System Investigations

Applied investigations into failure modes, delegation, constraint, and system control.

AI Authorship Evaluator

Expand
Failure Mode: AI-assisted workflows shift control of decisions and reasoning away from the human without detection
Intervention: Rule-based comparison of expected vs observed authorship
Outcome: Undetected transfer of authorship control becomes impossible
  • - Task conditions define expected authorship range
  • - Evaluation measures agency, cognitive engagement, evidentiary control, and authorship integrity
  • - Observed authorship is computed using structured criteria
  • - System compares expected vs observed authorship
  • - Output quality and correctness are excluded from evaluation
→ GitHub Link

Advisory Drift Study

Expand
Failure Mode: Mid-process context injection replaces initial recommendations with new classifications and constraints
Intervention: Fixed interaction sequence with delayed context injection and post-hoc scoring
Outcome: Recommendation replacement without trace becomes impossible
  • - Initial recommendation is required before additional context
  • - Sessions are isolated with no iteration or mid-run edits
  • - Context injection occurs only after recommendation is established
  • - Outputs are captured before any scoring occurs
  • - Evaluation enforces classification shift, constraint durability, and urgency tracking
→ View Case Study (PDF)

Classroom System Study

Expand
Failure Mode: Instructional escalation is driven by staff discretion under delayed behavioral data
Intervention: Rule-governed instructional lanes with interval-based data capture
Outcome: Discretionary escalation outside defined pressure rules becomes impossible
  • - Instruction is limited to three lanes with rule-based transitions
  • - Escalation decisions are governed by predefined pressure rules
  • - Data is captured at fixed intervals with append-only entry
  • - Graph generation occurs automatically without manual transfer
  • - Classroom zones are mapped to instructional pressure states
→ View Case Study (PDF)

Failure Without Rescue

Expand
Failure Mode: Silent correction replaces rule-based outcomes during simulation runs
Intervention: Irreversible decision locking with delayed resolution sequencing
Outcome: Outcome alteration after submission becomes impossible
  • - State display is read-only with manual updates only
  • - Decision capture is append-only with no edit or retry paths
  • - Resolution executes only after decision lock with no external input
  • - Data flow is one-directional from decision → resolution → reveal
  • - Instructor cannot modify outcomes at any stage
→ View Case Study (PDF)

Governed Writing System

Expand
Failure Mode: AI systems generate and modify text without exposing student decisions or authorship
Intervention: Rule-enforced interaction flow with immutable checkpoints
Outcome: System-driven rewriting or inference of student work becomes impossible
  • - Progression requires explicit acknowledgment at each stage
  • - Drafts are generated only from student-provided ideas
  • - Drafts must be frozen before critique begins
  • - Critique is limited to question-based prompts only
  • - All interactions produce immutable logs
→ View Case Study (PDF)

Intent Before Movement — Training in Dungeon

Expand
Failure Mode: Execution proceeds on incomplete plans, producing mismatched outcomes
Intervention: Full instruction commitment before execution with non-intervening system behavior
Outcome: Mid-execution correction or system inference becomes impossible
  • - Execution begins only after full instruction submission
  • - System executes instructions exactly with no correction
  • - No hints, previews, or feedback are provided during execution
  • - Environment is constrained to fixed grid and command set
  • - Revision occurs only after full execution completes
→ GitHub Link

Making Intent Legible — DALL·E

Expand
Failure Mode: Image generation resolves underspecified prompts through untracked assumptions
Intervention: Pre-generation intent gating with schema validation
Outcome: Generation from underspecified or unvalidated intent becomes impossible
  • - Generation is blocked until required fields are validated
  • - Interaction isolates decisions one question at a time
  • - Output is structured into JSON representations
  • - Tiered schemas enforce increasing constraint depth
  • - Guardrails constrain topic scope during prompt construction
→ View Case Study (PDF)