STRUCTURED_V1 derived `references` from every number-shaped token in a body.
A consumer measured 12 false references out of 12 on their corpus, ten of them
version numbers in titles read as document numbers. Measured again here before
changing anything, on two corpora, through `derive_document_structure` itself:
corpus A a consumer's normative bundles 2 561 docs -> 2 838 subjects
corpus B this repository's own docs/ 28 docs -> 559 subjects
The reported class reproduces, and two larger ones they did not report turn up:
702 of corpus A's subjects (24.7 %) are hexadecimal fragments of UUIDs read out
of `id="..."` attributes in embedded MathML, and corpus B lifts tokens out of
escape sequences in quoted source (`\n60` -> `N60`), licence identifiers
(`AGPL-3.0` -> `3.0`) and package pins.
A derived subject is now a bundle-local link target, or a number immediately
preceded by a cue from a closed set. Positive rather than a blacklist because
the data forces it: `V221` is a genuine document number in corpus A and
`V0.3.0` is a software version in corpus B, and they are the same token shape,
so only the words in front of them can tell them apart. The cue matches at a
word boundary (a Norwegian compound ending in `-klasse` otherwise satisfies the
cue `se`, which admitted 86 class designations) and the window is NFC-normalised
so a cue survives a decomposed filesystem. Fragment-only and brace-carrying link
targets go too: neither can name a concept, so neither is a pending pointer.
After: 2 838 -> 1 279 (A) and 559 -> 72 (B). Hand-classified against the
occurrence that actually passed the gate: 30 of 30 sampled genuine on A, 60 of
60 on B. Residual known falses: 9 of B's 72, all illustrative link targets in
prose about link syntax.
The prefix-resolution rule stays open, per the order's condition: of 2 589
documents, 2 562 carry a number and 0 of those are dotted, so unique-prefix
match has no data here to be defended against.
Cost stated rather than hidden: a corpus phrasing cross-references outside this
vocabulary derives nothing and must declare `references`. A missing reference is
visible to the reader; a false one is not.
Record: docs/plan/references-cue-rule.md
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ATyA5Lx53N1rKzwMnuMoem
Four decisions: Door C projects and never derives (D-C1), the projection is
key-agnostic so it is not about numbering (D-C2), an unrenderable value drops
the facet and never the concept (D-C3), and a faceted entry is refreshed
because it is the first entry that can go stale (D-C4).
Also records a measurement about the measuring. The mutation harness first
reported 11 of 11 killed; the number was entirely false, because it invoked
pytest with an unrecognised `--timeout` flag and every run exited non-zero. The
negative control caught it by being reported killed when it provably could not
be. Honest result is 9 of 11, with both survivors analysed as equivalent
mutants and the one real gap they exposed (profile threading pinned only by
coincidence) closed with a direct test.
Closes open item 6 of the Door B record, which asked exactly the question this
answers: whose claim does the frontmatter make.
A deep-research report arrived mid-session (.claude broadcast 20260826T223230Z)
on a near-identical problem shape -- numbered regulatory requirements with
cross-references. Two of its numbers bear on this work: index-selection
strategy contributed +38.0 points of accuracy, and edge inference gave NO
accuracy gain at 2.8x the cost.
The second is a negative finding worth inheriting rather than re-measuring. It
does not condemn what landed today: references and parent are EXTRACTED from
explicit tokens, and the one proposed relation is structural and costs a single
pass. It draws a line for later -- no semantic edge inference without measuring
that 2.8x against our own corpus first.
The report also states that no published source gives per-query token counts
for structured versus flat context, and none reports an indexed superseded-by
facet. That reframes today's 3.3x-6.4x cost dial: it is the tradeoff nobody has
published, which is a reason to measure it properly rather than to hide it.
Treated as a premise, not a result. An external number changes what is worth
trying next, not what this repo has proved.
Six decisions, each with the failure it prevents: confidence carried in the
format, certainty stated per field, supersession never inferred from one
document, the index as a projection rather than an append, facets as a new
profile, and Door B's ownership stamp left deliberately open.
Six findings reported as findings rather than as gaps, per the order: the cost
of closing the metadata gap is 3.3x-6.4x the flat index and whether that keeps
the OKF arm's cost advantage cannot be answered from this repo -- it needs
their bake-off re-run on their corpus; a reference to a parent-level number
does not resolve to a child, and the fix needs a real corpus before 'unique'
can be defended; OKF_V0_2's verified/status/stale_after are still expressible
and unwritten for the reason already recorded.
Every number carries its denominator, and the supersedes 0/55 carries a
known-positive control (50/55 on a corpus that has shared numbers) proving the
query can find what it reported missing.
