Enterprise copilots retrieve documents, summarize meetings, and accumulate organizational memory. What they are not evaluated on is whether any of it produces calibrated, target-specific foresight: will this commitment land by this review date? TargetSpace states that question as a sealed-forecast protocol with baselines and controls.
Version 1.0 — pre-pilot protocol proposal. Synthetic harness only. No human-subject results.
The category has climbed from meeting intelligence (transcribe, summarize) to memory systems (persistent recall across tools and threads). The rungs above — longitudinal evidence, target-specific predictive capability as TargetSpace defines and measures it, and validated assistance — require something recall does not test: forecasts about future observable states, sealed before the outcome exists. Memory is not understanding, and a copilot that recalls every commitment perfectly may still have no skill at forecasting which ones will slip.
Current rung: structured memory. Enterprise copilots are competitive on retrieval and summarization over organizational corpora. Next rung: longitudinal evidence — using an individual's history in correct temporal order to beat their own routine on forecasts, a property that must be demonstrated, not assumed.
Enterprise settings supply L0–L1 evidence tiers: organizational exhaust — calendars, issue trackers, commits, documents, and communications metadata — for consenting individuals evaluated under the flagship TS-Personal formulation. Whether each stream adds forecasting lift is an empirical question answered by the evidence-tier ablation, never assumed. TS-Enterprise, an organizational-level track, is research-status and not instantiated; v1.0 makes no commitment to it.
For a consenting employee under TS-Personal: forecast whether a specific project commitment resolves as {complete, defer, cancel, replace} by its pre-registered review date. The forecast is a full probability distribution over the four outcomes, sealed and hashed at time t; resolution is deterministic from the tracker state at the review date. Skill is scored against the R1 population-prior baseline and the R2 own-routine baseline, with calibration reported and the permutation specificity gate applied: forecasts rescored against a different employee's outcomes should lose their skill.
Consent boundary. Evaluation covers consenting individuals only, with the observation window and evidence tiers declared in the manifest. Workplace surveillance is a named risk in the protocol, not a use case.
Pick one copilot feature — organizational memory, retrieval, meeting summaries — and run feature-on vs. feature-off over the same sealed forecast tasks. The difference in skill, in bits, is the feature's measured contribution to foresight rather than to answer quality.
Raw organizational data never leaves the company. Forecasts are sealed inside the boundary; only sealed forecasts, resolved outcomes, and aggregate metrics are exported for scoring. The evidence manifest declares what was observed, at which tier.
One leaderboard row per condition: skill vs. the R1 population-prior baseline, skill vs. the R2 own-routine baseline, calibration, the permutation-gate result, and lift per evidence tier. A single headline number without the battery is not a result.
Does organizational context — dependency graphs, team load, meeting density — add forecasting skill over the individual R2 own-routine baseline, or is most of the signal already in the individual's history? A question for the evidence-tier ablation.
Can R2 be defined for a team or workflow — a recency-weighted routine of the unit itself — strongly enough to license an enterprise track? Without a defensible org-level R2, TS-Enterprise stays research-status.
How should consent be structured when the observer is an employer? Opt-in that is not freely refusable is not consent. The protocol treats workplace surveillance as a named risk and requires governance answers before any enterprise instantiation.
Which copilot features — retrieval, summarization, persistent memory — improve calibrated prospective skill rather than perceived answer quality? Feature A-B runs over sealed tasks can separate the two; current evaluations cannot.
Where does organizational evidence end and personal life begin — a calendar entry, a commit timestamp at midnight, a message header — and how is that boundary declared and enforced in the evidence manifest so that scoring can audit it?
These are open questions, not roadmap commitments. v1.0 instantiates TS-Personal synthetically; everything above is stated so it can be tested.