Internal AI feature

Glov Content Accelerator GCA

Multi-LLM pipeline for learning games — generate, evaluate, refine

GCA is the internal pipeline that produces and maintains curriculum content for Glov Games learning titles such as GANIXO. It uses independent models for each stage so that no single model both writes and approves its own output, then reconciles every new batch against the questions already published so players never see the same question twice.

How a question moves through the pipeline

Each stage is a separate agent with its own model and its own responsibility. A batch only reaches the backend after the judge approves it and the reconciliation step confirms it is not a duplicate of published content.

1. Plan A curriculum brief fixes the board, class, chapter and difficulty band. For Class 6 and above the brief defaults to the CBSE (India) syllabus.
2. Generate A generative model drafts questions, options, the marked answer and a worked explanation against that brief.
3. Judge A different model — the AI judge agent — independently recomputes the mathematics and checks scope, difficulty and consistency. Rejections return to generation with the specific defect.
4. Reconcile Approved questions are compared against everything already published in the same category. Duplicates are dropped and near-duplicates are rewritten so the pool keeps growing without repetition.
5. Publish Only a human-approved, reconciled batch is written to Amazon DynamoDB through owner-authorized (ADMINS) access.

Agents and their models

GCA is deliberately multi-model. The judge never shares a model with the generator, so a systematic mistake in one model cannot approve itself.

Skills applied

Each stage runs under a documented skill so output stays consistent and reviewable. Skills are the written rules an agent must follow — not prompts improvised at runtime.

• Curriculum alignment

  CBSE/NCERT chapters, terminology and difficulty for Class 6–12; board-neutral for Class 1–5. Rupee currency, metric units and CBSE answer conventions such as π as 22/7.

• Backend content boundary

  Reusable gameplay data — questions, options, answers, explanations, formulas, topics, difficulty bands and generator configuration — lives in the Amplify backend, never as the authoritative copy in frontend files.

• Independent validation

  A different model from the generator re-derives every answer. Arithmetic drift, inconsistent givens and off-syllabus content are rejected before review.

• Deduplication and reconciliation

  Normalized-prompt and semantic comparison against published content, so repeat generation tops up a question pool instead of duplicating it.

• Design system compliance

  Platform typography, colour, spacing and component standards, with accessibility, focus states and responsive behaviour verified across breakpoints.

• Release validation

  Schema and structural checks, dry-run seeding, and a read-only production query confirming record counts before content is considered live.

Reconciliation: new questions against published ones

Regenerating a topic must not duplicate it. Before publishing, GCA reconciles the new batch against existing content in the same category:

• Exact duplicates — removed when the normalized prompt and answer already exist in the category.
• Near-duplicates — flagged when a question tests the identical skill with only cosmetic variation, and returned for rewrite.
• Genuine additions — retained, so each run increases coverage rather than re-adding the same items.

Published writes are idempotent: a question already present is updated in place rather than inserted again, so re-running a batch is safe.

Quality gates

Why multiple models

A single model asked to both write and check its own work tends to approve it. GCA separates the two: generation and judging use different model families, and the judge's verdict — not the author's confidence — decides publication. Where a topic is better served by computed answers than authored ones, GCA publishes deterministic generator configuration instead, so answers are correct by construction.

Privacy and disclosure

GCA operates on curriculum content, not player data. It does not read player progress, cloud saves or personal information. Generated content is reviewed before publication, and curriculum errors can be reported to contact@glov.games.

This page documents an internal authoring tool. The models named above describe the pipeline's architecture; they are not used during gameplay.