Why this reading
The harder problem is no longer whether AI can write code. It is whether you can defend what it produced.
The newly launched nPhase statistical-computing architecture separates AI proposal, human approval and deterministic execution, while emphasizing version pinning, snapshots, provenance and validation.
ClinAgent reaches a compatible design from the research side: keep data-access tools thin, place clinical-programming expertise in richer skills, and validate each layer independently.
Reading order
Your 30-minute plan.
Find the difference between proposal, approval and execution.
Read governance, version pinning, snapshots, provenance and exception-based review.
Compare thin tools, thick skills, rule engines and agent-independent validation.
Connect the architecture to machine-readable, standards-based clinical workflows.
Redesign one SP task so its output remains defensible six months later.
Open-access sources
Fresh production architecture plus peer-reviewed clinical-programming design.
Brief background
Generation is only one layer of a regulated workflow.
The nPhase architecture separates three roles: AI proposes transformations, humans approve them, and deterministic systems execute them. It also describes immutable snapshots, pinned approved versions, provenance, SDTM auto-mapping, ADaM/TFL automation and integrated validation.
The important architectural lesson is independent of the vendor: a probabilistic model should not silently become the source of truth, execution engine, standards repository and approver at the same time.
ClinAgent's “Thin MCP, Thick Skills” approach makes a similar separation. Stateless tools expose data and files; skills carry prompts, clinical-programming rules, decision logic and deterministic checks.
A defensible TFL therefore needs more than a correct number. A reviewer should be able to recover the data snapshot, ADaM version, analysis definition, code or rule version, environment, validation evidence and approval record that produced it.
Read performance numbers carefully: nPhase's speed and accuracy figures are vendor-reported claims. They are useful for understanding the product's intended value proposition, but they are not independent comparative validation.
Key vocabulary
Fifteen terms for governed AI statistical programming.
| Term | 中文 | Meaning / use |
|---|---|---|
| defensible output | 可辩护的输出 | An output whose inputs, rules, execution, evidence, and approvals can be reconstructed and justified. |
| computational provenance | 计算谱系 / 计算来源 | A record of the data, code, environment, versions, and transformations that produced a result. |
| immutable snapshot | 不可变快照 | A preserved version of data, code, metadata, or environment state that cannot silently change after approval. |
| governed pipeline | 受治理的流程 | A workflow with controlled permissions, approved versions, validation rules, and auditable transitions. |
| human approval | 人工批准 | An explicit human decision that accepts, rejects, or redirects an AI proposal before controlled execution. |
| deterministic execution | 确定性执行 | Execution in which the same controlled inputs and rules produce the same result. |
| auto-mapping | 自动映射 | Automatic proposal of relationships between source data and target standard variables or domains. |
| auto-remediation | 自动修复 | A system-generated correction for a detected standards or validation issue, subject to defined controls. |
| version pinning | 版本锁定 | Fixing a workflow to an exact approved version of metadata, code, library, model, or dependency. |
| audit trail | 审计追踪 | A chronological record of actions, changes, approvals, and system events. |
| exception-based review | 基于异常的审阅 | A review model in which humans focus on uncertain or failed cases instead of rechecking every routine item. |
| domain logic | 领域逻辑 | Professional rules specific to clinical programming, CDISC, statistics, sponsor conventions, and study design. |
| stateless tool | 无状态工具 | A narrow tool that reads or transforms inputs without independently carrying workflow-specific professional judgment. |
| compliance infrastructure | 合规基础设施 | Controls such as logging, masking, access control, validation, and audit support around an automated workflow. |
| agent lock-in | 智能体锁定 | Dependence on one specific model or agent host such that the workflow cannot be moved or independently validated. |
Useful phrases
Language for governance, validation and production-readiness discussions.
- the central question is whether the output can be defended - The central question is whether the output can be defended, not merely generated.
- AI should propose while controlled systems execute - AI should propose while controlled systems execute.
- every approved version should be reconstructable - Every approved version should be reconstructable from its data, code, metadata, and environment.
- human review should concentrate on exceptions and judgment - Human review should concentrate on exceptions and judgment rather than repetitive reconstruction.
- domain logic should be separated from data access - Domain logic should be separated from data access.
- the same input should reproduce the same approved output - The same controlled input should reproduce the same approved output.
- a model suggestion is not the same as an executed transformation - A model suggestion is not the same as an executed transformation.
- automation must preserve provenance across handoffs - Automation must preserve provenance across handoffs.
- vendor efficiency claims require independent validation - Vendor efficiency claims require independent validation before they become evidence.
- production readiness depends on governance as much as capability - Production readiness depends on governance as much as capability.
Comprehension
Five questions.
- Why is “AI proposes, humans approve, deterministic systems execute” safer than one agent performing all three roles?
- What information belongs in the computational provenance of a final TFL?
- How does “Thin MCP, Thick Skills” separate data access from clinical-programming expertise?
- Why should vendor-reported efficiency gains be treated differently from independent evidence?
- Which SP activities are suitable for exception-based review, and which decisions should remain explicitly human-led?
Retelling
Say it three times.
- 30 seconds · Generation versus defensibility.
- 45 seconds · Source → AI proposal → human approval → deterministic execution → validation → provenance → release.
- 60 seconds · Divide responsibility among the LLM, skill, rule engine, SCE and programmer.
5-minute output task
Make one SP workflow defensible six months later.
- Minute 1: Choose SDTM mapping, ADaM derivation, SAS log QC, TFL programming, TFL QC or eSubmission.
- Minutes 2-3: Assign responsibilities to AI, skill/domain logic, deterministic execution, validation and human review.
- Minute 4: Define the exact provenance package that must be frozen or logged.
- Minute 5: Give a production go/no-go decision.
One sentence to keep
A production-grade SP agent should not merely create the right output once; it should produce a result whose source data, rules, execution, validation, versions, and approvals can all be reconstructed and defended.