TargetSpace Labs evaluates whether AI systems that claim to know, remember, and personalize actually build calibrated, target-specific predictive models of the people, teams, and organizations they serve — beyond routine, retrieval, and generic baselines.
The TargetSpace platform turns the benchmark into an end-to-end workflow you can walk through today: create a run, construct target-specific prediction instances from timestamped evidence, submit sealed predictions, resolve outcomes by pre-registered rule, score against R1/R2 and the validity stack, and report evidence efficiency — with a federated, privacy-preserving mode specified for what comes next.
Define the target type, evidence streams, horizon, and resolution rules — then watch timestamped evidence become target-specific prediction instances with fixed cutoffs.
Probability distributions are hashed and sealed in a tamper-evident registry before outcomes exist; deterministic rules resolve what actually happened.
Skill is scored against R1/R2 and the validity stack, and the report shows which evidence earned its cost — toward the minimum sufficient observation.
AI products increasingly claim to know users, remember context, personalize responses, or model teams. But most evaluations measure retrieval accuracy, user satisfaction, or task completion — and every one of those can score highly while the claimed model of the person does not exist.
Memory benchmarks grade recall against a record the system has already seen. That certifies storage, not modeling.
Most of anyone's behavior is habit. A system that memorizes routines looks personalized while capturing only repetition.
Population base rates make predictions look skilled while carrying no information about any specific target.
A claimed model of a person, a team, or an organization is latent — it cannot be read off summaries or embeddings. It has to be tested against the future.
TargetSpace evaluations are prospective: the system commits to sealed predictions about a specific target before the outcome exists, and the predictions are scored against what actually happens.
Assistants and memory systems that promise to know users better over time.
Pendants, glasses, and always-on capture products whose value is the longitudinal layer.
Copilots that claim to model teams, priorities, and organizational context.
Groups studying personal world models, longitudinal AI, user modeling, and agent memory.
Digital-phenotyping and sensing researchers who need prospective, baseline-controlled evaluation.
Personalization teams whose systems claim to adapt to individual learners.
The pilot program is open to companies and research labs. It is not currently open to individual builders.
A managed, end-to-end prospective evaluation of your system on your users or study cohort — design, sealing, baselines, scoring, and a full TargetSpace Report. Our highest-value engagement.
A one-time or recurring evaluation of a personalization, memory, or modeling claim — scoped to one product surface and delivered as a standardized report.
A readiness assessment against target-specificity, calibration, and evidence-minimization thresholds — the path toward a future certification mark.
TargetSpace Eval — a self-serve evaluation platform and API — now exists as a working prototype on synthetic data. Service details →
A standardized, decision-grade artifact your team, your board, and your customers can read — built on the same controls as the open benchmark.
Target-specific lift over R1 and R2 · calibration · wrong-target specificity · shuffled-history control · evidence ablation · minimum sufficient observation.
Whether the personal-model claim holds, which evidence earns its cost, what can be deleted without losing validated skill, risk and limitation notes, and certification readiness.
Agents are gaining persistent memory. Wearables and ambient capture are shipping. Copilots are accumulating organizational context. The market's central claim — the longer we observe, the better we know you — is testable, and almost nothing tests it. The teams that can prove their claim will own the trust the category currently lacks. TargetSpace shows builders what actually improves target-specific prediction: model architecture, memory design, observation streams, calibration, and evidence efficiency.
The measurement science — sealed prospective prediction, R1/R2 baselines, calibration gates, wrong-target permutation, evidence ablation, minimum sufficient observation — is an open research standard maintained at targetspace.org. Commercial work follows the same principles and never claims more than the protocol supports.
We are accepting a limited number of companies and research labs into the founding pilot cohort. Applications are reviewed manually.