COE: CO for Education

Application COE Status In Analysis (pilot planned) License CC BY 4.0 Base CO v1.1

COE is the instructor-side CO domain application. It does not ban AI. It redesigns assessment so that AI collaboration makes student judgment more visible, not less.

The core insight: if students will use AI regardless, design assessments where the interesting signal is how they use it. The COE Spectrum gives instructors four levels of progressive freedom, each shifting what gets assessed.

The COE Spectrum

Four levels of assessment, from full instructor control to full student autonomy:

Level	Name	Instructor defines	Student defines	What gets assessed
1	Full Constraint	All five system components (agents, knowledge, guardrails, workflow, learning)	Only prompts and decisions	Quality of thinking within constraints
2	Partial Freedom	L3 guardrails, L4 workflow	L1 agents, L2 knowledge base	Agent design + knowledge architecture + output
3	Full Freedom	Minimal safety constraints only	Everything	Full CO setup + output + deliberation journal
4	Meta	Nothing	Setup designed for someone else to use	Setup usability by others (knowledge transfer)

What progresses across levels

Level 1	Level 2	Level 3	Level 4
Following instructions	Designing experiments	Designing methodologies	Teaching

Each level increases the student’s responsibility for designing the AI collaboration setup. At Level 4, the student creates a setup that another student can use. This is the hardest assessment: it requires genuine understanding of both the domain and the methodology.

Agents

COE deploys two categories of agents:

Student-facing (configured by the instructor, used by the student):

Agent	Role
Domain Expert	Subject knowledge calibrated to the course level
Methodology Guide	Walks students through the assignment workflow

Assessment (instructor-side, evaluates student work):

Agent	Role
Rubric Scorer	Scores each rubric dimension against calibrated standards
Pattern Detector	Identifies collaboration patterns across student submissions
Anomaly Flagger	Detects statistical anomalies in submission characteristics
Quality Assessor	Evaluates overall work quality against course benchmarks
Summary Generator	Produces assessment summaries for instructor review

Anti-gaming

COE addresses the “students will just ask AI to do everything” problem at the architecture level:

Assessment targets shift. At Level 1, the output matters. At Level 3, the deliberation journal matters more than the output. At Level 4, the usability of the student’s setup is the assessment. Asking AI to “do everything” produces a poor deliberation journal and an unusable setup.

Audit trails exist. Every student-AI interaction produces a log. The Pattern Detector identifies when students bypass the intended workflow. The Anomaly Flagger catches statistical outliers (e.g., submissions completed in unusually short times, or with unusually uniform quality across rubric dimensions).

The rubric is transparent. Students see the rubric before they begin. This is not a trap. It is a signal: here is what matters. If what matters is judgment quality, students must demonstrate judgment. AI cannot do this for them.

Workflows

COE defines two parallel workflows:

Student workflow (5 phases):

1. Setup: receive the assignment, understand the rubric
2. Exploration: research the topic with AI assistance
3. Development: produce the deliverable
4. Refinement: revise based on self-assessment
5. Submission: package and submit

Instructor workflow (4 phases):

1. Calibration: define rubric, set scoring standards, seed anti-gaming config
2. AI Scoring: assessment agents score all submissions
3. Review: instructor reviews AI scores, adjusts where judgment disagrees
4. Moderation: inter-rater agreement check (threshold: >0.7 between calibration and scoring)

Three approval gates require human judgment before grades finalise.

Academic integrity verification

Every student action is traceable. COE uses cryptographic audit trails (based on the EATP trust protocol) so instructors can verify the full chain from submission back through the AI collaboration to the student’s original prompts and decisions.

• Every student action traces to the student (not to AI)
• Rules define what AI may do at each COE Spectrum level
• Tamper-evident records document the entire collaboration
• The instructor can verify who decided what, and when

Relationship to COL

COE and COL are complementary:

• COE is the Trust Plane: the instructor designs constraints, rubrics, and assessment criteria
• COL is the Execution Plane: the student works within those constraints to produce deliverables

This mirrors the Dual Plane Model from CARE.

Get started

COE is in analysis with a pilot planned. The links below give access to the current research implementation.

1. Download the COE workspace from GitHub and unzip it
2. Open Claude Desktop, switch to the Cowork tab, and open the unzipped folder
3. Type /start and describe your course and assessment goals

The full setup guide walks through each step.

• Source: github.com/terrene-foundation/co-education