Regulatory Submission Drafting

Submission writing is the case where the failure mode named in the third whitepaper is most exposed. The output is prose, the reader is a regulator, and the characteristic defect of a drafting model is an unsupported claim that reads exactly like a supported one. There is no build that fails, no test that goes red. The only thing standing between a fluent invention and a regulatory filing is whether somebody traced the sentence back to a source.

That makes the verification design the whole engagement. Everything else is arrangement.

Section 1The example at a glance

Workflow	Drafting defined sections of a regulatory submission from approved source documents and study reports
Front Door disposition	FD2 Controlled Regulated Workflow.
Regulatory relationship	R1 supports regulated work. Regulatory positioning and strategy remain unassisted.
Overlays	Overlay C (US pharmaceuticals and biological products), with Overlay E (clinical development and regulatory evidence) where clinical data is drawn on.
Amplification intent	Extend regulatory writer capacity for argument construction by reducing time spent assembling and cross-referencing source content.
Configuration baseline	Model, prompt and retrieval index versions are recorded at evaluation. A change to any of them re-opens evaluation at a defined scope.
Pilot boundary	Two section types, one submission, one therapeutic area, three named writers.

Section 2Front Door and regulatory routing

Routing returned R1 without redesign, but with a boundary that had to be drawn carefully. Drafting a section from approved sources supports regulated work. Deciding what position to take with a regulator, which arguments to make and which weaknesses to address directly, is regulatory reasoning and stays unassisted. The distinction is not always obvious in the text itself, because a drafting instruction can smuggle a positioning decision into it.

The design handles this by requiring the writer to set the argument structure before any drafting occurs. The framing gate is not a formality. It is where the regulatory judgement actually happens, and if it is skipped the route is no longer R1.

Section 3Material-step map

#	Material activity	Task class	Allocation	Consequence
1	Select and confirm approved source documents and versions	Retrieve	Human-led	Tier 2
2	Define section purpose, argument structure and prohibited positions	Human judgement	Unassisted gate	Tier 3
3	Retrieve supporting content and locate relevant passages	Retrieve	Collaborative	Tier 2
4	Extract data points, results and study parameters	Extract	Collaborative	Tier 2
5	Tabulate and convert units, populations and derived values	Transform	Collaborative	Tier 3
6	Draft section narrative within the approved structure	Draft	Collaborative	Tier 3
7	Trace every claim to an approved source	Human judgement	Unassisted gate	Tier 3
8	Verify every numerical value and derivation	Human judgement	Unassisted gate	Tier 3
9	Approve section for submission assembly	Human judgement	Unassisted gate	Tier 3

Consequence tiers are assessed before crediting any control. Tier 3 means an unchallenged error could affect patient safety, product quality, release or a regulatory position.

Section 4Substantive human gates

Gate	Position	Required human contribution	Artefact
Source gate	Before retrieval	Writer selects and confirms the approved source and version set, and records where the source estate is silent on a point the section must address.	Approved source list and corpus-silence note.
Framing gate	Before drafting	Writer defines the argument structure, the positions that may be taken and those that may not. This is the regulatory reasoning step and cannot be delegated.	Argument structure and prohibited-position statement.
Claim-trace gate	After drafting	Writer decomposes the draft into individual claims and traces each to a specific approved source passage. Claims that cannot be traced are removed, not softened.	Claim-by-claim trace with source and page reference.
Numerical fidelity gate	After drafting	Writer recomputes every derived value and checks every unit, population, denominator and rounding decision against the source.	Recomputation record and correction log.
Approval gate	Before submission assembly	Writer confirms the section states the intended position, addresses known weaknesses and contains no content beyond the traced set.	Approval decision, limitations and residual-issue note.

A gate is substantive only when passing it requires the person to produce something the machine did not. Approval alone is not a gate.

Section 5Prohibited use

The configured workflow must not:

• Determine regulatory position, strategy or how a weakness is addressed.
• Produce a claim that is not traceable to an approved source.
• Generate, estimate or infer a numerical value not present in the source.
• Draft in a section type or therapeutic area outside the pilot boundary.
• Substitute a summary of a source for the source itself during tracing.

Section 6Evaluation approach

The evaluation described here produces the evidence that feeds validation for intended use. It does not replace it.

The evaluation weighted Draft and Transform heavily, because those carry the Tier 3 consequence. Draft was assessed by claim decomposition against seeded unsupported claims, contradictions between sources, and cases where the correct behaviour was to report that the source estate did not support the point. Transform was assessed by recomputation, with seeded unit errors, denominator substitutions and rounding drift.

The affordability question in Part 3 mattered more here than in most workflows. Claim tracing is slow. If the ratio of verification effort to drafting effort exceeds one, the workflow costs more than writing the section by hand, whatever the drafting time saved. That measure was treated as a pass condition rather than a metric.

The pilot boundary was written as a context of use in the sense used in the FDA draft guidance of January 2025 on AI supporting regulatory decision-making: the specific question the workflow addresses, the section types in scope, and the influence the output has on the eventual regulatory decision. Model risk follows from that influence and the consequence of the decision, which is why the same configuration would need re-assessment for a section type carrying different weight.

Section 7Illustrative evaluation results

Measure	Unassisted baseline	AI-assisted result
Evaluation set and period	18 representative sections plus 21 constructed boundary and failure cases	Same set, ten-week window
Median writer effort per section	11.5 hours unassisted	7.4 hours AI-assisted
Verification Effort Ratio	Not applicable	0.71
Seeded unsupported claims detected at claim-trace gate	Not applicable	18 of 18
Seeded numerical errors detected	Not applicable	11 of 12
Corpus-silence cases correctly reported as unsupported	Not applicable	7 of 9
Untraceable claims removed rather than reworded	Not applicable	23
Positioning decisions originating outside the framing gate	Target: none	None observed

These figures are constructed to illustrate the shape of a defensible result. They are not measured outcomes and must not be cited as evidence of performance.

Section 8Capability and authorisation

Role	Minimum level	Required capability
Writers operating the workflow	Level 2	Prepare approved inputs, operate the configured workflow, recognise stop conditions.
Gate operators	Level 3	Draft and Transform verification competencies, claim decomposition and recomputation, within a Tier 3 ceiling.
Regulatory approver	Level 4	Challenge the evidence base, restrict scope, suspend or retire the workflow.
Managers	Role duty	Protect tracing time. This is the control most likely to erode first under submission deadlines.

Section 9Disposition

The measure that decided this was the Verification Effort Ratio, not the time saving. A workflow that halves drafting time and doubles review time has moved effort rather than reduced it, and in submission work it has moved effort from a step people are good at to one they are worse at. The ratio is the only number in the results table that would have stopped the pilot on its own.