Corpus

A small set of five-channel, seven-day lives with real user-turn operations, older background evidence, hidden expectations, and failure-inclusive scoring.

Observations, not answers

Scenarios ship as versioned, serializable lives. Adapters receive only public observations: stable source id, source timestamp, arrival timestamp, channel, sender/thread metadata, and original content. Hidden grading data — expected atomic facts, accepted alternatives, supporting observation ids, forbidden stale claims — never reaches the system under test.

Model time is explicit throughout: when evidence was written, when it arrived, when collection exposes it, and when the described fact applies. Expectations follow scheduled exposure, never omniscience. A question asked before an update is exposed cannot expect the update; ambiguous or conflicting evidence may warrant an uncertain answer.

Genuinely interleaved lives

Five streams — email, iMessage, WhatsApp, GroupMe, assistant conversations — with distinct senders, threads, language, and metadata. Each life carries recurring people and ongoing commitments, ordinary low-value traffic, and developments that only resolve by combining two or more channels: year-old email with a plausible old address, this week's group-chat planning, a later venue change. Broad requests with no obvious search string ("Anything fun this weekend?", "What should I sort out before my trip?") sit beside direct, complete-list, historical, cross-source, and unanswerable questions.

The motivating case: email books poker Friday at 7 at Sam's; the organizer moves it in GroupMe to Saturday at 8 at Jordan's; a participant repeats the old plan on WhatsApp; the organizer confirms in iMessage; you tell the assistant you're still going and bringing chips. Morning questions must recover Saturday, 8, Jordan's, attendance, and the chips — while keeping Friday as history. The answer follows the meaning and authority of the evidence, not newest-wins or channel priority.

Controls and roles

Matched controls change one meaning at a time: "moved" → "booked another," the decisive update removed, cancellation → non-attendance. Consolidated-source and distributed-source versions of selected histories test joining partial evidence versus reconciling conflicting complete statements.

Three corpus roles guard against overfitting: development cases for debugging, a small calibration set for reviewing grader and wiring, and held-out lives for measurement — split by story family, so a renamed poker night is never new evidence. Once a held-out life's details tune code or prompts, it is marked exposed and replaced.