assess-assignment

name	assess-assignment
description	Evaluate student submissions against SLTs and build assessment transactions for human signing. Agent reads, human judges.
argument-hint	<course-id> <module-code> [--demo]

Assess Assignment

Purpose: Agent reads student submissions for a course module, evaluates each against the module's Student Learning Targets, produces an assessment report with evidence-based reasoning, and builds credential transactions — but only the human signs. The agent does the reading; the human does the judging.

Usage

/assess-assignment <course-id> <module-code>
/assess-assignment <course-id> <module-code> --demo

--demo uses preloaded submissions for the Go course example (CF Office Hours demo).

---

Execution Flow

Phase 1: Load SLTs and Assignment

Fetch the module's Student Learning Targets — these are the assessment criteria:

andamio course slts <course-id> <module-code> --output json

Fetch the assignment prompt so we know what was asked:

andamio course intro <course-id> <module-code> --output json

Display the SLTs and assignment to the user:

## Module <code>: <title>

### Assignment
<assignment prompt>

### Student Learning Targets
1. <SLT text>
2. <SLT text>

If no SLTs found, stop: "This module has no SLTs — nothing to assess against."

Phase 2: Load Submissions

Production mode: Fetch pending assignment commitments for this course:

andamio teacher assignments list --course <course-id> --output json

Filter to:

• course_module_code matching the requested module
• commitment_status is "SUBMITTED" (skip PENDING_TX_COMMIT — those aren't on-chain yet)

Each commitment contains the student's submission in content.evidence (Tiptap JSON — extract the text). The student is identified by student_alias.

Demo mode (--demo): Use preloaded submissions:

Student	Submission
student-01	"This course is about the programming language, Go."
student-02	"This course is about a game called Go."
student-03	"This is what I yell at my children when we have to hurry."
student-04	"Golang"

If no submissions found (production mode), report and exit: "No pending submissions for this module."

Phase 3: Evaluate Each Submission

For each submission, assess against every SLT independently.

Scoring: Each SLT is scored as:

• Met — submission demonstrates the capability described in the SLT
• Not Met — submission contradicts or fails to demonstrate the capability
• Insufficient Evidence — submission is ambiguous; can't determine from this alone

For each score, provide evidence. Quote the specific part of the submission that supports the score. This is what the human reviewer needs to make their decision.

Example assessment:

### student-01
Submission: "This course is about the programming language, Go."

  SLT 1: "I know that this course is about Go, the programming language"
  Score: MET
  Evidence: Explicitly states "the programming language, Go" — directly demonstrates awareness.

  SLT 2: "I can distinguish between Go the programming language and the game of Go"
  Score: MET
  Evidence: By specifying "the programming language," the student implicitly distinguishes
  it from other meanings. Marginal — they didn't name alternatives, but the qualifier is clear.

  Result: PASS (2/2 SLTs met)

Be honest about borderline cases. "Golang" (student-04) demonstrates SLT 1 (it's a common name for the language) but is ambiguous on SLT 2 (doesn't explicitly distinguish). Say so. The human decides borderlines, not the agent.

Phase 4: Present Summary for Human Review

First, show only the summary table:

## Assessment Summary

| Student | SLT 1 | SLT 2 | Result | Recommendation |
|---------|-------|-------|--------|----------------|
| student-01 | Met | Met | PASS | Accept |
| student-02 | Not Met | Not Met | FAIL | Refuse |
| student-03 | Not Met | Not Met | FAIL | Refuse |
| student-04 | Met | Insufficient | BORDERLINE | Your call |

Passing: 1  Failing: 2  Borderline: 1

Then offer to show detailed reasoning:

Want to see the full reasoning? I can show the detailed assessment with evidence for each student before you decide.

Otherwise, tell me your decisions:

• Accept — issue the credential
• Refuse — no credential
• Override — change my assessment (you're the teacher)

Or say "approve all passing" to batch-accept passes and refuse fails. For borderline cases, you must decide — the agent flags them, the human calls them.

If the user asks to see reasoning, show the detailed per-student assessments from Phase 3 (submission text, per-SLT scores with quoted evidence, result). Then re-ask for decisions.

If the user gives decisions without asking for reasoning, proceed directly to Phase 5. Don't force the detail on them — the summary table may be enough.

Never auto-approve. Always wait for human judgment.

Phase 5: Build Transaction (on human decisions)

Build a single transaction containing ALL assignment decisions — both accepts and refuses. The API processes them as a batch.

API schema (POST /v2/tx/course/teacher/assignments/assess):

{
  "alias": "<teacher-alias>",
  "course_id": "<course-id>",
  "assignment_decisions": [
    {"alias": "<student-alias>", "outcome": "accept"},
    {"alias": "<student-alias>", "outcome": "refuse"}
  ]
}

Field reference:

• alias (top level): The teacher's on-chain alias (e.g., "manager-001")
• course_id: The course minting policy hash
• assignment_decisions: Array of AssignmentOutcome objects — one per student
• assignment_decisions[].alias: The student's on-chain alias (from student_alias in the commitment)
• assignment_decisions[].outcome: "accept" (issue credential) or "refuse" (deny credential)

Note: No module_code field — the protocol determines the module from the student's on-chain commitment.

andamio tx build /v2/tx/course/teacher/assignments/assess \
  --body '{
    "alias": "<teacher-alias>",
    "course_id": "<course-id>",
    "assignment_decisions": [
      {"alias": "student-01", "outcome": "accept"},
      {"alias": "student-02", "outcome": "refuse"},
      {"alias": "student-03", "outcome": "refuse"},
      {"alias": "student-04", "outcome": "accept"}
    ]
  }' \
  --output json

Capture the unsigned_tx CBOR from the response.

Present the transaction for signing:

## Transaction Ready

Decisions:
  - student-01: ACCEPT (credential will be issued)
  - student-02: REFUSE
  - student-03: REFUSE
  - student-04: ACCEPT (teacher override — borderline)

Unsigned TX: <cbor-hex>

To sign and submit:
  andamio tx sign --tx <cbor-hex> --skey <path-to-skey>
  andamio tx submit --tx <signed-cbor-hex>

Or I can sign if you provide your .skey path.

If the human provides their skey path:

andamio tx sign --tx <unsigned-cbor> --skey <path> --output json
andamio tx submit --tx <signed-cbor> --output json

Phase 6: Report

After all transactions are processed:

## Assessment Complete

| Student | Agent Recommendation | Teacher Decision | Outcome |
|---------|---------------------|-----------------|---------|
| student-01 | PASS | Accept | Credential issued |
| student-02 | FAIL | Refuse | No credential |
| student-03 | FAIL | Refuse | No credential |
| student-04 | BORDERLINE | Accept (override) | Credential issued |

Accepted: 2  Refused: 2  TX: <tx-hash>

---

The Demo (CF Office Hours, Mar 21)

The Go course example is designed to be funny but illustrative:

Two SLTs:

1. "I know that this course is about Go, the programming language"
2. "I can distinguish between Go, the programming language, and the game of Go"

Assignment: "What is this course about?"

Four submissions — two pass, one fails, one is borderline:

• "This course is about the programming language, Go." — clear pass
• "This course is about a game called Go." — clear fail (wrong Go)
• "This is what I yell at my children when we have to hurry." — clear fail (funny, wrong)
• "Golang" — borderline (correct language, but does it show distinction?)

The punchline: The agent reads, reasons, and recommends. The human reviews, overrides if needed, and signs. The credential lands on-chain — issued by a human who reviewed an agent's assessment. That's the model.

Demo talking points:

• The agent identified a borderline case ("Golang") and flagged it instead of deciding
• The human can override any assessment — the agent's recommendation is advisory
• The credential transaction requires a human signature — the protocol enforces the boundary
• SLTs define what "competence" means for this module — both humans and agents read the same criteria

---

Rules

• Never auto-sign. Always present the report and wait for human approval.
• Never auto-approve. Even clear passes get presented for human review.
• Be honest about borderlines. If evidence is ambiguous, say "Insufficient Evidence" and explain why. Don't round up.
• Show your reasoning. The human needs to understand WHY each SLT was scored. Quote the submission.
• One transaction for all decisions. The API accepts a batch — include all accepts AND refuses in a single assignment_decisions array.
• Include all students in the TX. Passing students get "outcome": "accept", failing students get "outcome": "refuse". The transaction records the teacher's decision for every submission.
• Only assess SUBMITTED commitments. Skip PENDING_TX_COMMIT (not yet on-chain) and any other non-submitted status.
• Use correct API terms. The outcome field is "outcome" (not "result"), with values "accept" / "refuse" (not "pass" / "fail"). The student field in decisions is "alias" (not "student_alias").
• The assessment is the demo. Make it clear, readable, and fun. The Go course is designed to get a laugh — lean into it.

---

Integration Points

Reads:

• Course SLTs via andamio course slts <course-id> <module-code>
• Assignment prompt via andamio course intro <course-id> <module-code>
• Pending submissions via andamio teacher assignments list --course <course-id> (production)
• Preloaded submissions (demo mode)

Writes:

• Assessment transactions via andamio tx build /v2/tx/course/teacher/assignments/assess
• Signed/submitted transactions via andamio tx sign and andamio tx submit

The architecture:

• Agent: reads submissions, evaluates against SLTs, produces report
• Human: reviews report, approves/rejects/overrides, signs transaction
• Protocol: credential lands on-chain, verifiable, issued by the human who signed it