Two rows from the recall-bench ledgers after #583, each a grammatical edge the span check gets wrong.
A coordination is read as a description. "SB Energy and SoftBank will build at least 10 GW of new energy generation" came out with subject SB Energy and subject_span: "SB Energy and SoftBank". Tier 4 says a bound name followed by lowercase words is a modifier inside a larger phrase; and is lowercase, so the span was subject_described and a true partner_with fact was dropped. In A and B, A is one item of a list, not a modifier: the fact holds of A.
A single capitalised word rebinds to any name containing it. "Stockholders approved the election of each of our ten (10) director nominees" came out with subject NVIDIA and subject_span: "Stockholders"; the same response declared 2026 Annual Meeting of Stockholders of NVIDIA Corporation. Tier 3 rebinds a span whose words are all inside another declared name, and lets a single word through when it looks like a proper name; a sentence-initial capital is not that evidence, and the fact moved from NVIDIA to the meeting.
What should change
and and & join the articles as grammar words: a bound name followed by a conjunction and then proper-looking names is one item of a list and is kept (SB Energy and SoftBank bound to SB Energy); followed by a conjunction and a description it is still a description (OpenAI and its investors); bound to the later item it is the prefixed tier as today.
- A single-word span does not rebind by containment at all; it rebinds only by exact or stem match (
slot_matches), as multi-word spans still rebind by containment (OpenAI's board → OpenAI's board of directors). No single-word containment rebind has been observed to be right in any ledger so far.
Acceptance
- The two sentences above, as unit tests, keep their facts on SB Energy and NVIDIA.
- The existing span tests hold;
Altman bound to OpenAI is now a coreference rather than a rebind, and the test says so.
Related: #582 / #583.
Two rows from the recall-bench ledgers after #583, each a grammatical edge the span check gets wrong.
A coordination is read as a description. "SB Energy and SoftBank will build at least 10 GW of new energy generation" came out with subject
SB Energyandsubject_span: "SB Energy and SoftBank". Tier 4 says a bound name followed by lowercase words is a modifier inside a larger phrase;andis lowercase, so the span wassubject_describedand a truepartner_withfact was dropped. InA and B,Ais one item of a list, not a modifier: the fact holds ofA.A single capitalised word rebinds to any name containing it. "Stockholders approved the election of each of our ten (10) director nominees" came out with subject
NVIDIAandsubject_span: "Stockholders"; the same response declared2026 Annual Meeting of Stockholders of NVIDIA Corporation. Tier 3 rebinds a span whose words are all inside another declared name, and lets a single word through when it looks like a proper name; a sentence-initial capital is not that evidence, and the fact moved from NVIDIA to the meeting.What should change
andand&join the articles as grammar words: a bound name followed by a conjunction and then proper-looking names is one item of a list and is kept (SB Energy and SoftBankbound to SB Energy); followed by a conjunction and a description it is still a description (OpenAI and its investors); bound to the later item it is the prefixed tier as today.slot_matches), as multi-word spans still rebind by containment (OpenAI's board→OpenAI's board of directors). No single-word containment rebind has been observed to be right in any ledger so far.Acceptance
Altmanbound to OpenAI is now a coreference rather than a rebind, and the test says so.Related: #582 / #583.