Standard operating procedure · RD5A · Research & Development

Model Development

Transactional procedure owned by Data Science & Modeling in Research & Development. 8 numbered steps, 2 of them a person deciding and 6 of them agent work … more

Research & Development manual Service view

Procedure code

RD5A

Transactional

Rule versions in force

v2.7 · v2.3 · v5.0 · v1.7

4 governing rules stacked over it

In force since

2025-02-27

Earliest effective date across those rules

Last reviewed

2026-08-14

Latest review date across those rules

Rules approved by

Director, Data Science

Rule health

3/4

0 under review · 1 superseded

There is no document register behind this header. The version, the dates and the approver names are read off the governing rules this procedure runs inside, which is the closest thing the estate holds to a document control record. Nobody wrote these procedures. There is no document library behind this page and no Word file anywhere in the estate. Every procedure below is generated from rows this platform already holds: the line and its ordered stations, the owning team and the person accountable for it, the policy items bound to that team, and the governance gates seeded against the function. The steps, the owners, the gates, the approvers, the dwell times and the volumes are read straight off those rows. The sentences that join them together are generated. Read a procedure as the estate describing itself, not as a document somebody signed.

Purpose and scope

What this procedure is for and where it stops

To run model development to one repeatable standard inside Data Science & Modeling, part of Research & Development. more

Entry condition

Runs transactional. Starts when a scheduled window opens.

Exit condition

The unit clears deploy, released by Director, Data Science. more

In scope

Every unit of model development raised inside Data Science & Modeling, from the trigger below through to the final step.
6 of the 8 steps, which run without a person present.
2 decisions that stop the line until a named person releases it.

Out of scope

Anything outside the envelope set by the owning team. It leaves the line at a gate and becomes the named approver's decision.
Work that belongs to another function. This line records no handoff, so nothing leaves it mid-flow.
The 4 steps carried on the shared AI GBS spine. They are done for this line but not by this function.

Who owns it

Runs the line

Modeling Orchestrator

Accountable person

Director, Data Science

The end of the escalation path. Above them sits R&D Chief of Staff for the function.

That team’s remit

Owns model work end to end inside Research & Development. more

That remit describes the whole team, not this one procedure, so its counts are larger than the ones on this page.

Autonomy

75% now · ceiling 83%

Training, validation and drift detection are fully agentic. more

The procedure, step by step

8 numbered steps · 2 of them stop for a person

A step is one station on the line. more

No.	Step	Who does it	What happens	System of record	Time on the step	In flight
01	Frame problem Agentworkflow	Feature Engineering Agent	Feature Engineering Agent performs this step unattended at the workflow layer, inside the envelope its owning team has set, and writes an evidence record naming the actor, the authority it acted under and the digests of what it read and wrote.	Product Lifecycle Management	44 minstandard 1.0 hrs	0
02	Engineer features GBSgbs	Feature Engineering Agent	Feature Engineering Agent performs this step unattended on the shared AI GBS transaction spine, outside the function, inside the envelope its owning team has set, and writes an evidence record naming the actor, the authority it acted under and the digests of what it read and wrote.	Product Lifecycle Management	2.5 hrsstandard 4.5 hrs	31
03	Train GBSgbs	Feature Engineering Agent	Feature Engineering Agent performs this step unattended on the shared AI GBS transaction spine, outside the function, inside the envelope its owning team has set, and writes an evidence record naming the actor, the authority it acted under and the digests of what it read and wrote.	Product Lifecycle Management	1.5 hrsstandard 3.5 hrs	13
04	Validate GBSgbs	Feature Engineering Agent	Feature Engineering Agent performs this step unattended on the shared AI GBS transaction spine, outside the function, inside the envelope its owning team has set, and writes an evidence record naming the actor, the authority it acted under and the digests of what it read and wrote.	Product Lifecycle Management	3.0 hrsstandard 4.0 hrs	27
05	Explainability pack Agentworkflow	Explainability Agent	Explainability Agent performs this step unattended at the workflow layer, inside the envelope its owning team has set, and writes an evidence record naming the actor, the authority it acted under and the digests of what it read and wrote.	Research Notebook and Experiment Data	1.4 hrsstandard 2.5 hrs	4
06	Reviewer approve HumanpolicyDecision	Director, Data Science Releases: Director, Data Science	Work stops. Director, Data Science reviews what the line has produced and either releases the unit or sends it back. Nothing downstream of this step runs until that decision is made.Holds until: Held for a named human when the unit exceeds the model decision envelope, when confidence falls below the floor, or when a policy clause is engaged.	Product Lifecycle Management	1.4 hrsstandard 2.4 hrs	9
07	Register GBSgbs	Feature Engineering Agent	Feature Engineering Agent performs this step unattended on the shared AI GBS transaction spine, outside the function, inside the envelope its owning team has set, and writes an evidence record naming the actor, the authority it acted under and the digests of what it read and wrote.	Product Lifecycle Management	2.2 hrsstandard 3.0 hrs	2
08	Deploy HumanpolicyDecision	Director, Data Science Releases: Director, Data Science	Work stops. Director, Data Science reviews what the line has produced and either releases the unit or sends it back. Nothing downstream of this step runs until that decision is made.Holds until: Held for a named human when the unit exceeds the model decision envelope, when confidence falls below the floor, or when a policy clause is engaged.	Product Lifecycle Management	3.0 hrsstandard 6.8 hrs	8

The swimlane

The same steps, laid out by who is holding the work

Three lanes: a person deciding, an agent working, and the shared services spine. more

Person2 stations and 28% of the dwellAgent2 stations and 14% of the dwellShared spine4 stations and 59% of the dwell5 crossings between lanes

Human

decides

Agent

does the work

GBS

shared spine

System

attributed by rule

Product Lifecycle Management

Research Notebook and Experiment Data

Product Lifecycle Management

Click any station to see who does it, what it costs in time, the rule that gates it and the system attributed to it.

Run the procedure

One unit walked along that swimlane on the happy viable agentified path

Press run and a single unit enters at station 01. Each station takes exactly the dwell this platform records for it and is held by whoever the estate says holds it. more

speed

worked clock

00:00

allowance clock

00:00

stations cleared

0/8

the happy path, worked time0 min of 16 hrs

the same run against its allowance0 min of 28 hrs

Human

decides · 2

Agent

does the work · 2

GBS

shared spine · 4

run log

Nothing has happened yet. more

where the worked time goes16 hrs a unit

Person4.4 hrs · 28%Agent2.1 hrs · 14%Shared spine9.2 hrs · 59%5 crossings between lanes

what makes this path viable

The lanes are the recorded lanes. 2 of the 8 stations stop for a person and every one of them still does on this run.
Every one of the 2 decisions is still taken by the people who hold them. The happy path is that each one is released first time, not that it is skipped.
The owning team is held to a 83% autonomy ceiling and this run sits at 72% unattended, inside it.
2 systems are attributed across the stations, and 0 of them are connected. Nothing in this run has been watched executing against a real system of record.

what would have to change

1 of the 2 decisions can be handed over inside the 83% ceiling. more

could move to the agent lane

06 Reviewer approve — held by Director, Data Science, 1.4 hrs a unit

stays with a person

08 Deploy — The last decision on the line. Something on every procedure ends with a person, so this one is never offered.

what this run is, and what it is not

This is a run, not a forecast. more

Happy means nothing goes wrong: no station raises an exception, no gate sends the unit back, no rework loop fires and nothing sits in a queue between stations. more

The second clock is the allowance. more

A decision may be handed to the agent lane only if two things hold … more

The run reads the station rows only. more

What this run does not model: arrival patterns, contention between procedures for the same person, shift patterns, weekends, learning curves, and the cost of anything. more

Run it again, going wrong

The same unit, the same stations, broken the way the estate already records breaking

This is the same unit on the same stations, breaking. more

No break below was chosen. more

One play head, two tracks, the same scale. The clean run stops when it runs out of stations; the other one keeps going.

As written

8 stations, nothing waiting, nothing sent back

0 min of 16 hrs

now Frame problem · Feature Engineering Agent · 44 min

As the ledger has it

the same stations, plus 6 categories the estate already counts

0 min of 2.1 days

now Frame problem · Feature Engineering Agent · 44 min

personagentshared spinework the break addsnobody working it

What breaks, and where it is counted

No break below was chosen. more

What the two runs say together

6 counted categories land on this procedure. more

What this run cannot claim

These breaks are shown together on one unit so the shape of a bad day fits on one screen. more

The separation register records a duty conflict against a sub-function and a named post, not against one procedure. more

Two clocks run. Worked time is the sum of the time the work is actually being done … more

Roles and responsibilities

4 named parties

The responsibility grid is assigned by a rule, not lifted from a chart. more

Party	RACI	Steps	Duty	Escalates to
Modeling Orchestrator orchestrator	Runs the line	0	Sequences this line, holds its timetable and decides what has to surface to a person. more	Raises to Director, Data Science when a unit falls outside the envelope or confidence drops below the floor.
Feature Engineering Agent agent A3% autonomy · ceiling A3% 98% clean · 5% overridden	Responsible	5	Runs the repeatable portion of data science & modeling inside the shared transaction spine at spine cost and spine controls.	Escalates to Director, Data Science when the unit falls outside the decision envelope or confidence drops below the floor.
Explainability Agent agent A3% autonomy · ceiling A3% 92% clean · 4% overridden	Responsible	1	Executes the data science & modeling step it owns, posts its evidence, and hands the unit to the next stage.	Escalates to Director, Data Science when the unit falls outside the decision envelope or confidence drops below the floor.
Director, Data Science person	Accountable and responsible	2	Holds 2 decisions on this line. more	Accountable to Director, Data Science for the decisions held here.

Decision points and escalation

2 places this procedure stops

06 · Reviewer approve

1.4 hrs

Held for a named human when the unit exceeds the model decision envelope, when confidence falls below the floor, or when a policy clause is engaged.

Director, Data Science

08 · Deploy

3.0 hrs

Held for a named human when the unit exceeds the model decision envelope, when confidence falls below the floor, or when a policy clause is engaged.

Director, Data Science

Handoffs out of this procedure

Nothing leaves this procedure for another function. It completes inside Data Science & Modeling.

The rules this procedure runs inside

4 governing rules

Data Science & Modeling Decision Envelope

active

RD5-POL-200 · envelope · vv2.7

Agents in this lane may act without a human when the model sits inside the stated value, risk and confidence envelope. more

Approved by Director, Data Science · in force 2026-06-19 · reviewed 2026-04-30

Binds Feature Engineering Agent, Model Training Agent, Validation Agent, Explainability Agent

EU · UK · US

Data Science & Modeling Evidence Standard

superseded

RD5-POL-201 · evidence · vv2.3

Every autonomous decision writes inputs, the rule version applied, the model and prompt version, the output and a reversal path. more

Approved by Director, Data Science · in force 2025-10-11 · reviewed 2026-06-25

Binds Feature Engineering Agent, Model Training Agent, Validation Agent, Explainability Agent

US · Canada

Data Science & Modeling Escalation Rule

active

RD5-POL-202 · escalation · vv5.0

Escalation is mandatory when confidence falls below the floor, when two options sit inside the confidence band, or when a break condition on a ratified position fires. more

Approved by Director, Data Science · in force 2025-02-27 · reviewed 2026-08-14

Binds Feature Engineering Agent, Model Training Agent, Validation Agent, Explainability Agent

US · Canada

Data Science & Modeling Autonomy Ceiling

active

RD5-POL-203 · ceiling · vv1.7

The ceiling for this lane is 83 on the platform scale. more

Approved by Director, Data Science · in force 2026-04-21 · reviewed 2026-06-29

Binds Feature Engineering Agent, Model Training Agent, Validation Agent, Explainability Agent

LATAM · US

Systems this procedure touches

2 named · 0 connected

System attribution is a rule, not a reading. more

System	Vendor	Domain	Criticality	Steps	Connection
Product Lifecycle Management	Siemens Teamcenter	engineering	high	7	modeled
Research Notebook and Experiment Data	Benchling	engineering	medium	1	not-connected

0 of the 2 systems on this procedure is connected to the platform, and none of the twenty-six in the estate register is. 0 of them are marked as taking a write back, which is a property of the system, not a live path from here.

How it performs

End to end

15.7 hrs

Summed from the time each step sits, standard 27.7 hrs — running at 57% of the standard.

Touchless

76%

Agent share

75%

A week

2,326

In flight

Effort a week

262 human minutes · 677 agent minutes

Volume multiplied by the time on each step. more

Assurance over this procedure

4 governance gates sit above it

The twenty-four controls and eighteen exception classes in this platform are held at estate level … more

Gate	What trips it	Who releases it	Clock	Waivable
Physical experiment authorization approval	Any laboratory experiment involving materials, equipment or subjects	Principal investigator and safety officer	2 business days	No
External data acquisition approval	Any new external dataset, license or third-party evidence source	Research governance and Legal	3 business days	No
Publication or external disclosure approval	Any material leaving the research boundary, including conference abstracts	CSO and General Counsel	5 business days	No
Reproduction sampling oversight	Independent re-derivation of 15% of reported internal results	Reproducibility board	Monthly	No

Next to it on the same team

Data Science & Modeling

Drift Response

RD5B

Model Retirement

RD5C

All of Research & Development

What this document is and is not

These procedures describe the modeled estate. more

Actions

What is waiting on a person

What one procedure document gives you, and what to check before you treat it as the standard.

151

named approvers appear across the library

Every decision point on every procedure carries a job title, not a queue. more

143

steps in the library hand their unit to another function

A handoff is where a procedure stops being one team’s problem. more

controls exist in the estate, and they belong to no procedure in particular

Controls and exception classes are held estate-wide. more

Operations

What this desk is allowed to start

A surface that only reports is not operable. more

Trigger and bound

A procedure document prints purpose, scope, entry and exit, the numbered steps with the actor on each, the roles and their RACI position, the swimlane, the decision points and escalation, the handoffs, the governing rules with version and approver, the systems by rule and the timings. more

Live observability

What the record shows right now

The whole library, by lane, so a single document can be read against the estate it came from. more

Current distribution

2,748 steps

Agent, inside the function1,38951%

Agent, on the shared spine85131%

Person deciding at a gate50818%

Is policy and strategy coming to fruition

Whether the written intent is holding here

Half. The document is complete and every actor on it is measured more

Not holding on the record

A procedure here carries a version, an approver, an in-force date and a last-review date. more