open/llm-ingestion-okf
1
0
Fork 0
Code Issues Releases 1 Activity

test(consume): hit@8 over six questions against a random-ranker baseline

Browse source 
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>
This commit is contained in:
Kjell Tore Guttormsen 2026-09-07 09:37:14 +02:00
commit d7751c0b9a
6 changed files with 598 additions and 39 deletions
Download patch file Download diff file Expand all files Collapse all files

185
docs/2026-09-07-okf-konsumskill-maaling.md Normal file
View file

@ -0,0 +1,185 @@
# The consumption pre-pass, measured — hit@8 on a 629-concept bundle
**Date:** 2026-09-07 · **Instrument:** `tools/okf_consume.py` at `51735fa`,
scored by `tools/okf_consume_measure.py` · **Corpus:** a 629-concept, 39-document
OKF v0.2 segmented bundle built by this library's own doors on 2026-09-03.
This is the first time `docs/consumption-contract.md` has been run against a
bundle. The contract has existed since S5 with no instrument, and a rule with no
instrument is a rule nobody can be shown to have broken. What follows is that
instrument's first numbers, with the two things a first number needs: the
denominator it was measured over, and the chance baseline it should be read
against.
**Read the headline with its granularity attached.** The order this work came
from asked for "≥ 5 spørsmål der **fasit-konseptet** er kjent" — gold at
**concept** granularity. The operator answered the open question on it
(2026-09-07 04:33Z) with option (a): gold is the top-level **document** the
question names, taken from the corpus's own document titles and never from
anything the ranker computes. **Five of the six rows below are therefore
DOCUMENT-granularity, which is a weaker claim than the order asked for.** One row
— the price form — has a gold document holding exactly one concept, so document
and concept granularity coincide there, and it is the only concept-level result
in this report.
## The result
| figure | value |
|---|---|
| questions | 6 positive + 2 known-negative controls |
| **hit@8** | **5 of 6**, every hit at **rank 1** |
| expected by chance | **1.35 of 6** |
| denominator | **629 concepts** per question, every one accounted for |
| `k` | 8 |
| known-negative controls | **2 of 2 FAILED** — see below |
| `spent` per question | 17 970 – 74 838 bytes (median 20 182), limit 120 000 |
| whole payload per question | 109 951 – 165 109 bytes |
| wall time per question | 0.51 – 0.56 s (629 concepts, 1.95 MB of concept text) |
Five hits at rank 1 against an expectation of 1.35 is a real signal and a small
one: **n = 6**. Nothing here supports a claim about the ranker's behaviour on a
seventh question, and this report makes none.
### Per-row, with the baseline each row must be read against
The per-question gold ids are **not in this file**. They name documents in a
consumer's corpus, and this repository keeps corpus paths and document titles out
of tracked files; they live in the local project directory beside the run, and go
to `portfolio-optimiser` by coord. What is publishable is the shape:
| row | gold concepts | hit@8 | rank | chance, analytic | chance, empirical |
|---|---|---|---|---|---|
| 1 — **the concept-level row** | 1 | ✅ | 1 | 0.013 | 0.013 |
| 2 | 49 | ✅ | 1 | 0.479 | 0.483 |
| 3 | 20 | ✅ | 1 | 0.229 | 0.231 |
| 4 | 43 | ✅ | 1 | 0.434 | 0.438 |
| 5 | 11 | ✅ | 1 | 0.132 | 0.134 |
| 6 | 5 | ❌ | — | 0.062 | 0.063 |
**Why every row carries its own baseline.** A document-prefix hit is *easier for
a large gold document*: with gold sets spanning 1 to 49 concepts, a raw count of
4/6 could be produced by a ranker that had learned nothing. Row 2 alone is 48 %
likely by chance; row 1 is 1.3 %. The two hardest rows (1 and 6, at 0.013 and
0.062) are one hit and one miss.
The empirical baseline is 20 000 draws at seed 20260907. It agrees with the
analytic form to within about half a percentage point — **not** to three decimal
places, and this report does not claim it does.
**The miss, diagnosed rather than absorbed.** The gold's best concept ranked
**11 overall**, just outside `k = 8`, and comes back at rank 11 for `k = 12` and
`k = 16`. Cause: the profile matches two tokens when they share four leading
characters, so a compound question term matches two much larger sibling
documents in the same subject area, and those two fill all eight slots. The
document prior does not rescue it (0.333 against 0.433 and 0.415). This is a
near-miss at the reported `k`, not an absence — and it is the shape of failure to
expect from this ranker: crowding by large same-prefix neighbours.
## The controls that failed, stated first rather than last
The order required a known-negative: a question with no answer in the bundle
must produce an empty or "not found" payload, never a confident guess. **Both
known-negative questions produced eight excerpts.** The control failed.
| control | tokens reaching nothing | delivered |
|---|---|---|
| the question chosen as planned | 2 of 6 | 8 |
| the most lexically disjoint question found | 4 of 6 | 8 |
**Why, measured rather than guessed.** The pre-pass has a per-concept relevance
rule (`no_lexical_match`) that withholds a concept the question does not reach,
and it works: it fires on 600-odd concepts per run. It does **not** work as a
whole-question gate, 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, `mange` 17, `brukes` 83,
`etablert` 38. Four candidate questions were probed and none reached zero on
every token. The first control's content words did genuinely match — one of them
reached 17 concepts describing welded seams in floor covering, which is a correct
lexical answer and a useless topical one.
**Consequence for anyone using this, stated in the skill as well as here: an
empty `excerpts` list is evidence of absence; a full one is not evidence of
presence.** When the delivered excerpts do not answer the question, the right
output is `[sourced-not-sufficient]` plus a statement that the cut found nothing
responsive — not eight answers.
**What would fix it, named and not built.** Rarity weighting: a token matching a
large fraction of the corpus carries little information, and requiring a
delivered concept to match at least one *informative* token would make the gate
work. Smoothed IDF is on the plan's own list of mechanisms worth lifting and
this build does not implement it. It is not attempted here because this step's
scope fence freezes the instrument before it is measured — tuning a ranker
against the questions it is being scored on produces a number about the tuning.
## Honesty limits
Six, and the first three matter most.
1. **n = 6.** Six questions is a small sample and the table says so on every
reading. The chance baseline is what keeps it from being meaningless, not
what makes it large.
2. **The question texts were written after the ranker existed.** The plan fixed
the gold documents' *size profile* before any ranker existed — its recorded
per-row baselines sum to 1.35, and the sizes used here reproduce that total
exactly, which is an independent check that this is the gold set the plan
profiled. But the plan recorded no question **texts**, so those were written
during execution. Three of the six gold documents also could not be pinned
uniquely from the recorded sizes and were chosen among equal-sized
candidates. The gold documents themselves came from the corpus's own titles,
per the operator's answer, and not from anything the ranker computes — but
this is not a blind evaluation and must not be quoted as one.
3. **One ranking change was made with a gold answer visible.** The stage-one
document prior originally *summed* lexical overlap across a document's units,
which measures document size: measured on this corpus, a 79-concept document
scored 6.0 (0.076 each) against the price document's 2.0 over 1 concept, and
the price gold fell outside the top 8. The prior is now a density. The fix is
justified by the scoring function's own arithmetic rather than by the answer
— a prior that grows with size is measuring size — but it was found by running
row 1, and the ranker is therefore not blind to that row.
4. **Five of six rows are DOCUMENT granularity**, not the concept granularity
the order asked for. Row 1 is the exception and is labelled throughout.
5. **Two of the three states this contract makes consumers distinguish have
denominator zero in this corpus.** `adjudicated` never occurs (all 618 present
values are `proposed`; 11 concepts carry no key and read as `unknown`), and
`verified` is absent on all 629 — anchored and unanchored — so every excerpt
is `unverified`. `type: verdict` is likewise absent on all 629, which makes
the § 9.1 exclusion **vacuous here**. All three are exercised only against a
synthetic fixture, and the skill states that rather than implying coverage.
6. **The contract checker is a lint, not the acceptance.** Measured: it returns
0 findings on a payload declaring `considered=0, withheld=0, delivered=0`
paired with the *unfilled* template. A green checker means the payload carries
what a claim would have to rest on — not that the claim is sound. The real
gates are the negative controls, the partition assertion, the placeholder scan
and the digest test, and none of those is something the checker performs.
## What the contract's own gates reported
| gate | result |
|---|---|
| SS 5.2, `considered == withheld + delivered` | closes on every run, by construction |
| SS 5.3, every withheld concept names its rule | 621 of 621 on the row-1 run, from a closed set of six |
| SS 7.4, the instrument reproduced a known-positive before reporting | 10 349 expected, 10 349 measured, plus a second route (`wc -c` = 10 060, delta 289) |
| SS 7.2, `spent` is the **delivered set** | 74 838 spent against a 120 000 limit on row 1 — where a whole-payload reading would have put 165 109 against the same limit and refused |
| SS 3.3, the ref is a content identity, not a declared version | `sha256-tree:` over the reachable tree; `--ref` asserts and never overrides |
| SS 9.2, no directory enumerated | the index walk reaches 629 concepts, exactly the set the forbidden method finds |
| contract check on a real payload | `conformant: 14 rules over 8 excerpts and 621 withheld entries, 0 findings`, exit 0 |
| negative control: broken identity | `NOT conformant`, 2 findings, exit 1 |
| negative control: missing payload | exit 2 |
## Reproduce
```sh
# One payload, and the checker on it.
.venv/bin/python tools/okf_consume.py <bundle> \
--question "Hvordan skal prisene fylles ut?" --out payload.json
.venv/bin/python tools/okf_contract_check.py \
--skill skills/okf-consume/SKILL.md --payload payload.json
# The whole table. The gold set is an INPUT and is not in this repository.
.venv/bin/python tools/okf_consume_measure.py <bundle> --questions <gold-set.json>
```
The gold-set file's shape is in `tools/okf_consume_measure.py`'s module
docstring. Determinism is asserted by the suite, not by inspection: two runs of
the same `(bundle, question, k, limit)` produce byte-identical output, and the
module reaches no clock.