The 'never duplicate an entry' rule has held across 1,761 entries - and the audit that checked it would not have caught the near-duplicate that prompted it
also called entry novelty audit · duplicate entries · one entry per thing · curation discipline
This project insists on one entry per thing and has never had a way to check. Comparing all 1,761 entries against each other finds no real duplicates, and the handful of similar-looking pairs turn out to be exactly the distinctions the rule exists to preserve, like a drug and the mechanism it works through.
The same thing, for a scientist
The standing brief says **never duplicate an entry**. Every other rule here has a checker; this one has been enforced by memory across 1,761 entries and many agents. Cycle 38 showed why that is fragile: the coordinator nearly committed a king entry restating `hyp-menin-dependency-tracks-available-mbm`, and caught it by chance. Pairwise IDF-weighted overlap over title plus `summary_expert` plus evidence claims, blocked on rare tokens: | | | |---|---:| | entries scored | 1,761 | | pairs at or above 0.60 | 23 | | not same-prefix-same-directory | **16** | | **true duplicates on reading** | **0** | **All 16 are separations the rule requires.** `drug-panobinostat` against `mechanism-panobinostat-depletes-h2bub-mll-r-all`, the drug and the mechanism it acts through; `gene-supt16h` against `protein-fact-complex`, subunit and complex; `gene-hla-dra` against `target-hla-dr-class-ii-heterodimer`, where the second entry's own title says no single gene symbol denotes the surface marker a flow report names; `fusion-kmt2a-abi1` against `gene-abi1`. **One entry per thing produces near-neighbours by design, and finding them is evidence the rule is working rather than failing.**
The audit would not have caught the thing that prompted it, and that is the finding
The cycle-38 near-duplicate shared a conclusion with hyp-menin-dependency-tracks-available-mbm, not a vocabulary. One draft was framed as menin dependence runs through wild-type MLL1, so drug response does not identify the route; the existing entry as menin dependency tracks an intact menin-binding motif in cis or in trans. The same claim in different words, and a lexical overlap test scores that low.
So this audit bounds one failure mode and not the dangerous one. What actually caught the near-miss was noticing that the load-bearing PMID was already cited in 34 entries — a citation-reuse signal, not a text-similarity one.
Which points at the cheap instrument that does not exist yet. Before writing an entry, count where its key citation is already used. A PMID appearing in 34 places is a strong signal that the conclusion is already written somewhere, and it costs one grep. For this corpus that is worth more than semantic similarity would be, because the graph is built around a small set of load-bearing papers.
| n entries | 1761 |
|---|---|
| pairs ge 060 | 23 |
| non neighbour pairs | 16 |
| true duplicates on reading | 0 |
| what caught the cycle 38 near miss | the load-bearing PMID was already cited in 34 entries - a citation-reuse signal, not text similarity |
example legitimate neighbours
[ "drug-panobinostat vs mechanism-panobinostat-depletes-h2bub-mll-r-all", "gene-supt16h vs protein-fact-complex", "gene-hla-dra vs target-hla-dr-class-ii-heterodimer", "fusion-kmt2a-abi1 vs gene-abi1" ]
- Do not cite this as showing the graph has no duplicates. It shows no pair restates another in substantially the same words. The cycle-38 near-miss shared a conclusion with different vocabulary and this instrument scores that low.
- Would a citation-reuse check be the better instrument? Before writing an entry, count where its key PMID is already cited. It costs one grep and would have caught the cycle-38 near-miss, which this audit would not.
- Are conclusions duplicated across differently-worded entries? Unmeasured, and it is the failure mode that matters.
- Should the 16 near-neighbour pairs be cross-linked? Several are not edged to each other, so a reader arriving at gene-supt16h may never find protein-fact-complex.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
Pairwise IDF-weighted overlap over 1,761 entries returns 23 pairs scoring at or above 0.60, of which 16 are not same-prefix-same-directory neighbours, and on reading all 16 are distinctions the one-entry-per-thing rule requires rather than duplicates.
entry-novelty-audit/run.py: pairwise IDF-weighted overlap over every entry title, expert summary and evidence claim in the graph · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
**Lexical overlap cannot see a duplicate written in different words**, which is the dangerous kind and the kind that prompted this audit, so a zero here bounds near-verbatim duplication only. Blocking on the six rarest tokens per entry means a pair sharing only common vocabulary can be missed, and the 0.60 floor is a choice rather than a derived threshold. Entries under 12 content words are skipped, so stubs are invisible. 'True duplicates on reading' is the coordinator's judgement over 16 pairs, not an independent adjudication.
Evidence that argues against this (1)
- A NULL FROM AN INSTRUMENT THAT CANNOT SEE THE MAIN FAILURE MODE IS WEAK EVIDENCE THAT THE RULE HELD. The corpus could contain paraphrased duplicates at any rate and this would report the same zero. The honest reading is narrower: no pair restates another in substantially the same words, and the discipline around gene/protein/complex/fusion separation is visibly intact. Whether conclusions are duplicated across differently-worded entries is unmeasured, and cycle 38 produced one instance of exactly that. Oncofusion knowledge graph (internal analysis) 2026
THE ENTRY THE COORDINATOR NEARLY DUPLICATED, and the reason this audit exists. It was caught by noticing its load-bearing PMID was already cited in 34 entries, not by any text-similarity check.