 -> 1308.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two consumer-reported defects, one rebuild.
The pre-pass could not see a requirement number: `_TOKEN_SPLIT_RE` split
`10.2-2` into digit runs and `MIN_TOKEN_LENGTH` removed them, so a question
naming a requirement reached the ranker carrying only the word every concept
in a standards bundle carries. Measured on three real bundles (446, 1133 and
270 concepts), the named requirement was withheld `below_k` in three of three.
Numeric groups joined by `.` or `-` are now held together, dash variants fold
to the ASCII hyphen, and the noise floor is unchanged. The gold moves from
160 to 96, 143 to 9 and 100 to 35 -- a large move, and NOT a delivery: it is
still `below_k`, because `_overlap` is a count and an exact requirement number
is worth no more than a common verb. That weighting is a separate decision.
The rule was narrowed by a measurement: a version that joined alphanumeric
groups swallowed a document slug whole and cost a hit@8 row. An equality-only
variant was measured on all three bundles and falsified -- better on one,
worse on two.
The orphan gate destroyed a heading's name: a table opening directly below a
heading left that heading with an empty body, the orphan check dropped it, and
the surviving table block kept the mechanical `Tabell linje <n>`. A table that
orphans its heading now takes that heading's title and section number.
Conditioned on the drop, on adjacency, and carrying both members -- each of
the three measured or mutation-tested.
One K2 rebuild for both, from a frozen source tree: 629 concepts, `39 + 4 = 43
= N`, 2 of 629 ids moved and both moved BACK to the names the 2026-09-03
bundle carried, 1106 of 1108 files identical to it. New ref
sha256-tree:2f82fcfea91c3bd3f8ef7147f80cd613227d3ca7975c41d88810233f3f79ab4b
-- c26eed6a... is superseded. The regression the previous session measured is
closed: candidate rank 19 -> 10, and the delivering command is now
`--cost-vocabulary --k 12` inside the default budget at 58 907 o200k against
65 912 before. The specific question is unmoved at rank 1.
The tokeniser alone leaves the K2 control question byte-identical, measured
with the bundle held fixed and both published byte counts reproduced.
Report: docs/2026-09-08-kravnummer-tokenisering.md. 8 new tests, red first;
6 mutations, 6 red, one of them only after the survivor was read as code and
a missing fixture was added. Suite 1287 -> 1295.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two producer-side findings from the consumer's S7c acid test (ordre 20260908T063454Z-3648220855-from-.claude), both measured on K2 before and after, both with the corpus rebuilt from scratch.
FUNN 3 -- THE FORM. The converter's default markdown writer emits simple tables, which pad every cell out to the width of the widest cell in its column. Measured on the tender's price sheet: one 594-character prose cell produced a 67 244-character whitespace carpet with runs of up to 887 characters between a label and its amount, 19 integral amounts carrying a converter `.0`, and a header row naming one column. The bytes reached a live model in 2 of 11 prompts and 0 of 11 answers. The spreadsheet row now writes pipe tables with `--columns=1` (load-bearing: the pipe writer pads to a width computed from it, so at the default 72 a narrow table gains runs of up to 45). Same sheet after: 11 048 characters, longest run 2, one row per line, 0 artificial `.0`. Spreadsheet-only, and the scoping is pinned by three digests -- the same change moves the odt fixture 1366 -> 1105, so it can fail.
The `.0` rewrite is bounded twice: to a cell whose whole content is such a number, anchored between unescaped pipes, and skipped when the literal is in the workbook's shared string table -- the converter renders the number 92 and the TEXT "92.0" identically, so the output alone cannot tell them apart. Read with zipfile and xml.etree; no new dependency.
FUNN 2 -- THE LOG LINK. `link_log_in_root_index` (95eb271) is removed. Consumption contract SS 9.2 forbids a consumer from enumerating the bundle directory unless the profile says the index is derived, so the index tree is the entire map a consumer may use and everything it links is a document: their navigator returned 630 where our pre-pass counts 629, and a corpus run's own log was citable as content. The log is still written to the bundle root (SPEC section 9); `tools/okf_consume.py` keeps its exclusion for the bundles already built with the link.
K2 rebuilt twice. BEFORE reproduces the consumer's ref exactly (`sha256-tree:f14872a0...c8a92a`, 629 concepts) and their three consume figures to the token (57 289 / 62 149 / 58 401). AFTER: 629 concepts, `merged + coded rejections = 43 = N`, new ref `sha256-tree:c26eed6a...e3261f`, 627 of 629 concepts byte-identical, 1104 of 1108 files identical to the delivered bundle.
ONE REGRESSION, MEASURED AND NOT FIXED: on the mandate-shaped question with the vocabulary bridge the priced concept moves from candidate rank 10 to 19, so `--k 12` withholds it `below_k`; `--cost-vocabulary --k 20` delivers it at 65 912 o200k. The cause is measured rather than argued -- restoring only the concept's title on the new short body ranks it 10 again. The chain ends at the orphan check (`propose.py:461`), which drops the sheet heading once a table block opens two lines below it. That is the already-reported orphan gate, and changing it is a default-ON segmentation rule affecting every document type. The specific question is unaffected: rank 1 before and after. The priced excerpt's budget share falls from 56.5 % to 9.7 %.
11 new tests (RED first), 8 mutations, 8 red, with an unmutated control green each time. One mutation survived twice before the fixture could make it fire, and both survivals are written down. 1279 -> 1287 tests. mypy --strict clean on 28 files. ruff clean. Both proposer goldens byte-unchanged. One frozen literal moved with the fix and is reported rather than hidden.
Report: docs/2026-09-08-prisform-og-loggen-k2.md
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The prior measurement (docs/2026-09-08-blindsone-below-k-k2.md SS 3) found that
the budget, not the ranking, is the second lock on a mandate-shaped cost
question -- and that the same mechanism was a REGRESSION on the question that
works: raising `--k` to 16 evicted the gold concept, because the exact knapsack
maximises a SUM of fused scores and has no opinion about rank, so twenty small
excerpts out-value one that costs 56.5 % of the budget.
Measured here on the same 629-concept bundle, with the three known-positive
figures from `4c699fd` reproduced first:
- Corpus distribution, denominator 629: median excerpt 857 B, max 223 391 B,
3 concepts over the limit alone.
- Candidate rule (b), a corpus-derived budget, is FALSIFIED by two numbers: two
defensible derivations are 49x apart on the same corpus, the small one turns
the gold concept into `over_budget_alone` (13 refusals against 2), the large
one changes nothing at the default k. A budget is the consumer's constraint,
not a property of the corpus; `--limit` already belongs to the caller.
- Built instead, behind `--reserve-top-rank` (default OFF): the top-ranked
candidate gets its bytes before the pack runs, AFTER the `over_budget_alone`
pre-exclusion and never before, and the payload declares `budget.reserved`.
- It fixes the eviction: k=16 and k=24 deliver the gold concept at rank 1,
costing one and two excerpts, and 20.4 % / 27.3 % FEWER o200k tokens.
- It changes the delivered list in 2 of 24 measured combinations -- both of them
that eviction. In the other 22 the list, its order and `spent` are identical.
- It does NOT close the mandate-shaped blind spot: that concept ranks 10, not 1.
The one delivering command is `--cost-vocabulary --k 12 --limit 160000`
(62 149 tokens against 58 401), and that is a consumer's decision.
11 new tests (RED first), 7 mutations 7 red with an unmutated negative control
green before and after; two of the seven survived the first test set and the
tests were strengthened. Default payload byte-identical, both goldens unchanged.
Report: docs/2026-09-08-blindsone-laas2-budsjett-k2.md
Suite 1279 green, mypy --strict clean over 28 files, ruff clean.
Co-Authored-By: Claude <claude-opus-5>
The consumer report (portfolio-optimiser, S7 SS 2) found that a mandate-shaped
cost question withheld the corpus's one priced table under `below_k`. Measured
here, on a bundle proven byte-identical to a fresh HEAD rebuild:
- The mechanism is a VOCABULARY gap, not a `k` defect: two of three ranking
signals are exactly 0.0 and the concept is candidate 249 of 269.
- The k-sweep buys nothing: k in {8,12,16,24,32,64,128} all withhold it, at
+9.5 % tokens. It also found a regression -- for the question that WORKS,
k >= 16 EVICTS the gold concept, because one 67 838 B excerpt is 56.5 % of
the budget and the knapsack maximises a sum.
- Two proposed rules were falsified BEFORE any code: number/table density ranks
the priced table 178/165/46 of 269 (the form is unfilled, so it is
number-poor), and per-document spread puts its document 30th of 35.
Built instead, behind `--cost-vocabulary` (default OFF, DEFAULT byte-identical):
one declared cost/price/quantity vocabulary family that bridges a question and a
document naming money with different words. It moves the concept from candidate
rank 249 to 10 -- and does NOT close the blind spot: the budget still refuses
it, which is now a separately measured second lock.
Seven RED tests first; six mutations of the rule, six red (two survived the
first version of the tests and the tests were strengthened). Control: a question
with no cost term produces a byte-identical payload with the flag on, at every
k, on the real corpus. Known-positive: 164 987 B / 40 425 o200k tokens, equal to
the published pair.
Report: docs/2026-09-08-blindsone-below-k-k2.md
Suite 1268 green, mypy --strict clean over 28 files, both goldens unchanged.
Co-Authored-By: Claude <claude-opus-5>
Two producer-side defects from the S7 acid test (ordre 20260907T234741Z-9578626297-from-.claude), both reproduced on K2 before and after.
F1: `okf build --ingested-at` alone stamped only 11/629 concepts -- the unsegmented ones, which read the call's value directly. The 618 segmented concepts read `segment.ingested_at`, the plan's `proposed_at`, independently defaulted to `DEFAULT_STAMP`. `proposed_at` now falls back to `ingested_at` when omitted; neither flag passed still yields `DEFAULT_STAMP` for both.
F2: the consumption pre-pass's index walk counted a root-linked `log.md` (`corpus.link_log_in_root_index`, `95eb271`) as a concept, inflating a 629-concept K2 rebuild to 630 and letting the log rank and cut like real content. The link stays -- the contract is silent on `log.md` and `95eb271` already named it a LOCAL choice -- but the walk now treats `LOG_NAME` like the index itself: reachable, never a concept.
K2 rebuilt twice from the same corpus and diffed against the delivered `K2-bundle-20260903`: FOR (stashed fix, matching fbaac6d) reproduces po's numbers exactly -- 619/1108 files differ, 618 ingested_at-only, ref `sha256-tree:4ffd750c...`. ETTER (fix applied) leaves exactly 1 line differing (the deliberate log link, predating this fix) -- 0 files stamped 1970, 629/629 stamped 2026-09-03, ref `sha256-tree:f14872a0...`. The delivered bundle's ref is unchanged before and after (`sha256-tree:9a4e5561...a968b5`), since it carries no log link and the new branch never fires. Conservation identity holds both times: merged + coded rejections = 43 = N, 39/0/4.
1258 -> 1260 tests. mypy --strict clean on 28 files. ruff clean. Both goldens byte-unchanged.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Order 20260907T141419Z-2723316647-from-.claude: a 12-document, three-question
worksheet so the operator can answer the unit-of-knowledge question
docs/2026-09-07-k3-arm-e.md left open before any new arm is attempted. The
worksheet itself (concept splits per arm, extracted document structure,
blank checkboxes) is local-only and gitignored because it quotes corpus
content; this commit adds only the method record, which names no corpus
path or document title.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
STATE.md rewritten (local-only, gitignored) with the NESTE block, the board
line and the route scoring for the next session. Nothing else to stage: the
working tree is clean and every artifact this round produced is either already
committed or deliberately local.
Gate suite, run AFTER git add so nothing untracked was measured as green:
uv run pytest -q 1258 passed exit 0
uv run mypy --strict src/ tools/ 28 files, Success exit 0
uv run ruff check . All checks passed exit 0
uv run ruff format --check . 88 files formatted exit 0
Baseline at the start of the round was 1230 tests and 27 mypy files.
The next step is NOT another arm. It is the operator answering the unit
question, and the report says why: the verdict is a conditional in two clauses,
and the blind row carries one document's worth of information about this arm --
only 1 of Arm E's 3 movable documents falls in the fixed blind positions. A
threshold set on that measures the protocol, not the arm.
No push, as the order requires. Measured rather than assumed: 35 commits are
unpushed on origin/main..HEAD, 13 of them from this round, and 0 carry a
Claude-Session trailer (checked with git log --format=%B, not --oneline, which
prints subject lines only and could never have found one).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three rows rated in one round by one first-rater identity, then two blind
raters per arm as the order asked -- 36 first-rater verdicts, 36 blind ratings,
six k/6 figures. No threshold is set and none is implied.
First rater: Arm B 8/4/0/0, Arm D 4/5/0/3, Arm E 4/3/0/5. Against the direction
declared before the row was seen -- fewer `too fine` WITHOUT more `too coarse`
-- `too fine` goes 5 -> 3 and `too coarse` stays 4.
Blind raters: `too fine` is 1 of 6 in EVERY arm, Arm E included. The report
publishes both rows and states the verdict as a conditional, because they are
measurements of the same twelve documents and only one of them is favourable.
The disagreement is one position and it is legible. At position 10 both Arm E
blind raters kept `too fine` where the first rater moved to `correct`, and
their REASON changed rather than persisting: under Arm B and D they object that
thirteen table rows are severed from their header; under Arm E, where the table
is one concept with its header, they object that the table is severed from the
sentence that introduces it. Arm E fixed the first complaint and does not touch
the second.
And the denominator that governs how much that can say: the blind positions are
fixed at 0, 2, 4, 6, 8, 10, and only ONE of Arm E's three movable documents
falls in that set. The blind row carries one document's worth of information
about this arm. No threshold should be set on that, and the report says so.
The ceiling is measured this round rather than inferred. A committed instrument
extracted all 43 files and counted table ROWS in the text -- 3 of 39 documents,
57 rows -- because zero ENTRIES is not the same claim: the orphan check can
delete a table candidate before it becomes an entry, and one sample document
produces no plan at all so its zero would be an absence with no denominator.
The 57 reproduces the Arm D round's independent pre-work screen exactly, with
neither measurement derived from the other.
Controls are split by what they do on failure, which is this round's correction
to its own first plan: gating controls halt, predictions are reported whatever
they say. Only one prediction was allowed to halt -- a zero-diff run, which
would be a wiring bug wearing a null result's clothes, with the ceiling standing
ready as a plausible wrong explanation.
Nine limits are stated rather than left implicit, including three that weaken
the arm's own case: the character class was fitted to the same three documents
it is measured on; position 5 has the round's largest reduction (21 concepts to
6) and its label does not change; and pandoc's simple and multiline table forms
carry no pipes at all, so the ceiling is bounded by which form the converter
chose rather than by how many tables the corpus holds.
Pure ASCII (0 non-ASCII bytes), short document labels only (0 corpus filenames).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
36 blind ratings, six k/6 figures, as the order asked -- twice the Arm D
round's protocol, which used one rater per arm. Blindness is structural: six
separate subagent spawns, each given one arm-neutral file and told to read no
other. Rule names were stripped, because a rule:table-grid in the list would
have identified the arm.
The rows:
Arm B blind-1a 4/1/0/1 blind-1b 4/1/0/1 agreement 5/6, 5/6
Arm D blind-2a 2/1/0/3 blind-2b 3/1/0/2 agreement 6/6, 5/6
Arm E blind-3a 2/1/0/3 blind-3b 2/1/0/3 agreement 5/6, 5/6
Within-arm agreement exists for the first time, because there are two raters
per arm: B 6/6, D 5/6, E 6/6. It bounds self-consistency and nothing more --
every rater is an instance of the same model family.
The finding the first-rater row does not contain: `too fine` is 1 of 6 in
EVERY arm, Arm E included. Both Arm E raters kept `too fine` at position 10,
which the first rater moved to `correct`. Their reason changed rather than
disappearing. Under Arm B and D they object that thirteen table rows are
severed from their header; under Arm E, where the table is one concept with its
header, they object that the table is severed from the sentence that introduces
it. Arm E fixed the first complaint and does not touch the second.
And the denominator that governs how much any of this can say: the blind
positions are fixed at 0, 2, 4, 6, 8, 10, and only ONE of Arm E's three movable
documents falls in that set. Positions 5 and 11 were seen by no blind rater. So
"the blind row did not move" is partly a fact about where the protocol looks,
not only about the arm, and the report says so rather than letting the number
stand alone.
One disagreement is a limitation of the material rather than of a rater: at
position 6 the first rater calls Arm B `too coarse` because a chapter is
ABSORBED into a neighbour and does not exist as a concept, which a rater seeing
only the emitted concepts cannot detect.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
36 verdicts, one rater identity, one round, all three arms re-rated rather than
carried over. The rows:
Arm B 8 too coarse / 4 too fine / 0 duplicate / 0 correct
Arm D 4 / 5 / 0 / 3
Arm E 4 / 3 / 0 / 5
Against the direction criterion pre-registered in the brief before the row was
seen -- fewer `too fine` WITHOUT more `too coarse` -- `too fine` goes 5 -> 3
and `too coarse` stays 4. Both halves are published. No threshold is set and
none is implied.
Against the ceiling measured in step 9 (3 of 39 documents carry any table row,
so at most 3 of the 12 can move), the row moved 2 of the 3 it could. The third
is position 5, and it is the finding worth more than the two that moved: Arm E
takes it from 21 concepts to 6, which is the biggest single reduction in the
round, and the LABEL DOES NOT CHANGE. Each of its three references is still cut
into a 114-char title concept carrying a heading and no body, plus its table.
Joining the table rows removed most of the fragmentation and left the rest.
Two things this rating is not. It is not independent: this rater had read the
Arm D report's published rows, so Arm B reproducing 8/4/0/0 and Arm D
reproducing 4/5/0/3 is consistency, not confirmation. And positions 2, 6 and 8
are `correct` only under the reading that a top-level chapter is one unit of
knowledge -- the question Arm D surfaced, which is the operator's and is not
decided here.
Sources were verified before rating, not assumed: both archives reproduced
byte-for-byte in step 8.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
No code. The figures this round publishes, and the controls that let them be
read, are now in the local measurements file.
The ceiling stopped being an assumption. The committed instrument extracted all
43 files and counted table ROWS in the text itself, not entries in a plan:
3 of 39 extractable documents carry any table row, 57 rows in total. That 57
reproduces EXACTLY the independent pre-work screen the Arm D round published
(57 of 35 050 lines) without either measurement knowing about the other. Both
declared expectations -- 38 grid-rule lines across 3 documents -- hit their
value, so the shipped grammar is the grammar that was measured.
The Arm E run differs from the Arm D archive on exactly 3 of 33 plans, and
those three are the ones named before the rule was built: 04.json (position 5)
21 -> 6 entries, 02.json (position 10) 15 -> 3, 43.json (position 11) 2 -> 1.
The NN.json to position mapping is derived from each plan's own path prefix
against the committed draw, not from the glob index the file numbers come from.
The one outcome that would have been a wiring bug rather than a finding -- no
file differing at all -- did not occur.
Corpus totals: 709 -> 681 entries, 33 table-block candidates -> 5, and 5 of 681
entries carry rule:table-grid.
Path churn is 0, and the denominator is computed rather than declared: 676
spans are present in both arms and none of them received a different concept
path. The plan had declared 686, which was wrong twice over -- 28 entries are
removed, not 33, and a joined block's end moves so its pair matches nothing in
Arm D. Publishing a control with a wrong denominator is the failure this round
is supposed to be about, so the computed number is the one of record.
Nothing here gated. Step 8's controls halt; these are reported whatever they
say, because the predictions came from an exploratory replica and a replica may
not sit in judgement over the shipped rule.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
No code. This commit records that the Arm E round is allowed to be read.
Seven gating controls, all green, each one asking whether the SHIPPED rule is
the rule being measured:
- `src/` diff touches only `propose.py` and `cli.py`; 136 changed lines in
`propose.py`, most of them `#:` comments.
- The consumer bundle `K2-bundle-20260903` is unchanged against the literals
already published in the Arm D report: 1108 files and
9cd745194346cda0c70eab9c7136fa44506203bbe85bc17d7eff2766c6e9b4d1 under
LC_ALL=C. Pinned to a published value rather than compared before-and-after,
because a pair both taken this round cannot see a mutation that predates it.
- 0 network imports in the proposer and in both instruments.
- Arm B identity: the whole corpus re-run with no flags at
--proposed-at 2026-09-03T00:00:00Z is byte-identical to
K2-plans-baseline-20260907. diff -r exit 0.
- Arm D identity: the whole corpus re-run with --outline-run 3 at
--proposed-at 2026-09-07T00:00:00Z is byte-identical to
K2-plans-armd-20260907. diff -r exit 0.
- Door counts: 4 of 43 .err files record FAILED, so extractable is 39/43,
unchanged, and measured from the run artifacts rather than by building a
bundle nobody needs.
Three things about the identity controls are deliberate. They compare against
the ARCHIVES, not against a second run of the same binary, which is the only
way to detect a flag accidentally defaulted on. The json and .err counts are
asserted BEFORE the diff, so `diff -r` cannot exit 0 by comparing nothing. And
`_index.txt` is INCLUDED in the comparison rather than excluded: the Arm D
round's published block both omits its generation and excludes it, which would
let NN.json name a different document across two runs without anything saying
so.
The Arm B run is not bookkeeping. It is the corpus-level half of the promise
that --outline-run 0 with the flag absent is still Arm B, and it is what makes
the Arm B row rated later in this round rest on verified bytes rather than on
an archive nobody re-checked.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`tools/okf_table_measure.py` answers how far Arm E's join reaches and what it
costs, and it imports `find_candidates`, `_GRID_RULE`, `_TABLE_ROW` and both
rule names from the shipped module rather than carrying a copy.
`tests/test_table_measure.py` pins that with `is`, not `==`: `re.compile`
returns a distinct object for an equal pattern, so equality would be satisfied
by a pasted literal and only identity catches it. `draw_sample` is imported
from `okf_outline_measure` for the same reason -- one K3 draw in the
repository, not two that can disagree.
The column this round actually needs is the `|`-row count, for EVERY file
including the ones the door refuses. It is the ceiling: a document with no
table row cannot be moved by this arm. Asserting the ceiling from entry counts
instead would assume the orphan check kept every table candidate, which nobody
measured -- and one of the K3 sample documents produces no plan at all, so its
zero would be an absence with no denominator.
Constants split the way `okf_outline_measure.py` splits them. 38 grid-rule
lines across 3 documents is a DECLARED expectation measured before the rule was
written; a different value means the shipped grammar is not the measured one
and no figure below it may be read. The Arm D entry totals are REFERENCE
values, printed beside the measured ones and gating nothing.
A file the door refuses becomes a `Row` with its reason named rather than a
missing row, because `extractable / len(rows)` is the door count this round
reports as 39/43 and a silent drop would make it unmeasurable.
Tests first: 7 red (the module did not exist), then green. 1251 -> 1258.
[skip-docs] on the same measured precedent as the flag commit: `grep -c` for
"okf_outline_measure" and "okf_cid_measure" in README.md returns 0. A
measurement instrument is documented by its own docstring and by the round's
report, and it never enters the wheel.
ruff check: exit 0. ruff format --check: exit 0. mypy --strict src/ tools/:
28 files, Success. pytest -q: exit 0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The flag threads through `run` and `main` and takes no argument. Arm D's gate
is a run LENGTH where 0 means off; Arm E has no numeric parameter, so a boolean
is the honest shape and an integer would only manufacture a sweepable knob that
means nothing. `run` therefore adds no numeric validation, and the help says
why.
Both prose sites that enumerate the arms `okf build` does not expose are
updated: `src/llm_ingestion_okf/cli.py` and `CLAUDE.md`. The second was found
by review, not by grep of the first -- the same claim lives in two files and
only one of them is code.
The generalised attribution test earned itself in this commit. The first draft
of the Arm E help contained "byte-identical to Arm D -- Arm D rather than Arm
B", and argparse's rendering plus the test's ` --` chunk split meant the
attribution fell OUTSIDE the `table-grid` chunk. The test went red with the
truncated chunk printed, which is exactly the failure it exists to catch: a
whole-output grep would have been satisfied and the attribution would have been
unfindable in the option it belongs to. The clause is now parenthesised.
Arm C's marker check in `tests/test_cli_build.py` gains `rule:table-grid` and
is renamed to speak of all three arms, measured on the artifact rather than on
the flag: a flag `okf build` never passes is not evidence about what it emits.
[skip-docs] is the MEASURED precedent, not a convenience. `grep -c` for
"outline-run", "max-segment-chars", "Arm C" and "Arm D" returns 0 in both
README.md and CHANGELOG.md: an arm flag is documented in its constant's `#:`
comment, in `--help`, and in the round's measurement report, and it is off by
default so it makes no promise to a consumer. `--path-prefix`, which is a real
interface change, does have a CHANGELOG entry. The rule this follows is stated
at docs/2026-09-07-k3-arm-d.md: "interface and behaviour changes yes, arm flags
no." Arm E's report is docs/2026-09-07-k3-arm-e.md, later in this round.
Tests first: 3 red, then green (a fourth, the no-argument test, is honest in
its docstring that it is green before the flag exists too, because argparse
rejects an unknown option with the same code; it becomes evidence only once
the flag is real). 1248 -> 1251.
ruff check: exit 0. ruff format --check: exit 0. mypy --strict src/ tools/:
27 files, Success. pytest -q: exit 0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`build_plan` gains the keyword-only `table_grid` and threads it into
`find_candidates`, and the `derived` composition becomes an ordered build in
`_derived_names` rather than a two-branch conditional expression. Four
combinations exist now, and the order is itself a claim: the marker, the rule
that OPENED the span, Arm E's join, then Arm C's cut.
The test that earns its keep here is
`test_a_split_part_of_a_joined_block_names_all_three_rules_in_order`. Arm C
rebuilds every part as a fresh frozen `Candidate` from an explicit keyword
list, so a field not copied there is silently defaulted back -- the spans stay
joined, the plan looks right, every diff-based control passes, and only the
rule name is missing. That failure has no other detector.
The module docstring's claim that "each entry names exactly one" rule is
corrected rather than deleted: it is the ORIGIN that is single, not the list.
`tests/test_propose_segments.py`'s `len(rules) == 1` is deliberately left
alone -- it runs on the default fixture, where the original claim is still
true, and widening it would weaken a true assertion about Arm B.
Tests first: 5 red, then green. 1243 -> 1248.
ruff check: exit 0. ruff format --check: exit 0. mypy --strict src/ tools/:
27 files, Success. pytest -q: exit 0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Arm E, off by default. `find_candidates` gains a keyword-only `table_grid`
whose branch is not even evaluated when it is False, so the flag-off path is
byte-identical by construction rather than by test.
The defect it addresses is measured. The converter emits pandoc GRID tables,
whose rows are separated by `+---+---+` rule lines that `_TABLE_ROW` cannot
match, so `in_table` resets between every pair of rows and ONE table becomes
one concept per row group. On the K2 corpus that is 33 of Arm D's 709 entries,
on exactly 3 of the 33 documents that produce a plan -- and those three are the
K3 sample positions 5, 10 and 11, all three rated `too fine`.
Three points where this could have gone silently wrong, and what each cost:
- `Candidate` is frozen and `dataclasses.replace` is not imported, so the join
is recorded as a `set[int]` over `marked` and applied at the orphan-check
pass that already rebuilds every candidate. `subdivide` rebuilds them again
from an explicit keyword list, so `grid` is copied there too -- exactly the
trap `split` already has.
- `rule_pending` is cleared on the fall-through together with `in_table` and
`open_block`. A grid table ends with a bottom rule, which sets it; without
the clear, the NEXT table's first row would be recorded as a join although
nothing was joined. `test_two_tables_separated_by_a_blank_line_stay_two_concepts`
is built to catch precisely that: its second table has ONE row group, so it
cannot be joined, and the test asserts `grid is False` on it. A test that
checked only candidate counts would stay green through the defect.
- A rule line can never OPEN a block: it is reached only with `in_table` true.
So no surviving candidate's `start` moves, bodies only grow, and the orphan
check -- which is monotone in the line set -- cannot drop a candidate it
previously kept. Asserted as an offset, not a length.
Measured on the real corpus with this code, reproducing a prediction written
down before it was built: 21 -> 6, 15 -> 3, 2 -> 1 entries, and identical at
--outline-run 0 and 3, so Arm D and Arm E do not interact here.
Tests first: 8 red, then green. 1235 -> 1243.
ruff check: exit 0. ruff format --check: exit 0. mypy --strict src/ tools/:
27 files, Success. pytest -q: exit 0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two constants, no behaviour. `RULE_TABLE_GRID = "rule:table-grid"` joins
`RULE_NAMES`, and `_GRID_RULE` sits beside `_TABLE_ROW` as the grammar of a
pandoc grid-table rule line.
The rule is the author's, not upstream's, and the constant says so: `grep -c -i
"arm" docs/2026-09-02-k3-k4-k5-metode.md` is 0, so the definition was written
for order 20260907T075834Z-18584396-from-.claude. Its axis is a third one --
Arm C names SIZE, Arm D names what the DOCUMENT declared, and this names what
the CONVERTER emitted.
The character class `[-=:+]` is measured, not guessed. Across the three
grid-bearing documents of the K2 corpus, 38 of 38 lines whose stripped form
starts with `+` match this pattern, and those four characters are the complete
set occurring on them. The `:` is pandoc's column-alignment marker and is load
bearing: a first pass with `[-=+]` matched 37 of 38, and through that single
miss read one document as having two tables where it has one. The `\s*` on
both ends mirrors `_TABLE_ROW` because the loop iterates
`splitlines(keepends=True)` -- every line carries its `\n`, and an indented
rule line is a real shape.
Tests first: 2 red, then green. 1233 -> 1235.
ruff check: exit 0. ruff format --check: exit 0. mypy --strict src/: 17 files,
Success. pytest -q: exit 0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`test_the_help_attributes_arm_d_inside_its_own_option_chunk` ended with
`assert squeezed.count("not defined upstream") == 1`, with the comment "the
literal belongs to Arm D alone". That was true when it was written and it is a
guard worth keeping -- but it goes red the moment a SECOND arm is attributed
the same way, on an axis that has nothing to do with whether the new arm is
correct. Left as it was, the cheapest way past it during the Arm E round would
have been to weaken Arm E's attribution, which is the opposite of what the
guard is for.
The assertion is now derived from a declaration:
ARM_ATTRIBUTION = {"outline-run": True, "max-segment-chars": False}
Per chunk, an arm either carries the literal or must not; the whole-output
count is `sum(attributed)`. Arm C stays False on purpose -- its help says "not
defined in the K3 method file" instead, and the point of the check is that
each arm's attribution sits in its OWN chunk rather than anywhere in the file.
Semantically identical this commit (1 == 1). It is committed on its own, before
the flag it unblocks exists, so the generalisation is visible as a deliberate
change rather than as collateral inside a feature diff.
Tests first: not applicable -- this commit's product IS the test, and its
correctness is that it stays green against unmodified production code.
1233 -> 1233. ruff check: exit 0. ruff format --check: exit 0. pytest -q: exit 0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The K3 Arm E round (order 20260907T075834Z-18584396-from-.claude) measures a
rule that has not been written yet. This commit pins what the code does TODAY,
so the later claim "the default did not move" rests on a committed artifact
rather than on reading a diff.
Three characterization tests and one new golden fixture. Every count they
assert was measured against unmodified code before the test was written:
GRID_TABLE yields 3 candidates, TWO_TABLES_BLANK_SEPARATED yields 3,
TWO_TABLES_RULE_SEPARATED 2, PIPE_TABLE 1, PROSE_WITH_A_STRAY_RULE 1.
The second golden is not duplication. `propose-golden-default.json` is taken
over `OUTLINE_DOCUMENT`, which contains no `|` row and no `+` rule line, so no
table rule -- present or future -- can move its bytes. It is a guard that
cannot fire. `propose-golden-grid-default.json` is taken over a document that
has a pandoc grid table, and it is generated here from PRE-Arm-E code: a
fixture recorded after the change would pin post-change bytes and prove
nothing about the default.
Two fixture shapes are load bearing and say so in the band's comment. The
second table in TWO_TABLES_BLANK_SEPARATED has exactly ONE row group, so it
cannot be joined -- that is the only shape that catches a join whose pending
state is never cleared. And `propose()` is excluded from this band: it pops
`source` and discards every other kwarg, so an Arm E test written through it
would silently measure the default and pass.
Tests first: 1 red (the golden did not exist), then green. 1230 -> 1233.
ruff check: exit 0. ruff format --check: exit 0. pytest -q: exit 0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Order 20260907T074541Z-939882634-from-.claude asked for smoothed-IDF rarity
weighting so no_lexical_match gates on informative tokens, with both
known-negative controls (sveising/titan/vakuum; pingviner/Bouvetoeya) required
to reach delivered == 0 on the real K2 bundle (629 concepts, 39 documents).
Swept the threshold across the full observed IDF range (1.5-7.9) before
building anything, per Verifiseringsloven face 3: verify an input premise
against ground truth before acting on it. The frontier is empty. neg2 first
reaches 0 at T=4.6, and gold-hit reachability has ALREADY fallen to 0 of 6 at
that same T. neg1 does not reach 0 until T=6.4, by which point every positive
question has been unreachable for eighteen threshold steps. The two rarest
tokens in the entire gold set (titan idf 6.059, vakuum idf 6.347) belong to
the sveising question, so rarity weighting scores them as maximally
informative rather than as noise.
Two structural causes, not just a bad threshold pick: (1) df is measured over
the same four-character prefix matcher that collapses brannalarm/branntrapp/
branncelle/brannkonsept into one bucket, so a real topic word reads as the
LEAST rare term in its own question; (2) MIN_TOKEN_LENGTH=3 admits tokens
MIN_SHARED_PREFIX=4 can never match (hva/for/den are structurally df=0),
leaving some positive questions resting on a single surviving token.
No production code changed — the mechanism the order named is falsified
before implementation, so there is no red test to write against it. Returning
the order with this measurement as the reason.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
hit@8 = 5 of 6, every hit at rank 1, against a chance baseline of 1.35 of 6
over a denominator of 629 concepts per question. Wall time 0.51-0.56 s per
question; spent 17 970 - 74 838 bytes against a 120 000 limit.
Two things this measurement did NOT establish, both in the report:
- BOTH known-negative controls FAILED. A question the bundle has no answer to
still returns eight excerpts, because no natural Norwegian question is
lexically disjoint from a 629-concept corpus under a four-character
shared-prefix rule -- measured per token, the interrogative `hvor` reaches 40
concepts, `brukes` 83. So `no_lexical_match` works per concept and not as a
whole-question gate: an empty excerpt list is evidence of absence, a full one
is not evidence of presence. The fix is named (rarity weighting) and NOT
built, because this step's fence freezes the instrument before it is measured.
- The question texts were written during execution, after the ranker existed.
The plan recorded the gold documents' SIZE profile -- its per-row baselines
sum to 1.35 and the sizes used here reproduce that exactly, which is an
independent check that this is the set the plan profiled -- but it recorded no
question texts, and three of six gold documents could not be pinned uniquely
from the sizes. Not a blind evaluation, and the report says so.
The scorer is a tool, not a script in a document: `tools/okf_consume_measure.py`
takes the gold set as an INPUT because it is tracked in a public repository and
an answer key names a consumer's documents. hit_rank, both chance baselines and
the document-size census are unit-tested; the corpus run is a measurement.
Public-file rule, checked with a pattern DERIVED from the corpus's own 39
document names rather than hand-picked, and shown able to find first (67 hits on
the bundle's own index): zero corpus document names in any tracked file in this
repository. One leak was found and removed on the way -- a corpus concept name
in a code comment and a hardcoded corpus path in a test.
Suite run after git add: 1230 passed, mypy --strict clean on 27 files,
ruff clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Step 12's README section is brought forward to here because the docs gate is
right: a feat commit that ships a new command needs the command documented.
CLAUDE.md's Commands section gains the pre-pass beside `okf build`. Nothing
else moves.
Contract check against a real payload from the 629-concept bundle:
$ .venv/bin/python tools/okf_consume.py <K2-bundle> \
--question 'Hvordan skal prisene fylles ut?' --out /tmp/k2.json
$ .venv/bin/python tools/okf_contract_check.py \
--skill skills/okf-consume/SKILL.md --payload /tmp/k2.json
conformant: 14 rules over 8 excerpts and 621 withheld entries, 0 findings
exit=0
And the two negative controls, because a green checker proves little on its
own -- measured, it returns 0 findings on an empty payload paired with the
unfilled template:
broken denominator identity -> NOT conformant, 2 findings, exit=1
missing payload file -> exit=2
Placeholder scan, known-positive first: the DOTALL scan reports 20 occurrences
on the template and 0 on this copy. The shipped example payload is generated
from the in-repo golden bundle, not from the corpus, and a test regenerates it
byte for byte. No K2 concept path or document title reaches any tracked file
here, checked with a pattern shown able to find against the bundle's own index.
Suite run after git add: 1224 passed, mypy --strict clean on 26 files,
ruff clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two corrections the plan did not carry, both found by running the instrument:
- A question matching nothing still returned the top eight by tie-break -- a
confident guess wearing a denominator. The closed rule set gains
no_lexical_match, so the known-negative returns a measured empty set with
every considered concept named in withheld.
- The stage-one document prior summed overlap across a document's units, so it
measured document SIZE. Measured on K2 for the price question: the
competition document sums to 6.0 over 79 concepts (0.076 each), the price
document to 2.0 over 1. The prior is now a density; the price-form gold moves
from outside the top eight to rank 1. Found with that answer visible, which
the measurement document states beside the number.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`docs/2026-09-04-k2-pptx-odt-rtf.md` measured the corpus denominator for
these three rows at zero, so K2 for them was unmeasured rather than
passing. This measures them on the synthetic set committed in `a7b050b`,
whose hand count landed BEFORE any of these numbers existed.
Coverage 59/59 and pairing 20/20 for all three, standalone and product
alike, where "product" is scored off the persisted concept in the bundle
rather than an intermediate. The instrument's own denominator, derived
from each file's XML or stream, agrees with the hand count on all three:
59 strings, 20 pairable. Door B over the fixture directory persists 3/3
with nothing quarantined, rejected, failed or skipped, so K1b holds at
3 = N. There is no sub-100 % cell to explain, and that is a result about
three documents we wrote, not about the formats.
THE NEGATIVE CONTROL IS WHAT MAKES THOSE FIGURES MEAN ANYTHING. Removing
one label and its value from each converted text moves every row to
57/59 and 19/20 -- exactly the two strings and the one pair taken away.
`tests/test_fidelity.py` carries that control per reader, so it cannot
rot into a claim.
The order names `okf_fidelity` as the instrument and the instrument
refused all three types (`no source-string reader for '.pptx'`), so step
2 could not run without readers for them. Added here, stated in the
report as sitting outside the order's enumerated fence: this is the
measuring instrument, not extraction code. `extract.py` is untouched and
every `_EVIDENCE` row still reads `unmeasured` -- three synthetic
documents in one house style are exercise, not coverage, and the suite
asserts the rows are not promoted.
Two converter findings, reported and not fixed, both exit code 0 with no
warning. RTF cell paragraphs without `\pard\intbl` are read as each row
nested inside the previous one, 2076 characters where 117 were expected.
And the `\uN?` unicode escape -- THE FORM WORD EMITS -- loses the
character after it: `A\u248?BC` reads back as `AoC`. Any Word-produced
RTF carrying Norwegian letters loses one character after each of them,
silently. That is now a named limit of the `rtf` row rather than an
unknown.
Pandoc 3.9 and the host's 3.10.2 produce byte-identical output for all
three fixtures, so the numbers are not sitting on a version-fragile edge.
`~/corpora/` was neither read nor written; N there is still 43.
Suite 1141 passed, `ruff` clean, `mypy --strict src/ tools/` clean over
25 files, all after `git add`. Regenerating the fixtures leaves
`git diff --quiet` clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`docs/2026-09-04-k2-pptx-odt-rtf.md` measured the corpus denominator for
these three office rows and found it ZERO: `K2/trinn1` holds 43 files and
not one of them is a `pptx`, an `odt` or an `rtf`. So `extract._EVIDENCE`
calls those rows `unmeasured` in the strongest sense available -- they
work by construction and had never met a document at all.
This is the smallest thing that changes that without inventing a corpus.
One authored document -- a title, an intro, a 20-row label/value table, a
caption and a 4x4 grid -- laid out three times in three containers, so
the container and its reader are the only variable between the three
measurements. `k2-office-fasit.json` carries the hand count taken from
the AUTHORED content rather than from any converter's output: 56 cells,
20 pairs, 59 distinct strings, shared by all three. It is committed here,
before the measurement runs, because a fasit written afterwards is a
description of a result rather than a denominator for it.
No converter wrote any of these files. `make_k2_office.py` lays every
part by hand, for the reason `make_fixtures.py` already states and this
set inherits: a file written by the converter and then read by the
converter proves only that the converter agrees with itself, and stays
green through any conversion defect that is symmetric. The commissioning
order offered pandoc as one generator option; the committed fixture
policy forbids it and the policy wins.
Two converter behaviours were measured while laying the RTF out, both of
them structurally plausible input read silently wrong, exit code 0 and no
warning. Without `\pard\intbl` on cell paragraphs, consecutive
`\trowd...\row` rows come back as each row NESTED inside the previous
one: five label/value rows read as five levels of nested table, 2076
characters where 117 were expected. And the `\uN?` unicode escape -- the
form Word emits -- loses the character after it: `A\u248?BC` reads back
as `AoC` with the `B` gone, `A\u248?xBC` reads back as `AoBC`. The
fixture writes `\uN ?` with an explicit space, which round-trips. Neither
is worked around anywhere in `src/`.
The generator and the fasit live one level above `k2-office/` and that is
not tidiness: Door B walks its drop directory recursively, so anything
parked beside the three documents would enter the run and N would stop
being 3.
Three synthetic documents in one house style are not a corpus. The rows
stay `unmeasured` and the suite asserts that they do.
Suite 1132 passed (1127 + 5), `ruff check` and `ruff format --check`
clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
