Foundation test prep — the complete handbook

Everything you need to pass the Foundation test (Annotator or Collector). You can't invent what you've never learned — this guide gives you the ground truth.

The Foundation test is the OraData gateway. Two flavors — Annotator and Collector — and you can take both. Passing at least one is mandatory before paid missions.

How to use this guide

Questions test your grasp of platform rules, not rote memory. Read section by section, focusing on WHY each rule exists. If you understand the reasoning, every variant becomes obvious.

1. The Golden Rule — absolute pillar

• You can't annotate data you collected yourself.
• You can't validate an annotation you produced yourself.
• Sockpuppet account = permanent ban + wallet frozen.

2. When to signal, when to skip

The test has many “what do you do if…” variants — the right answer is almost always one of three:

1. Signal the admin — for any content issue (wrong gold answer, duplicate, disputed item, hateful content, unconsented face/plate, brief error).
2. Mark “unknown” or skip — for an item you can't handle honestly (illegible, outside your qualification, needs an expert).
3. Ask the admin for clarification — ambiguous brief, contradictory instructions, sensitive subject (drone, military zone, paid public subject).

3. Reliability score — how it works

• Minimum to stay active: 70% reliability.
• Above 95%: quality bonus on subsequent missions.
• Below threshold too long: 14-day grace period, then one-tier demotion (Experts never drop below Gold).
• Cohen's kappa minimum on an annotation batch: 0.75.
• Gold-item frequency in your workflow: 5–10% (that's how we calibrate your score).

Gold items have a known correct answer. You encounter them without knowing they're gold — that's how your score is calibrated honestly.

4. Annotation — technical rules

• Object detection (pedestrians, cars, signs) → bounding box.
• Audio with loud background noise → flag in metadata, don't ignore.
• Medical image + you're not qualified → skip.
• Annotation already validated by an arbiter expert → don't touch it.
• Identifiable faces / private names → blur + flag.
• Text labels → server-normalized, case-insensitive.
• AI pre-labeling → forbidden. You work by hand.

Annotator workflow

• Annotation rejected → fix per feedback, resubmit.
• You're 100% sure a gold item is wrong → report it (never “go along with the wrong answer”).
• Two identical images in a batch → report to admin (don't deduplicate yourself).
• Standard deadline per batch: 7 days. Can't finish → release for another annotator.
• Eye fatigue after 2h → come back tomorrow. Quality first.
• Instant rejection reason: illegal or hateful content.

5. Collection — capture technique

Format & quality

• Road / trail video → landscape orientation, never portrait.
• Gesture capture (Physical AI) → 4K minimum.
• Rare-language audio → 48 kHz min, cardioid direct mic, RAW (no filters).
• Stabilization → gimbal recommended for smartphone.
• Framing → respect the horizon, no wild high/low angles.
• Compression → ship the original, not a default-compressed version. Heavy 4K road: H.265.
• Video frame rate: 30-60 fps for gesture, 1080p minimum.

Metadata — mandatory

• GPS → mandatory on any road, flora, or agriculture capture.
• EXIF → keep, never strip.
• Flora → 5 angles per specimen (leaf, flower, fruit, bark, whole plant). Scale: 10 cm ruler.
• Filename → OraData convention, never freeform.
• Languages → ISO 639-3 code, not 639-1.

Collector workflow

• Same file on two different missions → NEVER.
• Blurry shot found after the fact → redo, don't “fix in post”.
• Client brief unclear → ask admin for rewording.
• Partial delivery (50%) → allowed ONLY with written justification.
• Standard batch delivery: within 48h.
• Color + white calibration → mandatory before every session.
• Source file retention: 90 days post-delivery. Never delete immediately, never publish.
• Reselling batch material online → FORBIDDEN.

6. Consent & privacy

• Speaker for a language dataset → written + signed consent.
• Child → written parental permission required.
• Passersby on a street photo → blur if no consent.
• Public / paid subject (actor, athlete) → forbidden without agreement.
• Minimum speaker age: 18, or any age with written guardian consent.
• Speaker requests anonymity → refuse to publish their identity, but document with pseudonym + written consent.

7. Safety — the absolute priority

• Forbidden or unsafe place → refuse.
• Slippery terrain → safety first, you finish later or not.
• Rain / crowd / unexpected disruption → stop and note in metadata.
• Vehicle breakdown → stop, secure, don't film.
• Drone → flight permit + authorized zone mandatory. Never “open”.

8. Quality & priorities

• OraData values precision > speed. Always.
• Rushing a mission for max profit → NO, quality first.
• Document to re-read quarterly → the OraData Ethics policy (and this guide).
• OraData tracks average time per item to adjust pricing — that's normal.

9. Payment — what counts

• Annotator → paid per validated batch unit.
• Language collector → per minute of validated audio.
• Flora collector → per validated specimen.
• Agriculture collector → per validated plot.
• Urgent after-hours medical review → extra fee.
• External resale of material → forbidden without OraData agreement.

10. The qualification test itself

• Minimum score: 80%.
• Failure cooldown: 12 hours (never 14 days, never 24h).
• Pool: 30 questions, 20 drawn at random per attempt.
• Underperforming annotator sanction: 14-day grace period, then one-tier demotion. No immediate ban (unless fraud).

The “nevers” to burn in

• Never validate or annotate your own work.
• Never publish without written consent.
• Never guess — skip or flag.
• Never deliver a duplicate.
• Never pre-label with an AI tool.
• Never modify an arbiter-validated annotation.
• Never upload without GPS on road / flora / agriculture.
• Never compress by default — ship the original.
• Never trade safety for a mission.

OraData · guide public · révisé 2026

Photo de couverture : Green Chameleon · Unsplash