Skip to content

Assessment rubric

This page describes the earlier fourteen-week layout and is being rewritten; where it differs, the syllabus applies.

Rubric version 0.9.3. This is an explicit English migration of the canonical eight-dimensional rubric. Each dimension has four bands: 1 Missing or untrustworthy, 2 Basic, 3 Professional, and 4 Exemplary. Scores depend on observable work and evidence, never a guess about whether writing or code was generated by AI.

Only the dimensions selected by an assignment count. They have equal weight unless the assignment explicitly provides reviewed positive weights for every selected dimension. Unselected dimensions do not enter the denominator. A grader records the rubric version, selected dimensions and either default equal weighting or the normalized assignment override with its source. Human items remain pending until actually assessed. The selected dimensions guide evidence review; the teaching team records the course mark in the teaching team's score sheet.

Specification and acceptance criteria

  1. No specification, or acceptance criteria cannot be executed or are unrelated to the need.
  2. Specification and acceptance criteria exist, but boundary cases or traceability to user needs are missing.
  3. Each feature has executable acceptance criteria traceable to real user research or approved equivalent discovery evidence, with boundary cases stated.
  4. Acceptance criteria cover exceptional paths and security constraints and are reused directly by the evaluator.

Testing strategy

  1. No tests, or assertions cover no observable behavior and cannot detect known defects.
  2. Unit tests exist, but the mutation score is below the threshold and there is no property-based testing.
  3. Tests precede agent implementation; the mutation threshold is met and key logic has property-based testing.
  4. The mutation score is substantially above the threshold; fuzzing or contracts are present, and test failure messages directly help locate the problem.

Security

  1. Scans were not run, or unresolved high-severity alerts or leaked secrets remain.
  2. Scans ran, but some alerts have no disposition record and lockfiles are incomplete.
  3. SAST, secret scanning and dependency audits have no remaining findings or each alert has a disposition record; lockfiles are complete.
  4. LLM features have injection probes and mitigations; permissions are minimized and explained.

Review quality

  1. No review of agent PRs, or the review merely repeats the diff.
  2. Review comments are general and identify no real defect.
  3. Review comments identify real defects and suggest corrections.
  4. Review systematically covers correctness, security and maintainability and records problems the compared review, such as an AI review, missed.

Human verification evidence

  1. A merged PR lacks a causal explanation, or its explanation contradicts the diff.
  2. The explanation says what changed but does not explain why or where it might fail.
  3. Every merged PR explains why the change was made, what might fail and what was personally verified.
  4. A new reader can reproduce the verification; the decision log records rejected alternatives and the reasons for rejecting them.

Agent workflow design

  1. No context files or configuration; the student cannot explain how the agent was used.
  2. Context files and configuration exist, but evaluation gates or cost records are missing.
  3. Context files, skills, hooks, permission policies and evaluation gates are evidenced; costs are recorded.
  4. A/B or cost comparisons informed workflow changes, and reusable team conventions exist.

Process integrity

  1. After the evidence-completion deadline, verification, explanation or the single AI-use note explicitly required by this assignment is still missing; the process cannot be checked.
  2. One AI-use note and some verification and explanation exist, but required evidence or applicable individual work records are incomplete.
  3. One record connects the work product, verification result, personal explanation and one AI-use note. Applicable work records match version history. Required trajectory fragments or approved alternative evidence are supplied only when the assignment explicitly requires them; a commit trailer on every commit and separate disclosure files are not required.
  4. A third party can audit process evidence and reproduce key decisions.

Reflection and calibration

  1. Reflection or calibration records are absent, or statements contradict observable submission evidence without explanation.
  2. Reflection is general; calibration records are incomplete.
  3. The gap between predictions and measurements narrows over the semester; reflection identifies specific personal mistakes.
  4. Reflection proposes verifiable improvements and carries them out in later milestones.

Use the rubric to verify a change

For a course-authored example, generate a small function from an acceptance criterion, then run its assertions through the course launcher (Week 0 setup). Keep the original failing case, the corrected result and your own explanation. Identify which published assignment dimensions this evidence supports and what is still missing. A passing example does not establish every rubric dimension. Use the assignment's allowed scaffold and published scope; optional extensions do not become conditions for full marks. The rubric does not change the course's assessment weights or pass gates.