Meeting recorders, notetakers, and meeting-intelligence systems transcribe, summarize, and extract action items — and are evaluated on transcript accuracy and summary quality. None of that tests whether the system has learned anything about the people in the room. TargetSpace poses the missing question: given weeks of a team's meetings, can the system forecast what a specific person does next, better than base rates and better than that person's own routine?
Version 1.0 — pre-pilot protocol proposal. Synthetic harness only. No human-subject results.
Personal-AI product categories are converging on one ladder: recording → structured memory → longitudinal evidence → target-specific predictive capability → validated assistance, with the predictive-capability rung measured by TargetSpace. Meeting AI is the ladder's natural entry point: it already captures the among the richest recurring evidence streams in knowledge work — who said what to whom, and what was promised — and its products are now adding cross-meeting memory. But a summary of a meeting is not understanding of the people in it; the ladder is climbed by forecasting, not by recall.
Current rung: recording, extending into structured memory (search and recall across past meetings). Next rung: longitudinal context — treating weeks of meetings as one evidence stream about specific people, evaluated by whether it improves sealed, calibrated forecasts of their observable next states.
Meeting AI maps onto the flagship TS-Personal track with an unusually strong tier-2 stream. The category's central empirical question is whether that stream earns its cost.
Meeting audio, ASR transcripts, and speaker turns: who committed to what, in which words, under whose eyes. High-bandwidth, high-cost, privacy-sensitive — and its forecasting value is a hypothesis to be tested by evidence-tier ablation, never assumed.
Calendar and communications metadata plus user-authored notes and task lists: cheap, ambient, already logged. The ablation asks how much of the forecast skill survives when the L2 stream is removed — and how much the L2 stream adds on top. Self-reported notes stay in as an auxiliary channel and baseline, not a discard.
At a sealed time t, before the outcome exists, forecast whether a commitment assigned to a named owner in a recorded meeting resolves as {complete, defer, cancel, replace} by its stated deadline, using evidence only up to t. The answer space is discrete; resolution is deterministic, from pre-registered rules over observable workspace state at the deadline.
The forecast is scored with proper scoring rules against the R1 population-prior baseline (base rates across all owners) and the R2 own-routine baseline (this owner's recency-weighted habit). Calibration is reported alongside skill, and the permutation specificity gate requires that skill collapse when the same forecasts are matched against a different owner's history. A meeting-AI system earns target-specific credit only if it clears all four checks.
Pick one feature — say, conditioning on meeting audio or on the generated summary — and run it on/off as an A-B over the same sealed task set: same owners, same commitments, same seal times. The only difference between arms is the feature; the difference in skill is the feature's measured value.
Raw data never leaves the company. Audio, transcripts, calendars, and notes stay inside the operator's boundary; only sealed forecasts, resolved outcomes, and aggregate metrics are exported for scoring. The protocol needs the forecasts, not the meetings.
One leaderboard row: skill vs R1, skill vs R2 (in bits), calibration, the permutation result, and the breakdown per evidence tier. Not a demo, not a testimonial — a number another lab could check.
None of these have empirical answers yet; each is directly addressable by the v1.0 protocol.
Does meeting audio add calibrated forecasting lift over calendar and task metadata alone, or is most of the signal already present in the L0–L1 exhaust? Passive rich capture is the hypothesis; the evidence-tier ablation is the test.
Does conditioning on the generated summary beat conditioning on the raw transcript — or does summarization discard exactly the forward-looking evidence (hedges, phrasing, hesitation) that carries forecast skill?
How do ASR word-error and diarization error propagate into forecast skill and calibration, and can the degradation be normalized per capture device so that scores remain comparable across hardware?
Which meeting features carry forward-looking information about resolution: the assigned owner, the phrasing of the commitment, who was present when it was made, or the time pressure at the moment of assignment?
Does accumulating weeks of meetings beat single-meeting context — and by how many bits of skill over the R2 own-routine baseline? The answer locates where longitudinal context starts to pay in this category.