# Indoor instructions: authored evaluation sample

For human collection, annotation or review, [start a project on 9jaTesters](https://www.9jatesters.com/order?pack=other). To contribute, [join the 9jaTesters workforce](https://www.9jatesters.com/signup/email?role=tester). Mention 9jaBots or 9jaRobotics and the collection you want to build. These sample downloads do not create an assignment.

Download: [indoor-language-v0.1.json](./indoor-language-v0.1.json).

This pack contains **24 authored fixtures**, with 12 tagged Nigerian English (`en-NG`) and 12 tagged Nigerian Pidgin (`pcm`). Each language has six navigation tasks, three stopping tasks and three clarification tasks. Each language also has six development fixtures and six evaluation fixtures.

These examples demonstrate a language annotation and scoring format. They are **not collected phone videos, crowdsourced demonstrations, a representative language benchmark or robot-control training data**. Pidgin wording and expected labels have not received human language review. The public evaluation split is a diagnostic challenge set; it is not a private, statistically independent held-out test.

## What each task measures

Each task describes an instruction in a fictional indoor map with four named destinations: reception, room 01, store and charger. The expected output is one of:

- `navigate`, with one known destination ID;
- `stop`, with no destination;
- `clarify`, with no destination, when the instruction leaves the goal unresolved or requests an undefined destination.

Context is part of the input. Corrections replace an earlier destination. A negated destination with a clear affirmative alternative does not make the whole instruction ambiguous. These challenge cases can expose limitations in a small command grammar. They do not measure physical motion, collision avoidance, emergency stopping, perception, speech recognition or motor control.

## Version 1.0 contract

The pack fields are `schemaVersion`, `id`, `name`, `description`, `provenance` and `tasks`. Schema version `1.0` describes the annotation format; `indoor-language-v0.1` identifies this particular fixture release.

`provenance` records declared source kind, review status, consent status and license text. Each task contains `id`, `language`, `instruction`, `context`, `expected` and `split`. Task IDs are unique. Supported destination IDs are `reception`, `room-01`, `store` and `charger`.

The importer accepts 1–500 tasks, requires text of at most 2,000 characters per text field, limits IDs to 120 letters/numbers/dots/hyphens/underscores, and rejects unknown fields. An ID must start with a letter or number. The JSON import limit is 2,500,000 characters. Stop and clarification expectations cannot contain a destination. The schema carries no robot control configuration.

Source and permissions metadata are declarations, not proof verified by the software. Human review and recording consent are separate: a reviewed annotation may still await consent. A draft created in the task tool declares a proposed human-recorded source and pending permissions; it does not attach a recording or certify that one exists. Recording, participant permission, reuse rights and any sensitive material require their own documented evidence before a real collection is shared or used.

## Rights

**Evaluation sample; contact Ranked Technologies for redistribution or training rights.**

No Creative Commons, open-data, commercial-training or robot-training license is granted by this download. A structurally valid import does not confer rights. Contact: support@ranked.ng.

## Reproduction

The application source exports this pack as `SEED_PACK` from `src/lib/task-data.ts`. The focused unit tests compare the public JSON download with that source and exercise schema validation, source declarations and export/import round trips:

```sh
npm test -- src/lib/task-data.test.ts
```

Evaluate complete instructions together with their context. Report the model or baseline, configuration, pack version, evaluated task IDs, missing outputs and mismatches. A successful language prediction is not a successful robot deployment.

## Naija mobility vision and robotics workforce pack

Download the [Naija Mobility Vision & Robotics task pack](./naija-mobility-vision-tasks-v0.1.json) or read its [assignment guide](./naija-mobility-vision-tasks-v0.1.md). It contains exactly 100 9jaTesters tasks across scene capture, vision annotation, object capture, charger observation, route scenarios and quality review. The pack is designed to create a reviewed evidence loop into 9jaRobotics and 9jaBench before hardware is available.

This is an assignment pack, not a public dataset. It defaults to private collection and grants no public, commercial-training or robot-training rights. Each real submission still needs documented consent or source licence, privacy review, provenance, worker terms and a release decision. Do not treat a charger observation as live telemetry or a route result as a safety guarantee.