Peer-reviewed computer-science research by high-school first authors. Every accepted paper is published together with its full source code, dataset citations, and a reproducible experimental artifact under CC BY 4.0 (text) and MIT / Apache 2.0 / BSD (code).
Volume 1 spans machine learning, algorithms, systems, cybersecurity, human-computer interaction, and computer vision. Every entry below is a first-author paper by high-school researchers, with a public GitHub artifact and a peer-reviewed reproducibility spot-check.
Editorial · pp. 1–4
An opening editorial framing JHCS's mission — why a specialist, peer-reviewed venue for high-school CS research with mandatory open code and artifacts raises the bar for what a student paper can be. Explains the CS-specific rubric (novelty, empirical rigor, reproducibility, software-engineering quality, clarity) and how the two-week review process is designed to teach as well as judge.
Northgate STEM Academy
Machine Learning · Natural Language Processing · pp. 5–22
Bias in large language models has been extensively studied at the token-completion level, but the specific problem of detecting demographic bias in AI-generated college-application essays — where subtle framing choices about family, race, and socioeconomic background can materially affect admissions outcomes — has received less attention. We construct a labeled dataset of 12,000 essay pairs using controlled prompt substitutions across gender, race/ethnicity, and family-income axes, and evaluate whether small (~350M parameter) open-source language models can be fine-tuned to detect these signals. A contrastive fine-tuning objective outperforms both a standard classification head and prompt-only baselines, reaching 0.87 macro-F1 on held-out test essays vs 0.71 for a zero-shot GPT-4-class baseline at approximately 1/400th the inference cost. We release the dataset, checkpoints, and evaluation harness under CC BY 4.0.
Riverside Preparatory High School
Systems · Robotics · pp. 23–38
Autonomous-driving research is dominated by industrial-grade platforms priced out of high-school classrooms. We describe TinyKart, a fully open-source autonomous scale go-kart built from a Raspberry Pi Zero 2W, a wide-angle USB camera, and off-the-shelf brushless components — total bill of materials $178. We benchmark three lane-following controllers over 220 timed laps on a custom-built indoor track: a hand-tuned PID controller, a model-predictive controller, and a knowledge-distilled ResNet-8 policy trained from expert demonstrations. The distilled neural policy reduces mean lap error by 41% and holds 30 fps on the Pi Zero. We release the full CAD, BOM, firmware, training pipeline, and a step-by-step build guide targeted at first-time robotics students, so any classroom can rebuild the platform in a single semester.
Meridian Academy of Sciences
Algorithms · Machine Learning · pp. 39–52
Music-recommendation pipelines on low-bandwidth mobile devices must select a small pool of candidate tracks from a large catalog before applying an expensive ranking model, but classical offline greedy selection is infeasible when the catalog exceeds device memory. We present a streaming (1 − 1/e)-approximation algorithm for coverage-based candidate selection that uses O(k log k) working memory in the size of the desired candidate pool. We prove the approximation guarantee under a natural diminishing-returns assumption on the coverage function and validate empirically on the LFM-1B listening dataset. The algorithm reduces peak memory by 8× versus offline greedy while retaining 96% of ranking-quality plateau, making on-device candidate generation practical for a class of consumer devices previously restricted to server-side pipelines.
Cascade STEM Academy
Cybersecurity · Privacy · pp. 53–72
Password reuse and two-factor-authentication adoption in high-school populations are underdocumented in the security literature, despite this cohort being disproportionately exposed to credential-stuffing attacks on gaming, social, and school-SSO platforms. We survey 3,412 U.S. high-school students on password practices across seven platform categories (Google, Instagram, TikTok, Discord, school SSO, gaming, banking) and cross-reference reported reuse against the Have-I-Been-Pwned breach corpus under a strict k-anonymity anonymization pipeline. 61% of respondents report reusing passwords across ≥3 platforms; only 22% have two-factor enabled on any account. Two-factor adoption correlates most strongly with prior personal breach experience (χ² = 42.1, p < .001), suggesting existing awareness campaigns underperform lived-experience learning. We release the anonymization pipeline, the survey instrument, and a reusable ethics protocol as a reference for future high-school-based security measurement work.
Blueline Preparatory
Human-Computer Interaction · Education Technology · pp. 73–94
Popular discourse on AI chatbots in schools centers on academic dishonesty, but empirical data on how students actually interact with these tools during homework is scarce. We deploy a locally-hosted Ollama chatbot to 217 high-school students across two semesters under an opt-in ethics protocol and log the full text of 84,132 turns. Contrary to the dominant framing, 68% of turns are meta-questions (asking how to approach a problem or verify reasoning) rather than answer-extraction. Students in the top quartile of chatbot usage show statistically significant pre-to-post gains on end-of-unit assessments (Cohen's d = 0.47) versus a matched-baseline comparison group. We open-source the deployment recipe, the redacted conversation corpus, and the analysis notebooks to enable replication in other classrooms.
Southport Technical High School
Computer Vision · Applied Machine Learning · pp. 95–112
Municipalities lack automated pipelines to detect traffic-signal malfunctions — dark phases, ghost greens, cross-conflict timing errors — despite the widespread availability of dashcam and cyclist-camera footage. We construct SignalScope, a weakly-supervised object-detection and temporal-inference pipeline that flags malfunction candidates from raw video. The system is trained on 18 hours of manually labeled clips plus 340 hours of unlabeled footage via pseudo-labeling, with a per-intersection temporal model that eliminates single-frame false positives. On a held-out set of 42 verified malfunction events across 17 intersections in a mid-sized U.S. city, SignalScope achieves 88% recall (37/42) with 0.14 false positives per operational hour. We release the annotations, model weights, and a public web viewer that lets city engineers browse flagged events.