A statement says its mood

[WaylandYang]

Pitfall 2 of docs/design/prior-work.md: a triple the sentence did not assert (OLLIE 2012; speculation is a property of the tuple, Dong et al. 2023).

What lands

A statement says its mood. A statement the passage requires, plans, expects, forecasts or makes conditional carries a qualifier keyed mood whose value is the passage's own words for that ("要", "should", "will", "is expected to", "if the merger closes"); the phrase stays as written. A verb that reports (announced, said, revealed) is not a mood: what was announced is stated. Alignment (0044 cut 2) will read the qualifier and never materialize a statement with a mood as a typed fact; nothing in the ledger changes, the qualifier is a statement_qualifiers row like any other.

On the NVDA batch the model writes 44 moods on 901 statements: will on the plans of the PORTS-Pike release, Outlook on every figure of the guidance table ("NVIDIA —GAAP operating expenses→ $9.2 {mood: Outlook}"), are expected to be on the outlook prose, subject to definitive agreements on the acquisition terms, If an emerging growth company on the cover-page election. No reporting verb was written as a mood.

Tried and withdrawn in the same change: a rule giving a subjectless sentence the addressee or the speaker the passage names as its subject. Under it the model returned an empty reply for a fifth of a filing's chunks, tables included, and the bullets of a release already take the company as their subject without it. The trade-in notice's directive paragraphs, whose addressee is outside the paragraph unit, stay an open item.

Table rows quoted without their leading pipe now still find their period in the chunk: the check looks at the line the located quote sits on, not at the quote's first character (the model often drops the | at the start of a row).

Measured

Same four NVDA documents, the same hour, the extraction endpoint with reasoning on (see below), the same fast judge:

four documents	#744	this change
statements	864	901
misworded	3%	5%
not stated	0%	0.3%
reads alone = false	4%	5%
prose only, misworded	6%	9%

The extra misworded statements are the securities boilerplate ("Section 21E —created→ safe harbor", "partners —are→ forward-looking statements"), not the moods; within the 3-point band this branch set for itself.

The endpoint, not the contract, was the variable all afternoon. Three earlier runs of this change measured 10% misworded and a fifth of the chunks empty. A control with the #744 binary in the same hour reproduced that (460 statements, 7 empty table chunks, 33% misworded on the exhibit) while the proxy's thinking: disabled was honoured, and reproduced #744's own numbers (864 statements, 3%) once reasoning was switched on explicitly. Every earlier measurement of #731 to #744 had reasoning on by accident. Extraction with this model needs reasoning on; the cost is 17k to 40k completion tokens a call rather than 700, and one to two minutes a call. The bench notes record it; a run's completion-token count is now the first thing to check before comparing two runs.

Not in this change

Alignment reading mood; the subject of a directive sentence whose addressee is outside the chunk; a bench that measures run-to-run stability (ATOM 2025) as its own number.

🤖 Generated with Claude Code