Governance research · Research

Can teams detect drift between assistant SOPs and training examples?

A research design for finding when examples teach a different rule from the approved procedure.

Headline statistic

Example drift is detectable when each example has an owner, governing rule, creation date, review date, scenario, and decision boundary.

Methodology: Research question: whether active examples remain consistent with the current SOP, authority table, and tool behavior. Inventory examples used in one workflow, map instructions to the active SOP, and classify contradictions, missing scope, stale interfaces, and unauthorized exceptions. This is a bounded OutsourcingAssistant.com operating study, not a universal benchmark, causal claim, or evaluation based on nationality.

Key stats

  • Example drift is detectable when each example has an owner, governing rule, creation date, review date, scenario, and decision boundary.
  • The eligible population and excluded states must be declared before calculation.
  • Preparation, waiting, and owner-decision time remain separate measures.

Key takeaways

  • Define the operational decision before selecting a metric.
  • Preserve stopped, returned, waiting, and approved records.
  • Change one documented control after review, then observe another period.

Frame the question and decision

The question is whether active examples remain consistent with the current SOP, authority table, and tool behavior. State the queue, period, eligible records, exclusions, and person authorized to act before collection begins.

A Philippines-based assistant may maintain observations and flag gaps. The owner decides whether cases are comparable and whether evidence supports a change.

Frame the question and decision evidence table
ItemFindingSource note
Questionwhether active examples remain consistent with the current SOP, authority table, and tool behaviorLocal study protocol
OwnerPerson authorized to change the observed workflowNIST governance context

Collect a reconstructable record

Inventory examples used in one workflow, map instructions to the active SOP, and classify contradictions, missing scope, stale interfaces, and unauthorized exceptions.

Keep observations beside source fields and prior states so a later reviewer can reconstruct the case without relying on memory.

Collect a reconstructable record evidence table
ItemFindingSource note
Minimum recordItem, state, timestamp, source, owner, reason, dispositionUK Government measurement guidance
Publication dateSeptember 1, 2026OutsourcingAssistant.com batch record

Interpret the finding narrowly

Example drift is detectable when each example has an owner, governing rule, creation date, review date, scenario, and decision boundary.

A different example is not necessarily wrong; legitimate scenario variation must be separated from a contradiction in authority or evidence.

Interpret the finding narrowly evidence table
ItemFindingSource note
Supported useImprove the defined local workflowThis study design
Unsupported useUniversal ranking or causal attributionOECD measurement limits

Run a controlled pilot

Audit the ten most-used examples, remove or relabel stale items, then observe whether the same clarification questions recur.

Preserve the baseline, changed rule, window, and result. Stop interpretation when missing cases could reverse the conclusion.

Run a controlled pilot evidence table
ItemFindingSource note
PilotAudit the ten most-used examples, remove or relabel stale items, then observe whether the same clarification questions recur.Local operating proposal
Stop rulePause when missing or incomparable cases could materially change the resultResearch governance analysis

Conclusion and limitations

The defensible conclusion is modest: example drift is detectable when each example has an owner, governing rule, creation date, review date, scenario, and decision boundary. This can guide a local adjustment but cannot isolate every influence.

The assistant can collect evidence and prepare comparisons. The accountable owner retains access, policy, personnel, customer-impact, and public-claim decisions.

Conclusion and limitations evidence table
ItemFindingSource note
ConclusionExample drift is detectable when each example has an owner, governing rule, creation date, review date, scenario, and decision boundary.Source synthesis
Primary limitationA different example is not necessarily wrong; legitimate scenario variation must be separated from a contradiction in authority or evidence.Scope analysis

Related Research

Questions people ask

Does this produce a universal benchmark?

No. It describes a defined local queue and period; comparisons require compatible populations and methods.

What can the assistant own?

The assistant can maintain observations and prepare analysis. An authorized owner decides policy, access, and consequential interpretation.

Sources

  1. 1. NIST Cybersecurity Framework 2.0Governance and ownership context.
  2. 2. UK Government Service Manual: Measuring SuccessMeasurement design.
  3. 3. OECD Measuring ProductivityDefinitions and interpretation limits.

Explore research briefing support · Review the SOP handoff checklist