Educational software · research companion · current release v1.9.6
AI Playgrounds is a suite of 15 multilingual, offline-ready learner labs that make AI mechanisms inspectable in ordinary browsers. Across 13 Foundations/course-track labs and two Modern AI extensions, short activities ask learners to predict, change one condition, inspect the evidence, and explain the result.
Version boundary
The live product has continued to develop beyond the DOI-backed archive. The portfolio therefore identifies both artifacts without transferring claims or identifiers between them.
The current release contains 15 labs, all 15 Level-1 Quick Assigns, and tested English, Simplified Chinese, Vietnamese, and Spanish support across those learner surfaces.
The versioned GitHub release identifies the published website and downloads. The v1.0.1 DOI does not identify v1.9.6. The public repository distributes the website, classroom downloads, and public documentation; it is not the development source or full test suite.
The archived release preserves the earlier 12-applet bilingual boundary and the original 45-algorithm-check and 63-browser-case evidence lineage.
Those product and test counts describe the historical archive. They are not presented as totals for the current release.
Current product
The catalogue covers search and optimization, logical agents and SAT, probabilistic reasoning, supervised and unsupervised learning, neural networks, computer vision, reinforcement learning, adversarial search, Transformer language modeling, and agent tool-use/runtime protocols.
Each lab centers one bounded mechanism and connects learner controls to visual state, text-equivalent explanation, scenarios, teacher guidance, and fidelity disclosures where appropriate.
All 15 labs have standalone HTML downloads. Copy or rename one file and give it to students: its lab, four learner languages, and Quick Assign work without installation, neighbouring files, a server, an account, or an internet connection. Website versions may use neighbouring assets; optional links to teaching resources still need internet access.
All 15 labs and all 15 Quick Assigns support English, Simplified Chinese, Vietnamese, and Spanish under the tested release paths, including learner-state and response preservation during locale changes.
Some educator and support pages remain English/Chinese or English-only. The two current Level-2 Activity Pack canaries are English-only. Four-locale learner support is not a claim that the entire website or every teaching resource is available in four languages.
Pedagogy layer
Every lab now has one stable Level-1 Quick Assign, designed for approximately 10–15 minutes and routed through the applet's existing mechanism rather than a parallel worksheet simulation.
Fifteen permanent activity IDs give teachers canonical classroom links. Learners record mechanism-specific predictions, evidence, explanations, and transfer responses inside local browser surfaces that can be copied or printed.
Two longer pilots, NN-1 and CNN-1, extend the same inquiry spine across connected experiments. They are evidence for the resource architecture, not a claim that every lab already has a longer Activity Pack.
v1.9.6 classroom reliability
This release refines the existing 15-lab suite rather than adding more labs. Transformer Language Modeling, Agent Tool Use and Context Protocols, and Minimax/Alpha-Beta retain their scenario-led teaching layers alongside the foundational labs.
Quick Assign drafts and chosen evidence are preserved across reloads and supported language changes. Clear actions ask for confirmation, storage failures are made visible, and sharing experiment settings does not include learner answers.
Revised learning prompts distinguish prediction from observation and explanation. Targeted repairs address Minimax prediction/reveal behavior, reinforcement-learning updates, and Bayesian preset inference.
Featured experiments lead directly to native controls in all 15 labs. Keyboard focus and recovery paths are improved, while scenarios continue to drive the actual mechanism rather than a separate presentation-only simulation.
Repairs improve enlarged-text layouts, small-screen controls, translated assignments, and dark-theme readability. The NN-1 and CNN-1 Activity Pack pilots also improve long-answer printing and storage-failure handling.
Verification
Current assurance combines algorithm checks, lab-specific references, cross-runtime parity, localization and state contracts, deterministic composition, and browser behavior. The historical 45/63 pair is retained as provenance rather than reused as the present release total.
Evidence boundary
Software behavior, deterministic composition, selected algorithm/reference relations, release integrity, supported-path locale and state preservation, responsive interaction behavior, and representative contrast measurements.
Measured learning gains, classroom adoption, comparative educational effectiveness, universal learner preference, translation naturalness for every learner group, accessibility conformance, equitable access, or production-system equivalence.
The broader manual file/lab audit remains unfinished and continues after v1.9.6. Human learner, educator, and screen-reader validation is deferred. This is a tested software release, not a claim of a completed comprehensive audit, WCAG conformance, or validated learning outcomes.
The applets are inspectable teaching implementations. Their fidelity notes define which mechanism is represented and where the model stops; release verification does not convert a bounded teaching model into a claim about a full production solver, frontier language model, deployed agent, or hardware-performance benchmark.
Privacy and reuse
AI Playgrounds has no student accounts, submission backend, database, or student-upload endpoint. Built-in packet, Quick Assign, and Activity Pack responses remain in the learner's browser until copied, printed, exported, cleared, or otherwise handled by the learner or teacher.
The canonical site uses privacy-minimized aggregate analytics for canonical page/event identifiers and allow-listed campaign attribution. It excludes learner answers, experiment values, query-state payloads, names, grades, and general referrer URLs. Global Privacy Control, Do Not Track, local opt-out, offline copies, localhost, and non-canonical mirrors disable project analytics.
Role demonstrated
Product design, front-end implementation, bounded AI mechanism modeling, multilingual instructional design, teacher-resource architecture, deterministic and browser verification, release engineering, research provenance, and privacy-minimized analytics.