YPAI proprietary data collection technology
We built the systembehind the data.
Every modality on one platform. Every accepted asset carries the evidence that proves why it passed.
a-013 · one session · enters here speech · video · image · LiDAR
Source qualification ·
Collection conditions ·
Technical requirements ·
Consent and agreements ·
Automated quality control ·
Annotation and review ·
Permitted use ·
Dataset lineage ·
Delivery integrity ·
Withdrawal path ·
sha256 ·this tab · preview
Station 01 · Specification
Your requirement,as a control on the floor.
The brief becomes a set of controls. They run on the participant's device before a take starts and again on every upload. Pick a requirement and the floor shows where it is measured.
- 1 AUDIO-0042 Effective source resolution must be at least 23 bits.
- 2 VIDEO-0018 Audio and video within 40 ms of each other for the whole take.
- 3 IMAGE-0007 No upscaled or recompressed frames; EXIF chain intact.
- 4 LIDAR-0003 Camera and LiDAR clocks within 1 ms under PTP.
- 5 SRC-0021 Participant age 55 or over.
- 6 RGT-0004 No biometric identification.
- 7 RET-0002 + RET-0009 Retention 30 days after rejection; QC evidence kept six years.
observed · this tab 23.4bits
measurement threshold≥ 23 verdictPASS runsdevice gate · central qc
compiler finding contradiction · two retention rules govern the same rejected capture · requires customer decision before activation
Station 02 · Capture
One collection mesh.Every modality, measured at capture.
effective bits · ≥ 23 23.4bits
The measured noise floor is 9.4 dB above the project limit. Move away from the ventilation source and repeat the five-second room test.
RETRY SEGMENT · preview · measured in this tab
Station 03 · Quality
Measured four ways.Each stands alone.
Four banks read every upload. Each keeps its own reading and its own digest. A failing control stays visible beside a passing one.
Deterministic
Schema, sample rate, effective bits, frame rate, checksums, required streams, timestamps.
14 / 14 · this tab
sha256 3b7e…91a0Statistical
Coverage of the programme against its target, and agreement between annotators.
specimen programme · 60 assets · κ 0.71
sha256 c410…2e8dLearned
The transcript and the detector in this tab; the scene read by a model on the EU profile.
WER 0.06 · whisper-tiny.en · yolov10n · this tab
sha256 77a1…0c55Cross-modal
Speech against mouth motion, LiDAR against camera, timestamps against the manifest clock.
2 disagreements · 0 beyond tolerance · this tab
sha256 9c1d…44f2Station 04 · Annotation
Labelled by people.Checked by models. Both on the record.
Boxes, tracks, masks, cuboids, transcripts, entities and preferences, made in YPAI's own review console with model-assisted first passes. The model's reading is marker A, the annotator's is marker B, and the disagreement between them is the instrument. Two of the readers run in your browser.
review item · segment 3 of 12 · the box on the pedestrian
reviewer sees
asset · project instructions · relevant evidence (4) · pseudonymous reference
withheld
full identity · home address · contact details · unrelated demographics · other projects
Two annotators decided. You are the third.
consensuswaiting for your decision
agreementκ 0.71
credentialvisitor · unqualified · preview
rubricr-2026.09
decision·
Station 05 · Agreements and rights
MSA msa_v2 ACTIVE ·
Project SOW sow_v4 ACTIVE ·
DPA dpa_v3 ACTIVE ·
Processing instructions pi_v2 ACTIVE ·
Security schedule sec_v2 ACTIVE ·
Technical specification tech_v4 ACTIVE ·
Acceptance policy acc_v1 ACTIVE ·
Contributor agreement ca_v3 ACTIVE ·
Consent package consent_v3 ACTIVE ·
Privacy notice pn_v2 ACTIVE ·
Dataset licence dl_v1 ACTIVE ·
Delivery certificate dc_v1 ACTIVE ·
S1 I agree to audio recording. signed ·
S2 I agree to video recording. signed ·
S3 I agree to transcription and annotation. signed ·
S4 I agree to the stated training purpose. signed ·
statement hashes computed in this tab from the statement text · in the built page they are checked against the ledger's
Consent where it was given.Rights where they are used.
Twelve agreements and four statements are bound to the session at capture, on the contributor's own device. Every later use is a question put to those rights, answered with a decision and a reason, and any statement can be withdrawn with its impact traced.
rights runtime · ask the asset
operation
purpose
model family
recipient
ALLOW
Every permission for this operation is present, the recipient region is approved and no prohibition applies.
duties attach · retain_in_approved_region · delete_after_contract_end
rights as odrl 2.2 in a w3c vc 2.0 · signed key ypai-2026-09 · evaluated in this tab over the specimen's rights
withdrawal ·
affectedsource assets 1 · transcripts 0 · annotations 3 · dataset versions 2 · deliveries 1 · recipients 1 · training runs 1
future useBLOCKED · dataset inclusion BLOCKED
raw deletionSCHEDULED
derived dataREVIEW REQUIRED
model remediationASSESSMENT OPEN
evidence ledgerPRESERVED
e7a2…9c14·withdrawal · by you
Earlier model influence outlives the source file. Retraining, dataset reconstruction, rollback and unlearning are separate actions, each recorded on its own.
Station 06 · Record
One record.Every decision on it.
Everything the five stations above did to a-013 is on one signed record: the readings, the labels, the agreements, the decisions you made on this page. Change any of them and the digest changes with it.
record digest · now · this tab · on load ·
identity a-013 · multimodal session · v1
contract pc-2026-081 v1 · 6 / 6 requirements
origin ypai collect · phone, camera, LiDAR rig · 48 kHz / 24-bit · 1080p30
quality 14 deterministic · 3 statistical · 3 learned · 2 cross-modal · 3 human
annotation 1 box · 3 tracks · 1 mask · 1 cuboid · 2 annotators · rubric r-2026.09
agreements 12 active · 4 statements
privacy purpose collect_data · eu-west-1
rights 3 permissions · 4 prohibitions · 2 duties
integrity sha256 3b7e…91a0 · ed25519 · merkle 9c1d…
lineage 2 datasets · 1 delivery · 1 training run
governance retention 180 d · withdrawal 0 · hold none
VDA-3
Attested: collection, QC, annotation and review are signed. VDA-4 requires device attestation, which this specimen lacks.
Station 07 · Dataset and delivery
Constraints in,a lockfile out. Verified where it lands.
A dataset is a query over records. Change a constraint and the field re-lays; the lockfile pins what was included and why. The delivery carries a verifier your own team runs on its own machines.
acceptedHours
participantIsolation
dedup
languages en 0.80 · es 0.05 · fr 0.03 · pt 0.03 · ar 0.03 · hi 0.02 · zh 0.02 · ja 0.01 · ko 0.01
global: quality.verdict == PASS · rights.allow("train_asr") · agreementCoverage == COMPLETE · withdrawalStatus == ACTIVE
·
dataset.lock
·
splits participant-isolated · agreement snapshot msa_v2 · rights state 2026-09-20 · evidence root 9c1d… · croissant 1.0 · specimen programme
ypai verify delivery-2026-09-20.json
Manifest integrity PASS
Cryptographic signatures VALID · ed25519 · key ypai-2026-09
Transparency inclusion not logged · preview
Project contract v1 · coverage 6 / 6
Annotation and review coverage 1 / 1
Consent coverage 4 / 4
Dataset lineage COMPLETE
Known blocking exceptions 0
contract coverage6/ 6
public key ypai-2026-09 · MCowBQYDK2VwAyEA5Qf2…preview · verified in your browser
Station 08 · Feedback
yolov10n · 0 persons · missed · rooftop cell · recorded 2026-09-18 · /image-data/
A measured missbecomes the next brief.
intake.request 3 ms which condition did the detector miss a person in?
agent.plan 640 ms retrieve release 02 records where held ≠ found · nova 2 lite
retrieve.records 212 ms 6 hits · titan embeddings · cosine
agent.validate 910 ms cites r-11 r-14 · gap: person on a rooftop viewpoint
measure 1 765 ms eu-west-1 · 1 402 in / 96 out · preview
Collect 300 frames of people seen from rooftop and first-floor viewpoints at dusk; accept a frame when the person is fully inside it and a held box exists.
→ becomes requirement IMAGE-0011 in the next contract · station 01
Station 09 · Every modality
The modality changes.The controls governing acceptance hold.
measuredeffective bits · SNR · clipping · intelligibility · dialect · transcript alignmentmeasured on this page · this tab
plate pending · series slot Where the control model fits
Foundation and multimodal model teams
When performance depends on specific cohorts, environments, behaviours, languages or failure cases.
Speech and voice AI
When acoustic conditions, speaker profile, language, accent, recording chain and transcript quality must meet defined criteria.
Synthetic-media and authenticity teams
When genuine, manipulated and generated media must be collected or evaluated under controlled conditions with traceable provenance.
Robotics and Physical AI
When multiple sensors, timestamps, human actions, environments and task outcomes must remain aligned.
Multilingual and multi-market programmes
When sourcing, qualification, privacy, rights and delivery requirements differ across languages and jurisdictions.
Rights-sensitive AI programmes
When your team needs to know whether an asset is technically useful and whether it may be used for the intended purpose.