Reconstruction differences
Required by spec section 10: every deviation from the expected counts, and every internal contradiction found in the specification itself, is recorded here rather than silently absorbed.
Build: dataset3/, sources read from ICLR27/dataset/.
Summary
| Stage | Expected | Actual | Status |
|---|---|---|---|
| Raw cloze prompts | 1,783,541 | 1,783,541 | match |
| Five-model consensus | 10,601 | 10,601 | match |
| Candidate-Known Facts | 8,107 | 8,107 | match |
| Grouped (subject, relation) | 6,185 | 6,185 | match |
| Functional conflicts removed | 5 | 5 | match |
| Enriched Facts | 6,180 | 6,180 | match |
| Core-source facts | 3,705 | 3,705 | match |
| Benchmark facts | 2,592 | 2,592 | match |
| Benchmark relations | 21 | 21 | match |
| Migration duplicates removed | 10,323 | 10,323 | match |
| Evaluation queries | 44,418 | 44,416 | deviation D1 |
| — context | 6,831 | 6,829 | deviation D1 |
| — all other conditions | as specified | as specified | match |
Everything except D1 reproduces exactly.
D1 — Two context queries expose the gold answer (defect in the source data)
Deviation: context 6,831 → 6,829; total 44,418 → 44,416.
Spec 9.4 requires that context "must not introduce a second valid target", and
spec 13.9 that "no gold answer is exposed in open-ended main conditions". Two
queries migrated from eval_dataset/queryD_context.jsonl violate both, because
the legacy distractor sampler picked an entity whose surface form contains the
gold answer:
Some people confuse States Reorganisation Act with Indian Act.
In which jurisdiction does States Reorganisation Act apply?
gold = India <- exposed by the distractor "Indian Act"
Taiwan (ROC) Resident Certificate is a legal term in Taiwan.
Indian Act is a legal term in Canada.
In which jurisdiction does Permanent account number apply?
gold = India <- same cause
Spec 10 allows a count change when "the implementation discovers a documented defect in the old data". This is such a defect.
Action taken. src/build_eval_bank.py applies a general rule rather than a
hard-coded exception: for every open-ended main-forward query, the normalized
gold must not appear in the query text after removing all mentions of the
subject. Queries failing the rule are dropped and counted in
stage4_report.json as dropped_gold_exposed.
To restore 44,418 exactly, delete that guard — but the two queries are unanswerable-by-design artefacts and any model scores them correct by copying.
C1 — Spec 6.2 contradicts the required benchmark size (unresolved by design)
Not a count deviation. An internal contradiction in the specification.
Spec 6.2 lists relations "generally unsuitable" for the main benchmark, naming among them developer of and manufacturer of. Spec 1, 6.2 and 15 simultaneously require exactly 2,592 facts across 21 relations.
Those two relations contribute:
| Relation | Facts |
|---|---|
| manufacturer (P176) | 441 |
| developer (P178) | 405 |
| total | 846 |
Excluding them yields 1,746 facts across 19 relations, which fails the acceptance criteria in spec 15. Keeping them satisfies the counts but contradicts spec 6.2's guidance.
Action taken. The counts win, because spec 15 states them as acceptance criteria and spec 10 forbids silent changes. Both relations are retained and tagged:
"spec_6_2_flag": "unsuitable_named_in_spec"
configs/relations.yaml exposes options.drop_spec_6_2_flagged. Setting it to
true reproduces the spec-6.2-compliant 1,746-fact variant; the selection script
then reports the resulting size instead of failing.
Recommendation. Report the headline number on all 2,592 and a robustness number on the 1,746 subset. See F1, which gives an independent empirical reason to distrust exactly these relations.
F1 — 28.9% of benchmark facts are answerable by string copying
Finding, not a deviation. No counts changed.
For 750 of 2,592 benchmark facts (28.9%) the object string is literally contained in the subject name, so the query can be answered by copying a substring instead of retrieving knowledge:
Airbus A318 -> manufacturer Airbus
Adobe Acrobat -> developer Adobe
Amazon Music -> owned by Amazon
Agriculture and Agri-Food Canada -> jurisdiction Canada
Concentration by relation:
| Relation | Flagged | Total | Share |
|---|---|---|---|
| manufacturer | 387 | 441 | 87.8% |
| developer | 199 | 405 | 49.1% |
| owned_by | 52 | 89 | 58.4% |
| applies_to_jurisdiction | 31 | 54 | 57.4% |
| sport | 17 | 81 | 21.0% |
| country | 17 | 263 | 6.5% |
| headquarters_location | 15 | 132 | 11.4% |
| (10 further relations) | 32 | — | <34% |
586 of the 750 come from manufacturer and developer — the two relations spec
6.2 already called unsuitable. F1 is therefore independent empirical support for
C1's recommendation.
Why this matters. A stability benchmark measures whether a retrieved fact survives rephrasing. A fact answerable by copying is stable for a reason that has nothing to do with knowledge, so it inflates every retention number and dampens the very effect the study is about.
Action taken. Spec 16 forbids silently dropping facts, so all 750 are kept and flagged on both the fact and the query record:
"answer_in_subject_surface": true
Downstream analysis should report retention with and without them.
F2 — 808 of the 8,107 candidate rows are exact duplicates
Finding, not a deviation.
The 8,107 Candidate-Known Facts contain only 7,299 distinct
(subject, relation, object, source) tuples; 808 rows are exact repeats, e.g.
("BMW 3 Series", P176, "BMW", counterfact) appears twice. This is inherited
from dataset/filter/prompts.jsonl.
No action needed: spec 5.1 grouping absorbs them, which is part of why 8,107 rows become 6,185 groups. Recorded so the 8,107 figure is not mistaken for 8,107 distinct facts.
N1 — Stage I GPU passes were not re-executed
The five-model first-token forward passes (1,783,541 prompts x 5 models) were
run previously; their per-prompt verdict vectors are stored as
dataset/filter/correct_<model>.npy.
src/build_candidate_known.py re-derives the candidate set from those vectors:
it re-reads the prompt table, re-computes the intersection, re-applies the
source-drop rule, and checksums every input into stage1_report.json. It does
not re-run inference.
The counts are verified, not assumed — 1,783,541 / 10,601 / 8,107 all reproduce,
and the result is checked for set equality against
dataset/filter/global_known_clean.jsonl.
src/filter_run.py regenerates the vectors from scratch when required.
N2 — Triple direction is not normalized for capital_of
Spec 2.2 requires a single triple direction, forward being
subject + relation -> object. The relation capital_of (P1376) is stored the
other way round: subject = city, object = country.
Normalizing it would fold those 125 facts into capital, changing both the
benchmark size and the relation count, which spec 10 forbids doing silently.
Action taken. capital_of keeps "direction": "inverse" and
"inverse_of": "capital" in its record, so any analysis can group or separate
the two directions explicitly. configs/relations.yaml exposes
options.normalize_inverse_direction for the alternative build.
Coverage gaps (spec 12, expected and unchanged)
Main-forward families do not cover every fact:
| Condition | Facts covered | Missing |
|---|---|---|
| anchor | 2,592 | 0 |
| paraphrase | 2,592 | 0 |
| format | 2,592 | 0 |
| context | 2,591 | 1 |
| multilingual | 2,402 | 190 |
Missing fact ids are listed in outputs/coverage_report.json. No query was
imputed. Downstream analyses that compare conditions per fact should use the
explicit valid intersection.