medlab / family
Human Comprehension
Cases in the observatory's Human Comprehension family.
cases9
-
2026-11-02
A workaround written down is assumed to still be the design
now fails loudly None mechanical. The paragraph now describes what the tool does, and the neighbouring section says the same thing, which is the only check this document has.
Human ComprehensionEvolutionDocumentation DriftContract Tests
-
2026-11-02
A review verdict belongs to the commit you are looking at
now fails loudly A posted verdict names the head it read, so the server binds it to that commit, and a standing refusal names the commit it was given at rather than the commit under merge.
InvariantsHuman ComprehensionContract TestsDecision Provenance
-
2026-11-02
A review on your own pull request is assumed to be invalid
now fails loudly The refusal covers verdicts only: an author may post a comment review on their own pull request, and a test requires that path to exist, so the handback has a channel.
InvariantsHuman ComprehensionIndependent ReviewContract Tests
-
2026-11-02
A list read at its page size is the oldest page, not the record
now fails loudly Both readers page until a page comes back short, bounded at twenty pages of fifty and failing closed at the bound; tests assert that an approval on a later page is found and a refusal on a later page still blocks.
Human ComprehensionInvariantsContract TestsExample-Based Tests
-
2026-11-02
A model's quotation of a sentence is the sentence
now fails loudly The fidelity check compares words after folding characters that carry no meaning, stores the source's own characters when a sentence matches that way, and reports a validator refusal as a refusal rather than as malformed output.
InvariantsHuman ComprehensionSchema ValidatorProperty-Based Testing
-
2026-11-02
A word is the first substring that matches it
now fails loudly Cloze construction selects a whole source word by position, and a test requires that a target which is a substring of a longer word is refused rather than cut.
Human ComprehensionInvariantsExample-Based TestsMetamorphic Testing
-
2026-11-02
A commit's CI result is assumed to be reachable through a pull request
now fails loudly `forge checks` answers for a pull request number, a revision, or the branch's head, and refuses an argument it cannot read as either rather than guessing.
Human ComprehensionInvariantsContract TestsExample-Based Tests
-
2026-10-26
A page that leaves no receipt is assumed not to have fired
now fails loudly The probe logs what the messaging service answered for every page, accepted with its message id or refused with the reason, so a page's delivery is a fact instead of an inference.
Human ComprehensionRuntimeSynthetic MonitoringDecision Provenance
-
2026-09-12
An AI reviewer that reports an approval has posted one
now fails loudly A review counts only when the forge API shows it submitted with a verdict; until the tool submits, the merge record is the evidence and the reviewer's own result is not.
ChangeHuman ComprehensionIndependent ReviewSecond-Agent Review