Project / Governance

The inference object, not only the raw signal, is the privacy boundary

Deleting or never storing raw audio and video is not enough if the layers derived from it persist. TargetSpace treats the derived experiential model — the transcripts, embeddings, maps, routines, and inferred goals a system builds about a target — as the object that requires governance, not only the raw signal it was built from.

  • VersionVersion 1.0 · pre-pilot protocol proposal
  • ScopeGovernance principles for the benchmark; binding on any real-data submission
  • DatasetNo raw public first-person dataset in v1.0
Status  pre-pilot protocol proposal Reporting  consent-first · aggregate-only no public raw first-person dataset in v1.0

The derived artifacts that require governance

Raw capture is the smallest part of the risk surface. A longitudinal system converts a recording into a durable, searchable model of a person. Each of the following derived artifacts is a governed object in its own right, whether or not the source media is retained:

Transcripts & OCR

Spoken words and read text lifted out of transient audio and video into persistent, indexable records.

Embeddings

Vector representations that remain queryable and re-identifiable long after the source signal is gone.

Scene labels

Places, activities, and contexts tagged from first-person capture.

Speaker & entity maps

Who was present, who said what, and how people and things relate across time.

Commitments

Promises, plans, and obligations extracted from ordinary conversation.

Routines

Recurring patterns of behaviour — the target's own baseline, R2, is itself a sensitive derived model.

Summaries

Condensed narratives that outlive the moments they compress and are easy to redistribute.

Inferred goals

Intentions and priorities attributed to a person — evidence in the benchmark, never a scored state.

Derived target-state models

The predictive representation of a specific person assembled from all of the above — the object TargetSpace measures, and the object that most needs governance.

Attention, affect, and inferred goals are treated as evidence, never scored states. A target state is scored only if it has a pre-registered observable resolution rule; otherwise it is evidence. Self-report may be evidence but is never the outcome label when it is also a model input.

Observe, do not intervene

TargetSpace scores sealed calibrated forecasts against outcomes that resolve on their own. During a benchmark window the system is an observer: it emits a sealed probability vector over a discrete answer space and waits for a deterministic, pre-registered resolution rule to settle the outcome.

Observe-not-intervene rule. A system under evaluation must not act on, nudge, or otherwise alter the target during a scoring window. Intervening would both contaminate the outcome and cross the line the benchmark is careful to keep: TargetSpace measures whether a forecast was skilful, not whether a system may act on a person. Causal claims require intervention and are outside the default benchmark.

Consent-first, local-first, aggregate-only

Real-data participation is bound by architecture, not just policy. The principles below are requirements for any submission that uses longitudinal passive-observation data.

Consent-first

Every target is a consenting participant with a specific, revocable purpose. Derived, consented artifacts only — never raw bystander media.

Federated & local-first

Local processing and open-source implementations are valuable mitigations. Data stays with the target where feasible; the protocol travels to the data.

Aggregate-only reporting

Results are reported as aggregate skill, calibration, and specificity statistics — never as a reconstructable record of an individual.

Local and federated processing reduces exposure but is incomplete against bystander consent, purpose drift, and derived-inference risk. It is a mitigation, not a solution.

Bystander and third-party consent

First-person capture records people who never opted in. A target consents; the colleague, family member, or stranger in frame does not. The framework names this risk explicitly rather than treating it as a deployment detail.

Risks the framework names

  • Bystander consent — people near the system become data subjects without agreeing
  • Third-party exposure — a target's records reveal information about others
  • Purpose drift — meeting notes quietly become behavioural prediction
  • Deletion ambiguity — erasing media may not erase the derived layers
  • Searchability — a forgotten remark becomes a queryable record
  • Power asymmetry — a memory can be interrogated by whoever holds power over the target
  • Inference leakage — goals and vulnerabilities surface through accumulation, not any single record

Beneficial uses, held in balance

  • Accessibility — support for people whose needs current tools do not meet
  • Memory support — recall assistance for those who benefit from it
  • Productivity — lower-friction capture and follow-through on one's own commitments
  • Care coordination — continuity across caregivers with consent
  • Personal knowledge management — a person's own record, under their control
  • Safety — accountability and situational support where warranted

These beneficial uses are real. Naming them does not weaken the risks above, and naming the risks does not deny the benefits — both hold at once, which is why the two are kept side by side.

No raw public first-person dataset in v1.0

Version 1.0 ships the protocol, a minimal synthetic harness, and the submission specification. It does not release a public raw first-person dataset, and no human pilot has run.

There are no empirical results, no real leaderboard, no official submissions, and no cross-domain validation in v1.0. Any illustrative rows shown elsewhere on this site are mock baselines. A frozen private split and a first consented federated pilot are planned for a later version, gated on the governance requirements above.

Benchmark validity is not deployment legitimacy

A TargetSpace score certifies one thing: calibrated prospective predictive skill on sealed, externally resolved outcomes. It measures whether a forecast beat the population prior and the target's own routine, stayed calibrated, and collapsed under target permutation. That is all it certifies.

Benchmark validity is not deployment legitimacy — a high score is no licence to act on, manipulate, surveil, or infer the inner life of a person.

A high score certifies target-specific predictive skill only. It is not evidence of understanding, inner life, or causation, and it confers no permission to act on a person. Whether a capable system may be deployed — and under what consent, oversight, and recourse — is a separate question the benchmark deliberately does not answer.