// ASSIGNMENT.EVALUATION
Assignment Evaluation Built for Persistent Learning Memory
Every assignment response generates learning signals. CrazyGoldFish structures them as persistent learning memory your platform can act on.
// FIELD.VALIDATION
// THE.SYSTEM
Assignment Evaluation as Infrastructure, Not a Tool
Assignment evaluation generates signals — not just scores. Every response contains information: what the student understood, where they went wrong, how their thinking has evolved. A grading tool captures the score. CrazyGoldFish captures the signal.
CrazyGoldFish's Assignment Evaluation sits inside your platform as a persistent evaluation layer — processing each response through a structured pipeline, structuring signals as learning memory, and making that memory available to every downstream system that needs it.
// CAPABILITIES
What Assignment Evaluation Delivers
Rubric-Aligned Evaluation
Responses evaluated against configurable rubrics — not fixed answer keys. Each criterion produces a structured signal: score, confidence, and evidence extracted from the response.
Multi-Format Signal Capture
Text, image, and document responses all processed in a single pipeline. Every format produces the same structured signal output — consistent, reusable, and schema-aligned.
Persistent Learning Memory
Every evaluated response adds to the learner's memory layer — building a longitudinal record of where understanding has grown and where gaps persist.
Evaluator Drift Elimination
CGF's evaluation engine maintains consistency across every response — same rubric, same confidence band, same signal structure at any volume.
Feedback Signal Generation
Feedback derived from per-criterion evaluation — not generic copy. 90%+ accuracy at production load, delivered instantly.
// HOW.IT.WORKS
From Response to Structured Signal
Signal Capture
The student response is ingested — text, image, or document. Format is normalised and passed to the evaluation pipeline with rubric context and learner history.
Evaluation
The response is evaluated criterion by criterion against the configured rubric. Each criterion produces a signal: score, confidence, and evidence extracted from the response.
Persistent Memory
Evaluation signals are structured and stored in the learner's memory layer — updating their learning state, recording gap recurrence, and flagging areas for follow-up.
Structured Learning Intelligence Output
The full evaluation output — scores, signals, feedback, gap flags — is returned via API for your platform to surface, act on, or route downstream.
# Step 1: Create assignment + configure model answer
POST /v1/assignment
→ { "assignmentId": "asgn_001" }
POST /v1/assignment/{id}/model-answer
← rubric + step-wise marks
# Step 2: Submit student response
POST /v1/assignment-student-answer
→ { "sheetId": "ASN_2847" }
PATCH /v1/assignment-student-answer/{id}
← upload PDF / JPEG / document
# Step 3: Retrieve evaluation + signals
GET /v1/assignment-student-answer/score
→ {
"submission_id": "ASN_2847",
"criteria": [
{ "name": "concept_accuracy", "score": 8, "max": 10, "confidence": 0.96 },
{ "name": "argument_structure", "score": 6, "max": 10, "confidence": 0.91 }
],
"signals_captured": 7,
"learning_state": "UPDATED",
"gap_flags": ["argument_structure"],
"feedback_generated": true
}// WHY.THIS.WORKS
Validated at Production Scale. Built to Be Stricter Than a Teacher.
// RUBRIC.ADHERENCE
3.91
marks stricter than teachers, on average
The gap is not error. It is consistency.
// QWK.STANDARD
1.000
median Quadratic Weighted Kappa
The psychometric gold standard. Met at the median question.
// INFRASTRUCTURE.GRADE
99.6%
question-to-rubric mapping success
Evaluation accuracy starts with pipeline reliability.
HUMAN.IN.THE.LOOP
Policy-aware. Human-verified. Infrastructure-grade.
CGF's evaluation system is human-in-the-loop by design. Teachers remain the control point — AI handles volume, humans handle governance. This is not an automation layer. It is evaluation infrastructure with built-in human authority at every control point.
// GOVERNANCE
Human-in-the-Loop at Every Control Point
Every evaluation runs through configurable confidence thresholds. Low-confidence responses are flagged for human review before scores are finalised. Evaluation policy — rubrics, marking schemes, escalation rules — lives in your control layer, not in the model.
Confidence-Gated Human Review
Responses below your configured confidence threshold are automatically routed for human review before scores are released. Human-in-the-loop is not optional — it is built into the evaluation pipeline.
Policy-Aware Evaluation Rules
Marking schemes, rubric weights, and escalation thresholds are all configurable at the institution or platform level. Evaluation policy lives in your control layer, not in the model.
Full Audit Trail
Every evaluation decision — model score, confidence value, human override if applicable — is logged with a timestamp and actor record. Disputes are traceable. Decisions are defensible.
// USE.CASES
See Assignment Evaluation in Your Context
EdTech Platforms
Scale subjective assignment evaluation across your learner base without sacrificing signal quality.
See use case →Coaching & Test-Prep Networks
Consistent evaluation across every subjective response — no evaluator drift, full longitudinal memory.
See use case →LMS / ERP Platforms
Integrate structured assignment evaluation output directly into your platform data layer.
See use case →Institutions & Public Programs
Run large-scale assignment evaluation with audit trails, human oversight, and configurable marking policies.
See use case →// COMMON.QUESTIONS
CGF's Assignment Evaluation product converts every student submission into structured learning signals — not just scores. It evaluates responses criterion by criterion with 93% scoring reliability (ICC), captures concept-level signals, and builds a persistent learning memory layer. Every assignment becomes a signal your platform can act on.
What is assignment evaluation in CrazyGoldFish?
CrazyGoldFish's Assignment Evaluation product evaluates student assignment responses against configurable rubrics — capturing structured signals from every submission. Unlike a grading tool that returns a score, it builds a persistent learning memory layer: every evaluated response updates the learner's signal profile, recording understanding, gaps, and trends across the full assignment history.
How does CGF handle subjective assignment responses?
CGF evaluates subjective responses criterion by criterion against a configured rubric — not as a holistic impression. Each criterion produces a score, a confidence value, and evidence extracted from the response. Responses below your configured confidence threshold are automatically routed for human review before scores are released.
Is CGF's Assignment Evaluation a grading tool or evaluation infrastructure?
It is evaluation infrastructure. A grading tool processes a response and returns a score. CGF processes a response, captures structured learning signals, updates the learner's memory layer, generates feedback, and exposes all outputs via a consistent API schema — it is a component your platform builds on, not a widget you embed.
What accuracy does CGF achieve on assignment evaluation?
CGF achieves 93% reliability in total scoring alignment (ICC) and 89.7% adjudicated correctness vs. teacher 82.8% — a +6.9% accuracy advantage on the same rubric. These figures are from Phase 1 production deployment, validated in the India AI Impact Summit 2026 Compendium (MeitY).
How does CGF integrate with an existing assignment workflow?
CGF integrates via REST API. Submissions are sent to the evaluation endpoint with rubric context and learner ID. Evaluation results — scores, signals, feedback, and confidence values — are returned in a structured JSON response. Full API documentation at docs.crazygoldfish.com.
// RELATED · RESEARCH & PLAYBOOKS
// NEXT.STEP
Ready to Make Every Assignment Count?
CGF's assignment evaluation infrastructure captures learning signals at submission, builds a persistent memory layer per learner, and feeds structured intelligence back into teaching. Book a demo to see it in your stack.