Under AI Conditions 2.3B.2-3

The Structural Problem

When team decisions are made by humans who follow published processes, transparency and auditability are difficult but achievable — the decision traces back to the human who made it, the criteria applied, the reasoning documented. When those decisions are mediated by AI systems — screening resumes, recommending technical direction, flagging performance, routing communication — the transparency breaks.

The break is structural, not merely technical. The system producing the decision cannot explain why it produced it in human-interpretable terms, may optimize for objectives the team did not intend, may encode biases from training data the team did not create, and can change behavior with a training update no one authorized. When AI is mediating team decisions, encoded objective law must become stricter, selective transparency must become more rigorous, and the consequences of failure multiply — because errors are systemic, affecting many people simultaneously rather than one at a time.

Phase I's stress test (1.3.9) identified the strain condition on epistemology and ethics under adversarial pressure. AI-mediated team decisions are a direct instance of that condition: the team's epistemic infrastructure is partially delegated to a system that operates faster than human review, at a resolution humans cannot inspect, under conditions that shift without announcement.

Four Load Cases

The Screening Problem

An AI system screens job applications, performance data, or candidate pools. The structural risk: the system makes decisions affecting people's livelihoods without human comprehension of how the decision was reached. A person is not hired. Why? The system flagged something. What? No one can say.

Structural requirements: Every person affected by an AI-mediated decision must know the system was involved. Rejections must carry an explanation, even if approximate ("the system identified a mismatch between your experience and the role requirement; you can request human review"). Any AI-generated ranking affecting hiring, promotion, or compensation must pass through a human review step before the decision is final — not rubber-stamping, but contextual verification. The AI system itself must be auditable: what data it uses, what patterns it recognizes, whether those patterns correlate with protected characteristics, and whether its behavior has changed since last audit. Both the AI's recommendation and the human's decision are logged — so that disputes can trace the full decision chain.

The Measurement Problem

An AI system measures team productivity, individual performance, or system health by analyzing logs, communications, or work artifacts. The structural risk: the system generates metrics that influence decisions (raises, promotions, discipline) but the basis for those metrics is opaque. No one knows exactly what is being measured, whether the measurement is valid, or whether confounding factors are distorting the result.

Structural requirements: Publish the calculation — what raw data is collected, how it is transformed into a metric, what the metric claims to measure, and what it is not measuring that matters. Validate the metric against reality before using it for decisions: have humans independently assess the same thing the AI measures and compare results. Maintain a human interpretation layer — the metric is data, not a decision. Ensure the person being measured can see what data is collected about them, how it is processed, and what metric resulted.

The Automation Problem

An AI system automates decisions that normally require human judgment — writing performance feedback, generating improvement plans, recommending termination. The structural risk: the system produces output that appears authoritative but reflects no actual human deliberation. A person receives AI-generated feedback that is generic, off-base, or harmful. The feedback shapes their perception without any human having thought about their situation.

Structural requirements: AI can draft but cannot decide. Outputs too consequential to automate — performance reviews, discipline decisions, termination letters, career feedback — require human authorship of the binding content. AI-generated and human-authored content must be clearly marked. Any document affecting a person's career must carry a human signature from someone who takes responsibility for its accuracy. The person affected has a right to request human review of any AI-generated feedback.

The Decision Drift Problem

An AI system's behavior changes without announcement or authorization — a training update shifts weights, adversarial input affects reasoning, something in the deployed system differs from what was tested. The structural risk: yesterday's reasonable decisions become today's different decisions, affecting people before anyone notices.

Structural requirements: Quarterly capability auditing — run the system against known test cases and check for consistency, new capabilities, degraded capabilities, and behavioral drift. Maintain a version history logging every update, what changed, who authorized it, and the expected impact. Monitor for decision drift — are recent decisions statistically different from historical ones? Maintain immediate rollback capability so that a drifting system can be reverted and affected decisions re-evaluated.

The Enforce-in-Code Decision Tree Under AI Conditions

When an AI system is involved in any enforce-in-code decision, the five questions from the decision tree (2.3B.2-1) become stricter. An additional Question 0 — Is this decision being mediated by an AI system? — triggers the stricter standard.

Question 1 requires that conditions be clear not only to humans but to the AI system — vague human-readable criteria like "person is ready for promotion" will be silently reinterpreted by the system into proxy metrics that may be wrong. Question 2 requires criteria that are measurable without interpretation — "code quality" is too subjective for an AI system even if humans can roughly agree on it. Question 4 requires a default that protects people if the AI system fails — goes offline, produces garbage output, or drifts. Question 5 requires verification through test cases and logs, not through monitoring individuals to check whether the AI is behaving.

The Transparency Matrix Under AI Conditions

The transparency matrix adds two requirements when AI systems are involved.

AI decision log. Every decision made or mediated by an AI system is logged with system version and date, input data used, system output or recommendation, whether the output was accepted, modified, or overridden, human reasoning if overridden, and timestamp. This log is auditable by the integrity officer and accessible to the person affected (their own entries).

Personal data used by AI. If an AI system uses personal data to inform a decision, the person must know what data was used, how it was used, and can request access to their own data and dispute inaccuracies. Training data sourced from team members requires explicit consent — historical data is not repurposed without it.