llm-ingestion-okf/docs/plan/structure-derivation.md

# Structure derivation: numbering, hierarchy, cross-references, supersession

Answers order `20260826T124341Z-181848459` (from `.claude`, 2026-08-26). The
premise it waited on — the Door B / Door A capability gap — was closed in
`dc9ea59`: the answer was title derivation, not a door change and not a profile
on Door B alone.

The operator's requirement, quoted rather than paraphrased:

> "Og akkurat dette MÅ llm-ingestion-okf kunne lage automatisk og så ideelt som
> mulig når dokumenter legges inn i en okf bundle innboks (som kan skje en
> eller flere ganger, altså additivt)"

where "akkurat dette" is the form `.claude` described: documents that are
numbered, hierarchical, cross-referencing, and versioned such that one edition
supersedes another.

## What was measured, and what the measurement does and does not say

ms-ai-architect ran a pre-registered bake-off on 2026-08-26 over 55 documents
and 40 gold questions. Every arm retrieved 40/40, so the measurement could not
separate the arms on retrieval quality. The only axis that separated them was
trap exposure: 18/20 for the OKF-index arm against 8/20 for a frontmatter
head-scan over the same content. The reason was measured independently on both
sides — the flat `DEFAULT` index carries title/date/status/supersedes 0 times
while the documents in the same bundle carry them 55/55/55/5.

That is evidence against `DEFAULT`'s **index**, not against OKF. The metadata is
in the bundle; the index throws it away.

## Decisions

### D-S1. Every derived fact is marked, in the format

`derived` is a flat flow list naming exactly the fields this library inferred.
A field present and absent from that list was stated by the producer. The
format carries it, so this was not a case where "the format cannot express
confidence" — a flat list of strings round-trips through this library's
line-oriented parser and through the guard's T2 grammar alike.

Rationale: an unmarked heuristic is worse than no heuristic, because the
consumer cannot know when to doubt it. With the marker, a consumer that trusts
nothing derived can still use everything declared.

### D-S2. Certainty is stated per field, not per document

| Field | Source | Certainty |
|---|---|---|
| `title` | producer's `title` key | declared |
| | leading `# ` heading | **derived** |
| | filename stem | **derived** |
| `number` | producer's `number` key | declared |
| | leading number token of filename, then of title | **derived** |
| `parent` | arithmetic on `number` (drop last dotted part) | inherits `number`'s |
| `references` | producer's `references` key | declared |
| | bundle-local link targets + number mentions in the body | **derived** |
| `supersedes` | producer's `supersedes` key | declared |
| | same number + ordered `version` | **derived** |
| `version`, `status`, `date` | producer only | declared |

The number grammar is alpha-prefixed (`N500`, `V720`, `R610.4`) or dotted
numeric (`4.2.1`). A **bare integer is deliberately not a number**: admitting
`12-things.md` or `2026-notes.md` would stamp a document number on most of a
second brain that never had one.

A parent our own grammar could never admit is not emitted at all (`4.2` would
drop to `4`). An unresolved list that never clears is one a consumer learns to
ignore.

### D-S3. Supersession is never inferred from one document

Whether A supersedes B is a fact about a pair. `derive_document_structure`
records only what the producer declared; the bundle-level resolver may propose
a chain from a shared number plus ordered versions, and every edge it proposes
is marked derived. A group whose members do not all carry a version proposes
**nothing** — that is exactly the case where a guess would be indistinguishable
from a fact, and supersession is the relation a consumer is most likely to act
on.

### D-S4. The index is a projection, recomputed from the whole bundle

This is the answer to the additive requirement, and it is a design decision
rather than an algorithm: `resolve_structure` is a pure function of the whole
document set, and Door B rewrites the managed region of the index from it every
round. Nothing is diffed, so there is no diffing algorithm to prove correct.

Three of the order's four additive requirements fall out by construction:

- **rebuild == incremental** — both are the same function of the same files;
- **idempotence** — the concept name is the identity, so re-dropping a document
  replaces its entry rather than doubling it;
- **a round-3 document can supersede a round-1 document** — the round-1 entry is
  rewritten, which an append-only index could never do.

The fourth is explicit: an unresolved pointer is rendered with a `?` suffix
rather than omitted. During build-up, pointing at something not dropped yet is
the normal state; the dangerous version of it is the one that leaves no trace.

### D-S5. Facets arrive as a new profile (`STRUCTURED_V1`), never on `DEFAULT`

`DEFAULT` states commons' ingest-spec §6 index layer. Changing its rendered
bytes from here would be this repo editing a contract it does not own (O2), and
it would churn every golden fixture Door B has written. `entry_pattern` **is**
`link_pattern` when a policy carries no facets, so `DEFAULT` and `STRICT_V1`
match the same lines and emit the same bytes; a test asserts Door B's output is
byte-identical with and without the new parameter.

### D-S6. Door B keeps writing the literal `generated: true`

Not routed through `profile.ownership.stamp(...)`, which would move `DEFAULT`'s
bytes to the O2 mapping and orphan every bundle this door has already written.
Which stamp Door B should write is a separate question from this order's, and
answering it here would have answered it silently. **Open.**

### D-S7. Structural edge inference only — never semantic

Added the same day, from external evidence that arrived mid-session
(`.claude` broadcast `20260826T223230Z`, deep-research report on the OKF track).
A multimodal regulatory-document system on a near-identical problem shape
(numbered requirements, cross-references, tables and figures spanning pages,
arXiv 2606.29399) reports two numbers that bear directly on this work:

- index-selection strategy contributed **+38.0 percentage points** of accuracy —
  which is what a faceted index is;
- **edge inference gave no accuracy gain at 2.8x the cost**, and structural
  edges (References, Specifies) dominated single-hop questions while semantic
  edges only surfaced on composite multi-hop ones.

The second is a negative finding worth inheriting rather than re-measuring. It
does not condemn what is built here: `references` and `parent` are EXTRACTED
from explicit tokens rather than inferred, and the one proposed relation —
supersession from a shared number plus ordered versions — is structural and
costs a single pass. It does draw a line: **do not extend this module to
semantic edge inference**, and if anyone proposes it, the cost side of that
2.8x must be measured against our own corpus before any of it is built.

Treated as a premise, not a result. An external number is somebody else's
measurement; it changes what is worth trying next, not what this repo has
proved.

### D-S8. The cost dial is an original contribution, not only a cost

The same report states that no published source gives per-query token counts
for structured versus flat versus full-verbatim context, and that no source
reports an indexed superseded-by facet at all. The 3.3x–6.4x dial measured
below is therefore a number the literature does not have, and the `supersedes`
facet is not a re-implementation of a known technique. That reframes the cost
finding: it is the tradeoff nobody has published, which is a reason to measure
it properly on a real corpus rather than a reason to hide it.

## Verification

All commands are runnable and were run. Suite: 615 → 695.

```
.venv/bin/python -m pytest -q                 # 695 passed
.venv/bin/ruff check . && .venv/bin/ruff format --check .
.venv/bin/mypy --strict src/
```

### Mutation testing, with proof the mutation was applied

Each run asserts the anchor was present and that exactly one `# MUTANT` marker
landed in the file before the suite was read; a run whose anchor was absent
reports "NO RESULT READ" rather than a green suite. The first attempt reverted
mutations with `git checkout --`, which discarded uncommitted work; the harness
now snapshots the file text instead.

| # | Mutation | Result |
|---|---|---|
| M1 | index entry drops its facet tail | 6 failed |
| M2 | `unresolved` always returns empty | 2 failed |
| M3 | title derivation not marked derived | 3 failed |
| M4 | resolver iterates unsorted | **survived — equivalent** |
| M5 | self-reference guard removed | 1 failed |
| M6 | unnamed facet key accepted | 1 failed |
| M7 | index append-only (never removes a managed line) | 3 failed |
| M8 | reprojection claims any managed line | 1 failed |
| M9 | derived block emitted unsorted | **survived — equivalent** |
| M9b | *both* sorts removed | 1 failed (after adding the ordering test) |
| M10 | unresolved marker dropped | 2 failed |
| M11 | facet validation skipped | 1 failed |
| M12 | derived title not threaded to the entry | 1 failed |

M4 and M9 are equivalent mutants, not test gaps: `edges` is sorted before it is
returned, and `documents` is populated from `sorted(glob(...))`, so the loop
order cannot reach the output. M9 nonetheless showed the ordering property was
pinned only as a side effect of a byte comparison, so a test that states it
directly was added; M9b confirms it kills the mutation that removes both sorts.

### Cost, with the denominator stated

Measured on a **synthetic** 55-document corpus shaped like the one the bake-off
described — numbered, hierarchical, cross-referencing, versioned. This is a
measurement of **this library's emission**, not a re-run of anyone's bake-off,
and it is synthetic precisely because a private consumer's documents do not
come into this repo.

Facet coverage in the index, 55 documents, same counting query both sides:

| Facet | `DEFAULT` | `STRUCTURED_V1` |
|---|---|---|
| number / parent / status / date / version / references / derived | 0/55 | 55/55 |
| supersedes | 0/55 | 0/55 (corpus has no shared numbers) |

The `supersedes` zero is a property of that corpus, not of the code. Known
positive control, same query and same counter, on a corpus where documents do
share a number: **50/55** — 5 chains of 11 versions give 10 edges each, which
is the arithmetic the code should produce.

Index size for the same 55 documents, by facet key set — the dial a consumer
sets:

| Facet keys | Index chars | vs flat |
|---|---|---|
| none (`DEFAULT`) | 1 540 | 1.00x |
| status, supersedes, derived | 5 122 | 3.33x |
| number, status, supersedes, derived | 6 002 | 3.90x |
| number, parent, status, date, supersedes, derived | 7 817 | 5.08x |
| all eight (`STRUCTURED_V1`) | 9 797 | 6.36x |

## What was not achieved, and what would be needed

Reported as the order asked — what was tried, what happened, what would have to
change — rather than as "not possible".

1. **The cost of closing the gap is 3.3x–6.4x the flat index.** The bake-off's
   OKF arm won on cost (6 031 characters against 21 879 for the head-scan it
   lost to on traps). Whether a faceted index keeps that advantage **cannot be
   answered from here**: it needs their bake-off re-run against
   `STRUCTURED_V1`, on their corpus, with their questions. We own the emission
   measurement; the arm comparison is theirs. The facet key set is the dial,
   and a three-key set is the cheapest configuration that still carries status
   and supersession.

2. **A reference to a parent-level number does not resolve to a child.** In the
   synthetic corpus, `N130` in the body of one document does not resolve to a
   document numbered `N130.2`; it is reported unresolved (`N130?`). This is
   honest but incomplete — the fix is a prefix-resolution rule
   (`subject` resolves to the unique document whose number starts with it), and
   it was left out because "unique" has to be defined against a real corpus
   before it can be defended. Deferred, not forgotten.

3. **`OKF_V0_2`'s `verified` / `status` / `stale_after` are still declared and
   never written.** The order asked whether the alignment doc's "expressible"
   reading is still right. It is, and for the reason already recorded: their
   v0.2 values are block lists of mappings, which this library's value model
   cannot round-trip until D1's structured reader lands. `STRUCTURED_V1`'s
   `status` is a **scalar** facet and is a different key in a different profile;
   it is not that field arriving early.

4. **Door B's ownership stamp is still the literal `true`** (D-S6). Open, and
   deliberately not decided inside this order.

5. **A faceted index cannot be judged against a directory**
   (`entries_match_directory`) in the same call, because `STRUCTURED_V1` does
   not set it. Not attempted; no consumer has asked.

6. ~~No profile is applied to Door C.~~ **Closed** — see the Door C section
   below. The objection stated here is the one the answer is built on: Door C
   PROJECTS the sender's declared facts and derives nothing.


---

# Door C: the index carries the sender's facts, and only theirs

Ordered by vegnormal-okf (`20260826T224500Z-873805419`) on a measurement, not a
preference. A pre-registered reasoning bake-off on 2026-08-27 ran the same
bundle through two index shapes:

| arm | index shape | hits |
| --- | --- | --- |
| B | `DEFAULT` profile's `index.md` | **0 of 8** |
| — | faceted index over the same frontmatter | **25 of 29** |

Same bundle, same concept files, same model. The `DEFAULT` arm did not answer
wrongly; it **abstained**, because the fact it needed was never in the file it
was given. Measured on their `index.md`: 30 974 characters describing 269
requirements, with **0 occurrences** of any of the eight facts (title, date,
status, supersedes, erstatter, req_number, seksjon, kravtype) that the concepts
in the same bundle carry.

`FacetPolicy` and `STRUCTURED_V1` already did this. They did not reach Door C.

## Decisions

### D-C1. Door C PROJECTS, and never DERIVES

Door B derives structure from documents this repo's own operator dropped. Door C
merges documents a **third party** wrote, verbatim, and the entire door is built
on refusing to put anything of ours inside their bytes.

Running the deriver here would not have broken the verbatim guarantee where
anyone would look for it. The concept file would still be byte-identical. It
would have broken it in the index entry *about* the concept — where our
inference is printed next to their name and reads as **their** claim. Forging an
attribution is not a smaller failure than editing a file; it is a quieter one.

So every facet in a Door C entry is a value the sender wrote in their own
frontmatter, copied. Nothing is inferred from their body, their filename, or
their neighbours in the bundle. Where the sender carries `derived`, **their**
list travels unchanged — so a reader can still see which of the sender's facts
the sender inferred, a distinction we would erase by adding inferences of our
own beside them. That is the whole ownership stamp at this door: the concept is
verbatim, and so is the index's account of what the concept claims.

`resolve_structure` is deliberately not called. It computes supersession across
a whole document set, and at this door that set is somebody else's.

### D-C2. The projection is key-agnostic, which is why it is not about numbering

The loop asks the policy which keys to carry and never what a key *means*. It
reads a value only to check the policy can render it.

Operator directive 2026-08-27, verbatim: *"alt vi lager rundt OKF publiseres på
Forgejo og skal virke for alt av innhold"* — so a design that only works for
numbered norms is wrong even when it scores well on N100/N200/N500. Nothing in
`_project_facets` can key off a numbering scheme, and the suite says so with an
unnumbered Norwegian meeting note (`status` + `date`, no number, no parent, no
version) and with a document declaring none of the policy's keys, which renders
the bare link rather than a separator with nothing after it.

The same property pays a second time: a consumer whose concepts are named by
UUID gets `title` into the index by **naming the key** in their own facet
policy, with no change here. Door C's index label is the concept path, so for
those senders the title travels as a facet or not at all.

### D-C3. An unrenderable value drops the FACET, never the CONCEPT — and is reported

The policy refuses a value carrying its own separator or joiner rather than
escaping it. At Door B that refuses the document, and rightly: the value is one
we derived and the operator can fix the source.

At Door C it must not. This door judges no shape and refuses no sender on form —
that is what the module docstring already commits to — so refusing a merge over
a semicolon in someone else's frontmatter is precisely the failure it names.
Dropping it silently is the other failure: the sender made a claim our index
does not show. So the facet is dropped, the concept merges verbatim, and the
drop is reported per concept and key in `ImportResult.unrendered_facets` — an
advisory over the merged set, like `unverified_references`, never a fifth
bucket.

### D-C4. A faceted entry is refreshed, because it is the first one that can go stale

Found by measurement during the work, not predicted. A flat entry carries a
label and a target, both stable, so it can never disagree with the file it
points at, and `link_in_index`'s early return on a present target is exactly
right. An entry carrying the concept's **facts** can disagree.

The path is reachable: the collision gate refuses an updated concept outright,
so the operator's only route to accept an update is to remove the merged file —
which the refusal message itself instructs — and import again. Measured after
doing so: the concept file said `gjeldende` while the index still said `utkast`.

An index that contradicts the bundle it indexes is worse than one that says
nothing, because the consumer reads the index and stops. A faceted entry for a
present target is now re-rendered in place, keyed on the policy's entry pattern
and the parsed target — never on a substring, so a curated line that merely
mentions the target survives verbatim, line ending included. With `facets=None`
none of this runs, so every unfaceted caller emits the bytes it always did.

This is the one addition beyond the order's three edits. It is here because it
is a divergence *this change introduced*, and shipping a measured file/index
contradiction would have been worse than the scope it costs.

## Verification

Suite **695 → 709**; `ruff check`, `ruff format --check`, `mypy --strict` clean.
`import_bundle`'s `profile` is keyword-only and defaults to `DEFAULT`, so the
171 branch bases built through this door emit the bytes they always did — pinned
by a test that feeds the DEFAULT path the same frontmatter that *would* produce
a facet tail under a faceted profile.

### Mutation testing, and the harness that lied first

11 mutations of the new code, plus a permanent **negative control** (a
behaviour-preserving local rename that must SURVIVE).

The control earned its place on its first run. The harness invoked
`pytest --timeout=120`; `pytest-timeout` is not installed, so every run exited
non-zero and every mutant was reported **killed — 11 of 11**. The number was
entirely false: zero mutations had actually been evaluated. The control is what
exposed it, by being reported killed when it provably could not be. A clean
sweep is worthless until the harness has been shown able to report a survivor.

The harness now asserts the suite passes on **unmutated** source before a single
"killed" is believed, and asserts per mutant that the anchor matched exactly
once and that the mutation landed on disk.

Honest result: **9 of 11 killed.** Both survivors are equivalent mutants, not
gaps:

- **M1** (`if not value` → `if value is None`) — `FacetPolicy.render` already
  filters falsy values (`profiles.py:423`), so the guard in `_project_facets` is
  redundant with the policy's own. The observable property is now pinned
  directly anyway: a key the sender declared empty produces no facet.
- **M8** (drop the `changed` guard before writing) — writes byte-identical
  content.

**M10** — reverting the `profile` threading in `import_filename` — survived the
first honest run, and that one *was* a gap: every profile shipped today reuses
`DEFAULT.paths`, so no existing test could tell a threaded profile from a
hard-coded `DEFAULT`. The property was held by coincidence. A profile owning its
own `import_prefix` now states it directly, and M10 is killed.

## What was not achieved

1. **`title` is not in `STRUCTURED_V1`'s facet set.** Widening it would move
   Door B's index bytes, so it stays a caller's key to name. For a sender whose
   concept paths are UUIDs this is the difference between a title in the index
   and none — flagged to vegnormal-okf rather than decided here.
2. **Door C's index is an append plus a per-target refresh, not a projection of
   the whole bundle** (contrast D-S4). A concept removed from the bundle by hand
   leaves its entry behind. Reprojection would mean enumerating the directory,
   which no profile has asked this door to do.
3. **Parent-prefix resolution is still deferred**, unchanged and for the
   unchanged reason: "unique prefix match" must be defined against a real
   corpus. `vegnormal-okf` has that corpus; the count is theirs to take.
4. **The before/after number is not ours to produce.** Whether the faceted index
   is worth its cost on a real corpus is answered by re-running the same arm
   with the same scoring, on their data.