One fact should not answer two questions

[jepegit]

RET-06 fired 42 times across the stress corpus: two cards whose answer is the same sentence of the source. The listener is asked two questions and given one fact, and retrieval practice is spent without being had.

one sentence answers 2 questions ('SEI resulting', 'ethylene carbonate')
one sentence answers 2 questions ('crystalline silicon', 'irreversible capacity')
one sentence answers 2 questions ('materials discovery', 'machine learning')

Where it came from

The section-closing cards already avoided this, through _unshared_support, and the reasoning written there is the right reasoning. The segment prompts drew through SupportPool.take instead, whose ledger is keyed per concept — correct for exposition, where one sentence can reasonably serve two ideas, and wrong for a question. A sentence already answering one was free to answer another.

Compared by span, not by text: the answers differ by construction, each opening on a lead that carries its own subject's name, which is exactly why comparing whole answers reports nothing and looks like a rule that passes.

A prompt with nothing unshared left is dropped rather than asked. Unlike a section's closing question, which STR-06 requires, a segment prompt is already optional and capped.

This emptied KNOWN_FINDINGS, which is that mechanism working

tests/unit/test_script_lint.py records rules that fire on the clean fixture plan together with the reason, and fails when an entry stops firing. The entry said:

a section whose concept has only one usable sentence still shares it; the planner prefers an unshared sentence and falls back rather than losing the section's question

True of the section-closing cards — and it said nothing about the segment prompts, where the sharing actually came from. The reason is kept in the file as a note beside STR-09, which left the same way.

Corpus

	before	after
RET-06	42	0
total warnings	902	864
errors	2	2

🤖 Generated with Claude Code