P18 parts D and E (order 20260914T105139Z), plus the two things measuring
them turned up.
DEL D -- five paid runs (gpt-4-1-mini, azure, PACE_SECONDS=2), same four
context sets, SAME parameters on all four (--max-rounds 3 --max-tokens
600000), plus one variance repeat of gate-nordvik. All four free
--live-dry-runs first: rc 0, and Grounding-offer numbers IDENTICAL to round 1
(435/272/982/3) -- the control that the Grounding structure did not change
what the gate measures.
MEASURED, round 1 -> round 2:
- runs that died on the token cap: 3 of 7 -> 0 of 5, and two sets now fit a
LOWER cap than round 1 had to give them;
- guessed read paths: 12 -> 0;
- wall time, the two sets whose parameters are directly comparable: n100
140.1 s -> 70 s, n500 73.5 s -> 69 s;
- 5 of 31 read_file calls opened documents BEYOND the default window, so the
window WAS widened -- the trace does not record which knob (finding 1);
- (a) grounded in a fasit concept: 0 of 26, UNCHANGED. That is the mission
gap, and DEL A did not close it.
The a4 falsification arm fails once in each round, on a different set. r761's
a4 is now rejected -- but NOT by B1: the model proposed "Kontraktsum" this
time, so the "appears nowhere" arm caught it, and B1's effect on that row is
proven offline, not live. NEW failure: tunnel-hauglia a4 VALIDATED on
"impulsventilator" (3/270 documents), and fv412 a1 on "bituminost barelag"
(4/1133). Both are ordinary Norwegian words from the standard's prose, not
cost codes. B1 cannot and should not fell them: this is an ANCHORING defect,
not a grounding one, and it is finding 2 with two named remedies and a
recommendation.
C2 isolated by re-judging round 1 with the new judge: kontrakt-sorasen goes
named=3 -> named=1, and the survivor is named_in_measure -- the model's own
words. Two of the three were the whole-base snippet artefact.
DEL E -- docs/2026-09-14-p18-stressrunde-2.md: round 1 against round 2, what
each fix bought (measured, never attributed), the B2 table, variance, and for
EACH remaining ugly finding a NAMED solution with an estimate.
THE MUTATION THAT FOUND A HOLE. B6 (revert run.py to compose ONE blob instead
of one document per concept file) left the WHOLE suite green: 1698 passed / 5
skipped. The composition arm drives _grounding_text with a Grounding it
builds ITSELF, so it cannot see what the RUN handed over -- and a blob has
exactly one boundary, so the floor can never be reached, the share can never
fire, and the measured defect is back intact. The rule is only as good as the
boundaries it is given.
Arm (h) is the gate that was missing: a crafted base with TWELVE concept
files all carrying the same token -- per document 12 of 17 and inert, as one
blob 1 of 1 and grounding -- with a control on a code only ONE file carries,
which must still validate. Measured RED against exactly that mutation. The
mutation was not dropped and the seam was not declared unwitnessed: it got a
witness.
Also: the debate's own bundle pointer (run.py _bundle_pointer) now explains
the window and the filter, alongside the tool description and the navigator
instruction updated in 9b47e5a -- a description that lies about the body IS
the model's instruction (the Fase 3 class). Golden transcript unaffected.
Mutations, all against the FULL suite in an isolated worktree, one at a time:
DEL A 7 of 7 red (control 1685/5), DEL B+C 9 of 10 red (control 1698/5), the
tenth being B6 above. Tables in the report s 9.
Verification: uv run pytest -q 1699 passed / 5 skipped (1670 on cfd9079; +29,
0 removed). ruff check + format clean, mypy clean (38 files). Golden
demo-transcript.stdout BYTE-UNCHANGED, shasum -a 1 of the CONTENT =
ea8c534773acdbe41ae68f2c55724d69aaf8be4f. No version bump, no push.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
P18 parts B and C (order 20260914T105139Z).
B1 -- stage 0b. P7 made it `item.code in grounding`: plain containment over
ONE concatenated string. P16 ran it against a delivered corpus and measured
what containment cannot tell apart: the falsification arm a4-indeksregulering
put 250 000 NOK on a single cost line coded R761 -- the knowledge base's OWN
NAME, carried by all 2 756 of its concept documents -- and the whole gate
said validated (stage 0 skipped, un-anchored run; checker approve).
The grounding is now carried as the DOCUMENTS it is made of (validator.
Grounding), not as a blob. A structure and not a second argument beside the
text: the boundaries and the text are one fact, and .text is derived, so the
gate and P8's report measure the same characters. run.py composes one
document per concept file where the base is already walked; generate.
_grounding_text folds each cost line in as a one-line document.
N and A are MEASURED, not chosen (14.09, four mounted vegnormal bases):
- every must_cite ref and mandate affected_code in the four context sets --
shortest real identifier is FOUR characters (12.1, 52.1), so N = 3 sits one
below the measurement and cannot refuse anything measured;
- document frequency of every code-shaped token per base -- 1 692 distinct
and NOT ONE reaches 5 %. Highest anywhere 6/446 (1.35 %), highest a fasit
names 3/446 (0.67 %), R761 2 756/2 756 (100 %). A = 0.05 therefore sits
3.7x above the highest real token and 20x below the defect.
Length is NOT what makes the defect inert (R761 is four characters); the
share is. And a share is not a measurement without a denominator big enough
to take one (ansikt 4): one of three is 33 %, so an ABSOLUTE floor of 10
documents gates it. Highest absolute count any real identifier reaches is 6,
and every fixture in the repo is far below 10 -- which is why every pre-P18
gate is UNTOUCHED by this rule rather than exempted from it. Grounding.of
(one document) can never reach the floor by construction.
The refusal NAMES the denominator ("appears in 2756 of the 2756 documents
this run was given"), because Step 5 feeds that reason verbatim into the next
attempt's prompt: a proposer told only "ungrounded" answers with another
token of the same kind.
B2 SPIKE (measured, NOT built) FELLED the order's own alternative: option (b)
"ground in what the run OPENED" was run over P16's 16 code rows -- R761
stands in every OPENED document too, so (b) would NOT have caught the defect,
while B1 makes it inert and still grounds the real process line 65
ASFALTDEKKER (29/2756 = 1.05 %). (b) is not a substitute for B1.
C1 -- --docs-dir is optional once --bundle-dir is given (P16 FUNN 2). On the
bundle path docs_dir is never read: retrieval, the chunk tool and the "no
citable content" check all live in the road branch. Bound ONCE from
--bundle-dir, which is byte-identically what the README already tells an
operator to type by hand. NOT the "--docs-dir omvei": no such path is opened
and the road branch still refuses without a real --docs-dir (own arm).
C2 -- the judge's snippet arm counts only under citation_scope == "narrowed",
as (a) already does (PM decision, P16 s 6.2). P16's reason for (b') being
clean -- snippets are bodies while ref/title live in frontmatter, 0 of 446
n100 bodies -- holds for "Krav 4.1.2-1" but NOT for R761, where a process
number like 12.1 stands in the bodies. Under a whole-base citation list that
mark was "cited" before any model call.
tests: test_inert_identifier_loadbearing.py (7 arms; known positive is P16's
OWN artefact replayed against the base that run was given, known negative is
26 of 26 fasit references still grounding), test_docs_dir_optional_
loadbearing.py (5 arms). test_stress_judge_loadbearing.py's snippet arm split
into narrowed/whole-base -- the pair is the discriminator, same snippet, same
mark, only the scope differs. The grounding tests migrate from str to
Grounding.of (the honest reading of a caller that declared no boundaries).
Verification: uv run pytest -q 1698 passed / 5 skipped (1685 after part A,
strict superset, 0 removed). ruff check + format clean, mypy clean (38
files). Golden demo-transcript.stdout BYTE-UNCHANGED, shasum -a 1 of the
CONTENT = ea8c534773acdbe41ae68f2c55724d69aaf8be4f. No version bump, no push.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
