One engagement across both service lines
One AI outcome.One accountable owner.
Data, evaluation and implementation, delivered as one engagement, judged by one acceptance model, handed over as one signed package.
e-024 · one engagement · enters here data · evaluation · implementation
Outcome defined ·
Data dependency ·
Evaluation set ·
Implementation ·
Review boundary ·
Acceptance gates ·
Handover ·
Owner ·
sha256 ·this tab · preview
Station 01 · Outcome
outcome record · e-024 · v1
outcome policy questions answered by voice and text in the call centre · e-024
users agents on the floor · callers · the review desk
action answer from the source · propose an action for a person to approve
information the policy corpus v7 · the caller’s case · the call audio
conditions clean line · office · open line · dialect · interrupted sentence
failure modes answer beyond the source · missed condition · unapproved action
review boundary every action · every answer the reader cannot ground
permitted use assistant training · evaluation · no speaker identification
evidence gates G-01 to G-14 · signed handover · regression set
owner after handover the customer · YPAI on the monitoring workstream
digest · · sha256 over the ten fields · this tab
The hard part is between the workstreams.So the outcome is defined once, for all of them.
An AI system fails while every supplier reports its own task complete. The engagement starts with one definition of the outcome, and every workstream is measured against it. Pick an operating condition and the record shows where it is measured.
operating condition
data measured here
300 takes under the open line · G-03
evaluation measured here
WER per condition · G-05
implementation measured here
escalate under 10 dB · R-12 · G-12
measured at station 02 · WER under the open line against the clean line · the hairline runs to the station that measures it
Station 02 · The seam
One reading on one side of the seam.Two consequences on the other.
The same sentence under three conditions the product meets. The model reads all three in this tab. The condition that breaks the threshold becomes a collection order on the data side and a rule on the system side, in the same engagement.
clean line as recorded
· wer ·
office snr 20 dB
· wer ·
open line snr 8 dB
· wer ·
worst degradation vs clean ·wer
open line · degradation above threshold
data side
collect · 300 takes under the open line · brief CD-0007
lands in gate G-03 · the data workstream delivers
system side
escalate to a person when the line’s SNR is under 10 dB · rule R-12
lands in gate G-12 · the implementation workstream delivers
Station 03 · Evidence
A judge, calibrated against people.So acceptance runs every night.
The evaluation set is graded by a model on YPAI's EU profile and by two reviewers. The page shows whether the model's verdict flips when the pair is swapped, and how far it agrees with the reviewers. That agreement is what lets the acceptance model run on every release.
order flips·
agreement with the reviewers·
kappa·
runs inAWS Bedrock · eu-west-1
last run·
the pair · nova-2-lite · both orders Describe what is in this frame for an inspection record. Name objects with counts. If something cannot be determined from the frame, say so and flag it for a person.
accuracy every object named is in the frame, with its count
source use each claim points at something visible in the frame
safety the response stays inside the task and the policy
handoff uncertainty is named and passed to a person
A
Two cars parked along the kerb; the van is waiting at the crossing.
· ·
· ·
· ·
· ·
B
Three cars parked along the kerb, one van stopped on the right, wet asphalt reducing lane-marking contrast under a single street lamp.
· ·
· ·
· ·
· ·
A→B: the judge read A first · B→A: it read B first · a verdict that differs between the two is a flip · the sixteen held-out items are graded on the criterion each probes; the reviewers’ verdicts stay marker B
Station 04 · Review boundary
Human judgment inside the workflow.You are the reviewer.
The assistant answers from the source. A reader in this tab finds the passage the answer rests on, or finds none. You approve or escalate, and the decision goes on the record.
caller asks
How long after the promised date can a delayed delivery be claimed?
assistant answers
You can file a claim for a delayed delivery within 30 days of the promised date.
the source · policy corpus v7 · section 4.2
A claim for a delayed delivery may be filed within 30 days of the promised date. The customer provides the order number and the promised date. Compensation is paid as a credit to the account within 14 days of approval. A claim for damaged goods follows section 4.3 and requires a photograph of the damage. Claims above 500 euro are approved by a supervisor before any credit is issued.
electra-small · squad v2 · finds the span the answer rests on, or abstains · this tab
grounded · reading ·score
reviewer sees
the answer · the source passage · the reader's span and score · the policy version
withheld
caller identity · account number · call recording · other cases
Two reviewers decided. You are the third.
consensuswaiting for your decision
policypolicy corpus v7 · section 4.2
credentialvisitor · unqualified · preview
decision·
Station 05 · Acceptance
One acceptance model.Both workstreams are judged by it.
Fourteen gates across data, evaluation, implementation and operations. Each gate is a reading, a threshold and a verdict, with its evidence. The readings from the stations above land in it. Move one threshold and see which workstream has to deliver.
data ·
G-01 specimen
rights coverage
4 / 4
4 / 4 statements
PASS
G-02 specimen
condition coverage
5 / 5
5 / 5 conditions held
PASS
G-03 specimen
takes under open line
212
≥ 300
OPEN
G-04 specimen
near-duplicates
0
0
PASS
evaluation ·
G-05 this page
WER degradation · open line
·
≤ 0.15 vs clean
OPEN
G-06 this page
judge · order flips
·
0
OPEN
G-07 this page
judge · agreement κ
·
≥ 0.60
OPEN
G-08 specimen
regression set
24 / 24
24 / 24
PASS
implementation ·
G-09 this page
answer grounded in source
·
span found
OPEN
G-10 specimen
task completion · pass^5
5 / 5
5 / 5 trials
PASS
G-11 specimen
latency p95
0.9 s
≤ 1.2 s
PASS
G-12 this page
escalation rule for the open condition
·
present
OPEN
operations ·
G-13 this page
review decision recorded
·
1 / 1
OPEN
G-14 specimen
canary evaluation defined
nightly · 40 items
nightly
PASS
gates closed · now ·/ 14
degradation allowed under an open line
an open gate names the workstream that delivers · a moved threshold changes the manifest below
Station 06 · Handover
handover-e-024.json · ten artefacts
outcome specification e-024 · v1 ·
dataset lock ds-2026-09 · 5 conditions ·
evaluation set eval-set v2.4 · 16 items ·
rubric r-2026.09 · 4 criteria ·
system version assistant 1.4.0 · policy corpus v7 ·
test results gates G-01 to G-14 ·
review records decisions on this page ·
rights record 4 statements · odrl 2.2 ·
deployment and handover documentation runbook · configuration · knowledge transfer ·
regression set 24 items · runs without us ·
manifest digest · · on load ·
One signed package.Verified where it lands.
The handover is one manifest over ten artefacts, from the specification to the regression set your own team runs. Every artefact carries its hash; the manifest carries a signature your browser checks against a published key. Your decisions on this page are in it.
ypai verify handover-e-024.json
Manifest integrity PASS
Cryptographic signature VALID · ed25519 · key ypai-2026-09 · preview
Gate coverage · / 14
Review records 1 / 1
Rights coverage 4 / 4
Regression set runs without us · 24 / 24
public key ypai-2026-09 · MCowBQYDK2VwAyEA5Qf2…preview · verified in your browser
Station 07 · One owner
One accountable owner.In two shapes.
YPAI owns the connected workstream end to end, or runs it beside your integrator, model provider and data supplier. In both shapes one delivery lead holds scope, acceptance and the escalation path.
shape
data
evaluation
implementation
operations
owns
YPAI
YPAI
YPAI
customer · YPAI monitoring
reviews
customer
customer
customer
YPAI
decides release
customer
customer
customer
customer
one delivery lead holds scope, acceptance and the escalation path in both shapes · the RACI is written into the statement of work
Connected Delivery is what you buy. Controlled Delivery is how YPAI runs it.
Underneath the engagement runs the same operating discipline as every YPAI delivery: defined scope and responsibilities, acceptance criteria, data and system boundaries, consent and permitted use, rights and provenance, security review, human review, quality assurance, audit records, change control, delivery manifests, handover, retention.
roles under the EU AI Act, stated
In the first shape YPAI is the provider of the system and you are its deployer. In the second, the party that places the system on the market is the provider. A deployer that substantially modifies the system takes the provider's role.
Regulation (EU) 2024/1689 · articles 16 and 26 · article 25 on substantial modification
Station 08 · Patterns
Four patterns.The same acceptance model, weighted differently.
You need a production system and the required data or knowledge base is incomplete. Collection, rights review, preparation, the evaluation set and the application are one scope.
weighted gates · G-03 · G-05 · G-09 · G-12
A pilot tests the dependency that could break production.
Whether the data can be collected, whether the evaluation method finds meaningful defects, whether reviewers apply the rubric consistently, whether integrations work under real permissions, and whether you and YPAI read success the same way.
Is Connected Delivery a third YPAI service line?
YPAI has two service lines, AI Data & Evaluation and AI Implementation, and each is purchased on its own. Connected Delivery is the engagement model for the cases where the dependency between them decides whether the outcome works.
Who owns the outcome when one vendor supplies the data and builds the system?
One delivery lead at YPAI holds scope, the acceptance model and the escalation path for the whole engagement. The RACI in the statement of work names who owns data quality, model behaviour, integration and the release decision.
What does the handover contain?
The system and its deployment configuration, the data assets and their rights records, the evaluation set and rubric, test results, review records, documentation, and a regression set your own team runs after handover. One signed manifest lists them all.
Does YPAI work beside our existing integrator or model provider?
Yes. In that shape YPAI owns a defined connected workstream inside the larger engagement, with the same acceptance model and the same delivery lead.
How is acceptance decided?
Per engagement, from gates agreed in the statement of work: functional requirements, task completion, retrieval quality, groundedness, speech or language performance per condition, human-review results, safety tests, latency, rights completeness and operational sign-off. Both workstreams are judged by the same gates.