S. Taneja
Project I · Deployed Platform

Teacher ONE: Deterministic Grounding & Epistemic Refusal

A retrieval-grounded study platform for secondary curricula engineered to enforce boundary fidelity. Comprising thirty modules, 101 interactive structural diagrams, 9,226 cryptographically pinned passages, and a pre-generative confidence gate calibrated to refuse ungrounded queries.

1. Executive Summary & Pedagogical Motivation

Teacher ONE originated from an empirical observation made during the deployment of an earlier educational system, Scholar Pro Connect. Conventional digital learning tools typically operate on the premise that student underperformance is caused by a scarcity of learning resources. Consequently, existing platforms aggregate vast repositories of lecture notes, video recordings, and problem sets. Observation of actual study habits revealed that students rarely suffer from an absolute deficit of material; rather, they lack mechanisms to evaluate the correctness of their comprehension in real time.

The introduction of unconstrained generative artificial intelligence into secondary education exacerbated this vulnerability. General-purpose large language models prioritize conversational fluency over domain accuracy. When presented with queries outside a prescribed examination syllabus, such models frequently generate authoritative, polished explanations that contain subtle factual inversions or extraneous theoretical concepts. For a secondary school student revising for national board examinations, an uncalibrated model answer presents an acute epistemic risk: plausible falsehoods are rehearsed, internalised, and reproduced under examination conditions.

This technical report terms this structural vulnerability boundary truncation: the pedagogical requirement that instructional explanation must terminate precisely at the perimeter of the evaluated curriculum, rather than at the boundaries of general human knowledge.

2. System Architecture & Structural Refusal Pipeline

To eliminate hallucination and ungrounded extrapolation, Teacher ONE inverts the standard generative framework. Generative synthesis is rendered strictly subordinate to deterministic document retrieval, and retrieval is conditioned upon an antecedent confidence gate.

Student Query Submission │ ├─▶ Lexical Retrieval Engine (BM25 Parameters: k1 = 1.5, b = 0.75) │ Corpus: 9,226 verified NCERT & formula passages (SHA-256 Pinned) │ ├─▶ Confidence Gate Evaluation (Similarity Metric) │ Score ≥ 0.60 → Admit: Ground generation exclusively in top-k passages │ Score ≥ 0.35 → Admit with explicit hedging disclaimer │ Score < 0.35 → Categorical Pre-Generative Refusal ("Syllabus Boundary Exceeded") │ └─▶ Constrained Synthesis Engine Generation permitted only upon passage admission; citations mandated

By enforcing refusal prior to prompt assembly, the generative language model is never exposed to queries that lack corpus support. Refusal operates as an architectural invariant rather than a post-hoc output filter, rendering the system resilient against prompt manipulation or conversational circumvention.

3. Technical Design Analysis & Trade-offs

Why employ lexical retrieval (BM25) over dense vector embeddings?

Lexical retrieval provides absolute determinism, mathematical interpretability, and zero runtime model overhead. Every generated answer can be audited against the exact passage indices that produced it. Furthermore, the BM25 index compiles to a lightweight data artifact rather than neural network weights, allowing local execution without GPU acceleration. The empirical trade-off is acknowledged: BM25 cannot recognize semantic equivalence when terminology diverges from the index, resulting in retrieval misses on heavily visual or non-textual questions.

Why enforce client-side execution over hosted API endpoints?

Client-side execution satisfies three primary requirements: absolute data privacy, curriculum permanence, and cost elimination. A student's revision query history and diagnostic performance records remain strictly local to the host machine. Additionally, local execution insulates the platform from remote API drift or model updates. Crucially, the marginal operational cost per student is reduced to zero, enabling distribution across economically constrained secondary schools. The packaged backend cold-starts in 455 ms with a memory footprint of 78.6 MB.

Why select Electron over lightweight desktop alternatives such as Tauri?

Guaranteeing offline functionality requires bundling both the retrieval index and the local execution runtime within the client application. While Tauri produces significantly smaller binary distributions, Electron offers established cross-platform packaging stability across Windows and macOS environments without requiring external system dependencies.

Why prioritize epistemic refusal at the expense of user convenience?

In academic assessment, providing no answer is strictly preferable to providing a plausible but incorrect answer. A refusal forces the student to consult primary course literature, whereas a hallucinated answer instils false confidence. The refusal rate represents the deliberate price of mathematical integrity within an automated instructional system.

4. Quantitative Benchmarks & Empirical Findings

Full experimental methodologies and statistical distributions are documented in the evaluation report. Key findings include:

  • Retrieval Efficiency (BM-1): The BM25 engine retrieves the true ground-truth passage at rank 1 in 97.33% of evaluated test cases, achieving top-3 recall of 100.0% across 75 standardized curriculum queries. Mean reciprocal rank (MRR) reaches 0.9867.
  • Gate Discrimination (BM-2): The confidence gate discriminates between in-syllabus and out-of-syllabus queries with an F1 score of 0.984 across 50 probe vectors.
  • Controlled A/B Evaluation (BM-4): Routing queries through the confidence gate elevates the refusal rate on unanswerable out-of-syllabus questions from 0.0% to 100.0%, demonstrating statistical significance at p = 3.1 × 10⁻⁵.

5. Systemic Limitations & Ongoing Research

  • Gate Calibration Over-Refusal: On answerable edge-case queries, the gate exhibits an 83.3% refusal rate under conservative safety thresholds. Re-calibrating these decision boundaries without compromising precision represents the principal open challenge.
  • Lexical Representation Ceilings: Subjects relying heavily on symbolic notation, such as Advanced Mathematics and Structural Biology, exhibit lower first-rank retrieval rates due to vocabulary mismatch.
  • Scope of Evaluation: Current metrics quantify retrieval precision and refusal fidelity; they do not measure long-term cognitive retention or student learning outcomes. Controlled longitudinal trials are required.

6. Version History & Iteration Milestones

v0.1
Static curriculum indexing and structured notes hierarchy. Initial prototype without automated inference.
v0.2
Integration of unconstrained generative model layer. Internal testing revealed significant hallucination on out-of-syllabus queries.
v0.5
Implementation of BM25 lexical retrieval over 9,226 verified NCERT textbook passages.
v0.8
Introduction of the pre-generative confidence gate. Execution of ablation studies confirming refusal behavior stems from gate calibration rather than model prompt instructions.
v1.0
Final desktop packaging and execution of standardized benchmark suites BM-1 through BM-6 against the release binary.

7. Technology Stack Specification

  • Electron
  • Next.js
  • React
  • TypeScript
  • Tailwind CSS
  • Python Local Engine
  • BM25 Search
  • Supabase
  • NextAuth
  • Cloudflare R2
  • Cloudflare Turnstile
Verifiable Artifacts & External References
Academic Correspondence

Inquiries regarding gate calibration and refusal architecture.

tcshaksham@imperialecc.com