Materiality Dial - Calibration Patterns by Domain
Chapter 13 of Loop Engineering introduces the 3×3 materiality dial - stakes (low / medium / high) crossed with autonomy posture (supervised / bounded / autonomous). The diagonal offers a recommended starting posture, not a universal safety boundary. Use the domain patterns below to calibrate that starting posture.
Each domain entry below names a likely dominant stake, a plausible posture for routine work, and controls worth discussing with domain professionals. Treat these as thought starters, then apply the rules, evidence standards, and accountability practices of the real setting. Pair them with the materiality prompts.
Software engineering
The materiality dial moves with side-effect class: file-only writes are low; shared-repo writes are medium; production deploys are high. Autonomy posture matches: agents can write files autonomously, should ask before pushing to shared branches, must surface every production-touching action. The compensating controls (tests, review, staged rollout) make this the most-tooled domain - calibration is mostly about which controls actually fire for this specific change.
Failure mode to guard against: ambient autonomy creep during long agent sessions. The agent gets used to acting; the user gets used to approving; the materiality of the actions creeps up without a re-verdict.
Healthcare (clinical decision support)
The dial sits permanently in supervised posture for any output that touches treatment, diagnosis, or medication decisions. Bounded autonomy is acceptable for adjacent work (intake summarization, scheduling, documentation drafting) where the clinician reviews before the artifact reaches the patient. Autonomous posture is essentially never appropriate for patient-facing surfaces - Lee & See (2004)'s under-reliance and over-reliance failure modes both have high consequence here.
Failure mode to guard against: automation-bias drift. Clinicians who use the AI assistant routinely begin to under-scrutinize its suggestions, especially during high-load shifts. Compensating: explicit "this requires your verification" framing at every suggestion; visible-verification on the audit trail.
Legal (research, drafting, citation)
The dial bifurcates by output destination: internal research memos can run bounded (associate reviews the AI's output before partner review); anything heading to a court filing or opposing counsel must be supervised (full citation verification before submission). The dominant compensating control is independent citation check - the maker / checker primitive is non-negotiable because hallucinated citations are a sanctionable failure mode, not just an embarrassment.
Failure mode to guard against: citation confabulation (the AI invents a plausible-looking case that does not exist). Verifiable stop: every citation in any AI-assisted output must be re-resolved against the primary source before the output leaves the firm.
Policy (analysis, briefing, recommendation)
Synthesis work (summarizing positions, mapping arguments, listing precedents) can run bounded if each claim is sourced. Recommendations - anything that names a preferred path - must be supervised because the materiality moves from "informational" to "consequential" the moment the document goes to a decision-maker. Compensating: every claim must cite its source with enough specificity that a reviewer can re-verify; sensitivity analysis names what changes if the underlying evidence is revised.
Failure mode to guard against: false-consensus framing. The AI synthesizes evidence in a way that flattens disagreement and presents a "clear best path" when the underlying evidence supports multiple paths. Compensating: require explicit alternatives generation (the Tenet VII pattern) in any synthesis.
Education (instructional design, assessment, feedback)
Curriculum and instructional-design work runs bounded - the educator reviews AI-drafted lesson plans before classroom use. Assessment work (grading, feedback generation) requires supervision because individual learner consequences and equity exposure both rise sharply when the AI's output is part of an evaluative record. Compensating: rubric must precede AI feedback so the AI is scoring against an externally defined standard; sample audits catch systematic biases the rubric did not anticipate.
Failure mode to guard against: systematic bias in feedback or grading that disadvantages a subgroup. Compensating: per-cohort sample audits with explicit demographic breakdowns; immediate refusal of any AI grading surface that cannot be audited this way.
Customer support (routing, response, resolution)
The materiality dial moves sharply with the action class. Routing may be suitable for bounded automation when misrouting is cheap to reverse and escalation is reliable. Draft responses can be reviewed according to risk, evaluation results, and customer impact; a model-generated confidence score alone should not decide whether a message is sent. Account changes - refunds, plan changes, identity-affecting actions - deserve stronger authorization because reversibility and consequence both jump.
Failure mode to guard against: the AI confidently produces a wrong refund or account change because the confidence threshold was poorly calibrated. Compensating: hard floor at supervised posture for any action with financial or identity consequences; the confidence threshold gates routing only, never side-effect actions.
Financial operations (reconciliation, disbursement, reporting)
Reconciliation work - matching transactions, flagging anomalies - can run bounded because the human review at exception-handling time catches the cases where the AI was wrong. Disbursement (any action that moves money) must be supervised because the action is functionally irreversible and the audit / regulatory exposure is large. Compensating: segregation of duties applies even when the AI is one of the "duties" - the AI cannot be both maker and checker on a transaction. Independent verification (different prompt, different model, or human) is mandatory above defined dollar thresholds.
Failure mode to guard against: ambient privilege escalation. The AI assistant is granted broader access to "make work smoother" and gradually accumulates permissions that should have triggered segregation-of-duties review. Compensating: explicit permission audit on a quarterly cadence; any new permission grant must be justified against the dial.
The pattern that holds across domains
The questions travel better than the answers. Each domain has different stakes, evidence standards, routine postures, and non-negotiable controls. The materiality dial provides a common way to ask what is at risk, how much autonomy is justified, and what must remain observable and reviewable.
When adapting the framework to a new domain, ask: what is the dominant stake, what posture does that stake permit, who has authority, and which controls cannot be skipped? Existing professional standards and domain expertise should shape the answer. The framework helps make that reasoning explicit; it does not replace it.
This page extends Chapter 13 of Loop Engineering. Corrections to domain calibrations welcome via errata - domain practitioners are the authoritative source for their own field's compensating-control requirements.