Field note 17

Decide when a correction should change an AI workflow

A localized matte red filler patch with shallow trowel marks sits within a broad continuous off-white plaster surface.
A correction is not a rule

In an illustrative print-service case, a reviewer sees "A4" in a job summary and changes it to "A3." The request really did specify A3, so the correction is justified. It does not follow that every future poster should default to A3.

For a process owner reviewing AI-assisted preparation, those are different decisions. One repairs a draft. The other changes what future requests will mean. Use the correction-scope worksheet below to keep them separate without losing useful feedback.

What the evidence supports

Microsoft Research's human-AI interaction guidelines recommend making errors easy to correct, adapting cautiously, accepting granular feedback and conveying how user actions affect future behavior.[1] NIST's AI RMF Playbook describes incorporating adjudicated feedback into system design and implementation, and includes outlining change-management requirements among its suggested actions.[2]

Our interpretation: record feedback promptly, but decide its scope before treating it as reusable instruction. The four dispositions below are a proposed MAJLS framework, not a regulatory taxonomy or a claim that feedback automatically trains a MAJLS agent. This is an operating design recommendation, not evidence of a deployed feature.

The correction-scope framework

Classify each comment by what it asks to change:

One comment can need two linked entries. "Fix this to A3 and always use A3 for posters" contains a local correction and a proposed default. Closing the first must not close the second.

Worksheet template: capture the scope before the edit

Use one row per feedback item. Keep comments and source references accessible only to authorized reviewers; avoid copying confidential feedback into broadly shared instructions.

Field What to record
Case and draft Case ID, draft version and exact disputed text
Feedback Verbatim correction, reviewer role and receipt record
Evidence Source ID/version, relevant value and any disagreement
Current-case impact Which sentence or field may change now, under which permission
Disposition Case fix, source request, rule proposal or unresolved preference
Human owner Who can decide this scope; a reviewer is not automatically that owner
State and receipt Feedback received, edit verified, source request pending, or proposal pending; link the actual decision separately

The worksheet's purpose is to make an instruction inspectable. A thumbs-down alone does not supply a replacement value, scope or authority.

Filled example: a fictional print-service register

Everything in this example is illustrative and synthetic, including records, permissions and outcomes. A preparation role drafts summaries for a human print coordinator. It cannot book, purchase, print or dispatch work. The assumed permission permits correcting a draft against a verified intake request, but not changing that request or the standing workflow.

Item / draft Reviewer comment and evidence Disposition / current-case treatment Owner and recorded state
F1 / PS-17 v1 Print coordinator: "Change A4 to A3." Intake PS-17 r1 explicitly says A3; draft says A4. Case fix. PS-17 v2 says A3; human coordinator compares it with r1. Other drafts unchanged. Print coordinator: case fix verified. Original v1 retained.
F2 / PS-17 v1 Same reviewer: "Make all posters A3." One A3 request is the only basis supplied. Rule proposal. No default added; linked to F1 without inheriting its closure. Print-service process owner: proposal pending, not approved.
F3 / PS-18 v1 Coordinator: "The tracker size is wrong." Intake r1 says A4; tracker r2 says A3. Source request. Preserve both values; affected summary stays disputed until authoritative record and correction are confirmed. Intake-record owner: request received, resolution pending.
F4 / PS-19 v1 Coordinator: "Use friendlier wording." No standing style instruction identified. Preference clarification. Record the requested phrase and intended audience before editing. Process owner: scope pending.
F5 / PS-19 v1 Second reviewer: "Keep it strictly factual." Conflicts with F4's requested wording. Unresolved preference. Keep both original comments; no last-comment-wins rule. Process owner: joint decision pending.

F1 is closed only for PS-17. F3 does not authorize a tracker update, and a later source correction would not itself approve a workflow change.

Rule-proposal template: test what would change

For F2, reject the unsupported blanket default as submitted. Offer this narrower proposal for a separate decision:

Proposal field Filled synthetic proposal P-01
Exact predicate Intake contains one explicit supported paper size and no conflicting source value.
Future behavior Copy that size into the summary; do not infer size from "poster."
Exceptions Missing, unsupported or conflicting sizes stay unresolved for the coordinator.
Basis / affected cases Intake-authority convention proposed by the process owner; size extraction in print summaries only. No policy approval yet.
Decision owner Human print-service process owner; intake-record owner resolves source disputes.
Version and state Candidate summary rule v2; awaiting approval. No implementation claimed.

Have a reviewer who did not write the proposal compare its expected outcomes with these counterexamples:

These are tabletop expectations, not results from a working integration. A failed counterexample returns P-01 for revision.

First-use checklist

Try the worksheet on one review session before adding standing instructions:

Consequential approval remains with the named human print-service process owner; correcting a summary never authorizes printing, purchasing or dispatch. Stop when source support, scope, permission or ownership is missing, or a counterexample fails. A model confidence score cannot resolve those conditions.

After any approved rule change, record implementation as unverified until an independent reviewer checks the actual stated version. For the next correction you receive, write down both the permitted edit and what must remain unchanged before touching the draft.

Sources

  1. Guidelines for human-AI interaction design
  2. NIST AI RMF Playbook: Govern

Discuss the process with MAJLS ↗