2026 · RESEARCH · QUESTION

Epistemic Guard

An instrument built to measure what AI assistance costs the person using it

TypeScript
VS Code Extension API
Node.js
LLM integration

Epistemic Guard is a research-grade reproduction of the VibeCheck Explanation Gate and the two-phase epistemic-debt experiment described in Sreecharan Sankaranarayanan (2026), "Mitigating Epistemic Debt in Generative AI-Scaffolded Novice Programming using Metacognitive Scripts", ACM Learning at Scale 26.

It is a VS Code extension for studying AI-assisted programming through metacognitive explanations and code-repair tasks, with LLM-based evaluation, telemetry, and automated experimental workflows for analysing developers' understanding of AI-generated code.

It has produced no results, and says so. The value on offer is a working instrument and a replication path, not a claim.

Problem

Unrestricted AI assistance lets a novice programmer produce working code they do not understand. The literature calls the gap epistemic debt: high functional utility masking low corrective competence.

Question

Ed(t) = Uf(t) − Cc(t) — what you can build, minus what you can fix without help. Can that gap be instrumented inside a real editor, in a way that would survive peer review?

Approach

Build the apparatus, not the finding. This is a research-grade reproduction of the VibeCheck Explanation Gate and the two-phase epistemic-debt experiment from Sankaranarayanan (2026), implemented as a VS Code extension: metacognitive explanation prompts, code-repair tasks, LLM-based evaluation, telemetry, and automated experimental workflow.

System

Open any stage to see what it does.

EXPLANATION GATE — Before AI-generated code is accepted, the participant must explain it. The gate is the intervention under study.

PHASE 1 — BUILD — Functional utility: what the participant can produce with assistance available.

PHASE 2 — REPAIR — Corrective competence: what they can fix once the assistance is withdrawn. The difference between the phases is the debt.

LLM EVALUATION — Automated scoring of explanations and repairs, so the protocol can run without a human grader in the loop for every trial.

TELEMETRY — Structured capture of the interaction inside the editor — what was accepted, when, and after how much reading.

ANALYSIS — A reproducible analysis pipeline. On a fresh clone it prints "No experimental data collected yet." — because none has been.

Demo

Install the extension, run the two-phase protocol on the bundled code-repair tasks, and inspect the telemetry the run produces. The analysis script executes on that data.

Result

No experimental data has been collected with this software. This is stated at the top of its README, deliberately.
The repository is the apparatus, verified against itself. It reproduces no published result.
Running the analysis on a fresh clone prints "No experimental data collected yet."

Zero collected trials — Stated by the repository README and enforced by the analysis script output.

Lesson

The honest README was harder to write than the code. There is real pressure to describe an instrument as though it had already produced a finding, and the sentence "this reproduces no published result" costs something to leave in. It is also the only sentence that makes the rest of the repository worth trusting.

Source

Connected

CodeRelay — Both are AI and SYSTEMS work, and both are built with TypeScript, VS Code Extension API, Node.js.
CausalGuard — Connected through SYSTEMS and AI and RESEARCH.
Bnlang for VS Code — Both are SYSTEMS work, and both are built with TypeScript, VS Code Extension API.
ContextGuard — Connected through AI and SYSTEMS.