Product Quickstart

Run TargetSpace against a product memory or context feature

A step-by-step recipe for AI product teams — note-taking and memory apps, wearable audio recorders, multimodal glasses, and enterprise assistants — to turn a memory or personalization claim into sealed, resolved, scored forecasts. It measures target-specific predictive skill and separates it from generic prediction or routine replay.

Release Version 1.0 Status pre-pilot protocol proposal Results none claimed no product has run this protocol yet

How to run TargetSpace against a product memory/context feature

A protocol usage guide, not an empirical claim

If a product asserts that its memory or context feature helps a specific user, TargetSpace asks it to make that claim falsifiable: forecast, before the outcome exists, how this user will transition next, then score the forecast against the population prior R1, the user's own routine R2, a calibration check, and a target-permutation control. The recipe below reuses the released v1.0 schemas and adds no new record types.

◉

Everything on this page is a protocol usage guide. No product has run TargetSpace, and no empirical results exist. A run over real users must satisfy the binding consent, privacy, and governance requirements — consenting adults only, on-device filtering, no raw-media export, aggregate-only reporting. A high score certifies calibrated prospective predictive skill about a consenting target only — never understanding, inner life, causation, or permission to act on a person.

The minimum-viable benchmark path

Seven steps, each mapped to one released schema

The smallest defensible run is seven steps. Each maps to a single schema file in the v1.0 harness.

Choose consenting target instancesSelect consenting users and a sealed evaluation window; register each as a target (participant.schema.json) with a routine profile and allowed tasks.

Choose a task familyPick one or more rows from the quickstart pack below. Each fixes a task type, a discrete answer space A, and a deterministic resolution rule.

Freeze evidence tiersDeclare which evidence tiers each condition may access (evidence_manifest.schema.json): modalities, evidence start/end time, hashed manifest. One manifest per tier condition drives the evidence-tier ablation.

Generate sealed forecasts before outcomes existBefore any outcome is observed, emit one forecast per instance (forecast.schema.json): a probability over every element of A, an evidence cutoff time t, and a hash sealing the payload. Stored write-once.

Resolve outcomes deterministicallyAt resolution time r, apply the task's resolution rule to later observable evidence and write the outcome (outcome.schema.json). Outcomes are withheld from the system under test until after sealing.

Score against R1 and R2Compute log score in bits (primary) and Brier per instance; report Skill over the population prior R1 and over the own-routine baseline R2, plus calibration. A claimed improvement must beat R2, not merely R1 or replay.

Run the permutation control and report the standard rowRe-score each system against matched wrong-target outcomes; target-specific skill must collapse. Emit one leaderboard row (leaderboard.schema.json) with the standard columns.

◉

The four-bar decision rule: a forecast earns target-specific credit only if it (1) beats the R1 population prior, (2) beats the R2 own-routine baseline, (3) stays calibrated, and (4) loses skill under a matched target permutation.

Quickstart task pack

Six product-relevant task families

Each family is a task type with a fixed answer space A and a deterministic, pre-registered resolution rule over later observable evidence. All are scored against R1/R2 with log score in bits, calibration, and the permutation gate.

Task family	Answer space	Resolution rule (deterministic)
Recurring-commitment completion	{complete, defer, cancel, replace}	Calendar/task status or a pre-registered behavioural marker at the next scheduled occurrence fixes the label.
Meeting / event realization	{attended, no-show, rescheduled, cancelled}	Attendance signal (join event, location match, or logged presence) within the event window.
Response-latency bucket	{<1h, 1–24h, 1–3d, >3d, no-response}	Time from a flagged inbound obligation to the first outbound reply, bucketed; no reply before the horizon resolves to no-response.
Task continuation vs. switch	{continue, switch, pause}	The active task at t versus the active task after the next transition boundary; continuation iff the same task persists past the window.
Priority maintained vs. displaced	{maintained, displaced}	Whether the top-priority commitment at t still receives sustained allocation after the window, or is supplanted by a newly dominant one.
Engagement vs. avoidance	{engaged, avoided}	Engaged iff a substantive action on the named obligation (reply sent, document edited, task advanced) occurs before the horizon; otherwise avoided.

No subjective report resolves an outcome. Self-report may be admitted as evidence, but is never the outcome label when it is also a model input. A target state is scored only if it has a pre-registered observable resolution rule; otherwise it is evidence, not a scored state.

Four product instantiations

The same protocol, mapped to real products

Each product maps to an evidence-tier band and a natural pair of task families. Higher tiers carry higher sensitivity, and the evidence-tier ablation measures whether they actually buy skill over a lower tier.

Note-taking / memory app L0–L1

Evidence = notes, tasks, calendar, and communication metadata already in the app. Natural tasks: recurring-commitment completion and response-latency bucket. Resolution is read from the app's own state at the horizon. No new sensing; the manifest declares tiers L0–L1 only.

Wearable audio recorder L2

Evidence = on-device transcripts and derived commitments/entities from ambient speech, not raw audio. Natural tasks: meeting/event realization and engagement vs. avoidance of a spoken obligation. Transcription stays local; only sealed forecasts and resolved labels leave the device.

Multimodal glasses L3–L4

Evidence adds embodied context — objects, screens, documents, and locations in view. Natural tasks: task continuation vs. switch and priority maintained vs. displaced, where visual attention disambiguates transitions. Bystander redaction and on-device filtering are mandatory, and the ablation measures whether L3–L4 buys skill over the L2 audio-only condition.

Enterprise assistant L0–L1

Evidence = consented organizational memory over trackers, commits, calendars, and communication metadata. Natural tasks: task continuation vs. switch and priority maintained vs. displaced for a tracked workstream. The target is a consenting individual or a defined unit; R2 is that unit's own routine, and skill must still collapse under permutation.

These are product formulations of the protocol, not validated empirical results. Attention, affect, and inferred goals stay evidence; they are never scored states.

The standard reporting row

What every run must report

A TargetSpace result is one row. Reporting a headline number without its controls is not a TargetSpace result.

Skill vs R1Bits gained over the population prior — the entry condition the system must clear to exceed base rates.
Skill vs R2Bits gained over the target's own routine — the headline number; a claimed improvement must beat R2.
Calibrationpass / warn / fail via intercept, slope, and reliability; ECE is a coarse diagnostic at small n, not a hard gate.
Permutationcollapses / survives when forecasts are scored against a matched wrong target.
Evidence-tier liftSkill as a function of the evidence rung — what each additional tier buys, measured rather than assumed.
Resolved forecastsNumber of forecasts with a determined outcome.
Independent targetsNumber of distinct targets contributing to the row.
HorizonTime between sealing t and resolution r.
Target domainWhich target class and task family the row covers.
Result statussynthetic / private / public.

⚠

A large forecast count from few targets is not a large independent sample. Many sealed forecasts drawn from a handful of targets do not license population-level claims; report the number of independent targets alongside the number of resolved forecasts, and stratify low-coverage targets separately rather than pooling them to inflate skill.

Before any real run

Consent and governance are binding, not optional

Any run over real users requires consenting adults, per-instance consent and eligibility gating, ethics or IRB review before recruitment, bystander redaction, local-first storage, on-device filtering, and aggregate-only reporting — no raw media or transcripts leave the client, and benchmark validity is kept strictly separate from any permission to deploy or act. See the governance requirements in full before recruiting a single target.