// FREQUENTLY.ASKED
Everything you need to understand how CrazyGoldFish works, who it is built for, and how to integrate it.
// WHAT.WE.ARE
CrazyGoldFish is evaluation infrastructure for education. We convert evaluation from a grading step into a system layer that captures, structures, and stores learning signals as persistent learning memory — so platforms and institutions can move from score generation to a memory of learning that compounds in value over time.
An AI grading tool automates marking and stops there. CrazyGoldFish operates as an infrastructure layer: policy-aware interpretation of responses, question-level signal capture, structured storage, human-in-the-loop review, and audit trails. It is not a front-end grading feature — it is a governed system layer designed for institutional and product workflows where reliability, traceability, and teacher authority are non-negotiable.
CrazyGoldFish is built for EdTech platforms embedding evaluation into product workflows, coaching and test-prep networks running high-volume subjective evaluation, LMS and ERP builders adding governed evaluation capabilities, and institutions handling large-scale assessment operations. It is not designed for individual teachers, student-facing self-serve tools, or small-scale manual grading workflows.
// HOW.IT.WORKS
CrazyGoldFish returns structured learning signals: a rubric-aligned mark, step-wise correctness flags, conceptual gap identifiers, and a structured feedback string. These outputs are directly consumable by your gradebook, recommendation engine, or reporting layer — no additional processing required.
CrazyGoldFish deterministically flags responses that fall outside expected scoring distributions for mandatory human review. Every evaluation is auditable — full reasoning traces are stored alongside scores so your team can inspect or override any output. Teacher authority is preserved by design, not as an optional add-on.
CrazyGoldFish evaluates typed open-ended responses, handwritten answer sheets (via OCR), step-by-step mathematical workings, and short-answer formats across subjects and languages. Hindi-medium handwritten evaluation is in production. Multi-language support is part of the core architecture.
// INTEGRATION
No major architectural changes. CrazyGoldFish integrates via REST API — you pass the question, rubric, and student response; the API returns structured evaluation outputs. It sits between your submission layer and your gradebook, requiring no schema changes to your existing platform.
Most integrations complete within one to two weeks for the core evaluation API. We provide Mintlify-hosted documentation, sandbox credentials, and direct engineering support during onboarding. We do not leave integrations to self-serve — a CGF engineer is available throughout the process.
No. CrazyGoldFish is usage-based pricing — you pay per evaluation. There is no minimum commitment for sandbox or pilot phases. Enterprise agreements with committed volumes are available for platforms running high-throughput workflows.
// RELIABILITY
CrazyGoldFish achieved 93% reliability in total scoring alignment (ICC), with 89.7% adjudicated correctness vs teacher 82.8% (+6.9pp). 87% of responses fall within tolerance (±6 marks). Source: India AI Impact Summit 2026 Compendium, MeitY.
CrazyGoldFish uses confidence banding. Responses outside expected distribution are flagged and routed for human review before a score is returned. The system does not guess — uncertainty is surfaced explicitly, not averaged away.
All evaluation outputs — scores, reasoning traces, rubric mappings, confidence signals, and human overrides — are stored in CGF's Memory Layer indefinitely. This enables longitudinal learner profiles, audit trails for dispute resolution, and programme-level aggregation across cohorts and time periods.
Book a 30-minute call. We'll walk through your evaluation workflow and confirm fit before you write a line of code.