@thanh01.pmt/project-assessment-kit

High-quality, evidence-first assessment engine for coding projects (Scratch, Web, Python, Unity, Roblox, IoT).

Features

• 2-Pass Relevance Selection (context-builder.ts):
  • Pass 1: Path heuristics via @thanh01.pmt/domain-kit (scoreByPath).
  • Pass 2: Content logic signal ranking (scoreByContent) with tiered budget allocation (Core: 12KB, Mid: 6KB, Tail: 2KB, total $\le 120$KB).
  • Eliminates naive 5KB/file truncations.
• Scratch Semantic Digest (scratch-digest.ts):
  • Parses Scratch 3.0 project.json into a dense structured summary (sprite inventory, block counts, custom procedures, variables, and event wiring).
  • Preserves core logic scripts while replacing bulky 60KB raw JSON dumps.
• Claim & Evidence Verifier (claim-verifier.ts):
  • Follows research-kit evidence pattern to check that all cited quotes in justifications exist verbatim in the supplied context.
  • Automatically tags ungrounded assertions and supports a single-pass retry repair loop.
• Academic & Objective Prompt Builder (prompt-builder.ts):
  • Anchors evaluation purely on rubric level descriptors without blunt hardcoded caps.
  • Retains unassigned learning objectives with [unassigned-LO] tags.

Installation

pnpm add @thanh01.pmt/project-assessment-kit

Quick Start

import { buildAssessmentContext, parseScratchProject, verifyReportGrounding } from '@thanh01.pmt/project-assessment-kit';

// 1. Build bounded context
const context = buildAssessmentContext(rawFiles, {
  maxTotalBytes: 120_000,
  coreBudget: 12_000,
});

// 2. Verify evidence grounding
const result = verifyReportGrounding(report, context.fullCode);
console.log('Fully grounded:', result.isFullyGrounded);