Module 3: Threshold Logic: AI Systems Pre-Threshold Signals and Early Warning

Learning Objectives

Identify early-warning signals in AI and algorithmic systems.
Distinguish false positives from genuine pre-threshold signals.
Operate the c-ECO State Machine through a AI Systems scenario.
Make intervention decisions under uncertainty with asymmetric error costs.
Design early-warning architecture for AI Systems CSAM work.

The Signal Detection Problem

The central challenge of Module 3 is distinguishing genuine approach to systemic limits from normal variability. In AI and algorithmic systems, no single indicator should be treated as magical. Pre-threshold governance depends on convergence among physical, institutional, contractual, and systemic signals.

Threshold Logic Principle

A signal becomes c-ECO-relevant when it alters the interpretation of trajectory, reversibility, or institutional duty. The question is not merely whether the signal is alarming; it is whether delay would reduce the capacity to stabilize the system.

Pre-Threshold Signal Classes

Physical / Technical

Model drift, performance degradation, and unexplained decision variance; data bias, representational failure, and dataset obsolescence.

Institutional

Coordination or capacity stress among model developers and deployers, data providers and infrastructure operators, affected users and communities.

Contractual

Failure of existing instruments to preserve reversibility, especially model drift Safe Mode clauses and human oversight covenants.

Systemic

Cascading exposure across acceptable error, drift, and harm boundaries, human oversight capacity limits, data integrity and bias thresholds.

Simulation Exercise: The Delayed Signal

Interactive Simulation Scenario

Your Role: Fellow assigned to advise a faculty panel on an automated decision system, model deployment, AI infrastructure dependency, algorithmic governance process, or compute-intensive platform with systemic risk or public-interest exposure.

The System: Models, training data, automated decisions, computational infrastructure, human oversight, governance controls, public impact, and institutional accountability.

Your Task: Monitor a staged evidence feed, classify signal deterioration, and identify the first defensible point for pre-threshold intervention. Each decision has asymmetric costs: early intervention may be costly, but late intervention may destroy reversibility.

Indicator	Round 1	Round 2	Round 3	Interpretation
Model drift, performance degradation, and unexplained decision variance	Visible	Worsening	Persistent	Tests P proximity
Data bias, representational failure, and dataset obsolescence	Stable	Accelerating	Critical	Tests ΔV
Compute and energy burden acceleration	Incomplete	Contested	Material	Tests σ
Automation dependency and human oversight erosion	Latent	Converging	Cascading	Tests Lr and Safe Mode

Decision Points

Simulation Decisions

1Round 1 — Monitoring or Mandate?

Is ordinary monitoring sufficient, or must the CSAM be revised immediately? Explain what evidence would change your answer.

2Round 2 — Amber or Red?

Signals begin to converge. Decide whether the case remains Amber or requires Red/Safe Mode conduct. Identify the actor with escalation responsibility.

3Round 3 — Cost of Waiting

Explain what reversibility has been lost by waiting. Draft a one-page intervention memo for cohort review.

State Machine Translation

State	Entry Logic	AI Systems Fellow Task
Green	Signals stable and reversibility adequate.	Verify monitoring scope and preserve evidence continuity.
Amber	Trajectory deterioration or uncertainty rise requires closer examination.	Update CSAM, increase monitoring frequency, and identify reversible options.
Red / Safe Mode	Threshold proximity, high uncertainty, or declining Lr makes delay unsafe.	Escalate through institutional channels and draft Safe Mode implications.
Black / Restoration First	Reversibility is severely impaired or boundary breach is imminent/confirmed.	Document loss of reversibility and prioritize stabilization or restoration logic.

Preparation Guide

Step 1 — 90 min: Review early warning concepts: critical slowing down, rising variance, spatial correlation, and institutional lag.

Step 2 — 90 min: Build a signal register using at least five AI Systems indicators.

Step 3 — 120 min: Prepare simulation decision rules for Green, Amber, Red, and Black states.

Step 4 — 60 min: Draft an intervention playbook for one actor: model developers and deployers, data providers and infrastructure operators, or affected users and communities.

Required Materials

Scientific and Governance Foundations

Scheffer et al., early-warning signals for critical transitions.
TFP Manual sections on State Machine, prudential bands, and asymmetric uncertainty.
NIST AI Risk Management Framework.
OECD AI principles.
ISO AI governance materials.

Assessment

Component	Weight	Standard
Pre-Simulation Signal Register	30%	Signals are classified by type, evidentiary quality, and TFP relevance.
Simulation Decisions	35%	Decisions reflect asymmetric error costs and preserve reversibility.
Intervention Memo	25%	Memo distinguishes monitoring, escalation, Safe Mode, and Restoration First.
Discussion	10%	Participation demonstrates disciplined judgment under uncertainty.