AI-augmented decision without harm-calibrated human involvement

Map each AI decision type to a harm tier, enforce the matching human role in the executor, hold high-harm decisions for review, and keep every decision repeatable.

An oversight matrix in config (decision type to harm severity and probability, and to human-in-the-loop, human-over-the-loop, or human-out-of-the-loop) that the decision executor reads before any outcome takes effect. High-harm decisions are written with status='pending_review' (or paused with interrupt_before / can_use_tool on an agent) and only a reviewer can confirm, amend, or reverse them; over-the-loop decisions apply but surface to a supervisor who can intervene; only low-harm decisions run unattended. Each decision stores what is needed to explain or repeat it: pinned model version, temperature=0 or a fixed seed, logged inputs and outputs, and reason codes or feature attributions.

Where it goes: 3 config and feature flags, 1 application source code, 15 agent action surface, 10 logs and telemetry.

What reviewers look for: a written harm-to-involvement mapping that the code actually consults; a pending_review queue, interrupt_before, or can_use_tool approval on high-harm paths before set_status, notify, payment, or tool execution; and per-decision records with pinned model version, inputs, outputs, and reason codes or attributions that let the result be explained or reproduced.

decision_types:
  loan_denial: {auto: true}
  account_freeze: {auto: true}

# harm = severity x probability -> human role
loan_denial:       {severity: high, probability: medium, mode: human_in_the_loop}
account_freeze:    {severity: high, probability: high,   mode: human_in_the_loop}
fraud_flag:        {severity: medium, probability: medium, mode: human_over_the_loop}
address_autofill:  {severity: low, probability: low,    mode: human_out_of_the_loop}

Engineering guidance, not legal advice. Each provision below may add its own details (a cadence, a deadline, a required notice element): open it for those.