# Infinite Sandbox educational assurance record Assurance version: 2026-07-22.3 Current status: **experimental, unreviewed, and not certified** Instructional challenge profiles: middle school (grades 6–8 candidate), high school (grades 9–12 candidate), and college (supporting-tool candidate) ## What this update establishes The game now has software-enforced educational guardrails: - Sixteen generative STEM concept families have a learner goal, an objective, at least two success criteria, a concept-specific equation policy, an explicit assessment boundary, and candidate standards references. - Every standards reference is stored as `candidate-unreviewed`. The program rejects a teaching plan if it claims certified alignment. - Formulas are checked against the chosen concept. Computed multiplication formulas must be numerically true, ratios must have positive terms, and fixed equations must match their reviewed local template. - Online search can add allowlisted Wikibooks page metadata, but downloaded text cannot create equations, assessments, actions, structures, or code. - Feedback distinguishes material, shape, attachment, location, size, height, sequence, and unsuccessful-action errors. - Repeated mismatches trigger more support. Success produces a transfer prompt instead of ending the learning sequence. - The evidence ledger says `observed_performance`, `practicing`, `demonstrated_performance`, or `retained_performance`; it never awards mastery. Higher stages require distinct plans, approaches, and contexts, independent attempts, and later retrieval evidence. - `level.dat` records the assurance version, evidence interpretation, data handling declaration, and external work still required. - Every teaching display begins with `AI MENTOR`; V overrides a recommendation without an evidence penalty, and B creates a non-identifying concern report. - Shift+H adjusts support and P adjusts pace. H toggles tutorial hints. There is no level selector. Completed mentor assessments move between three instructional challenge profiles using conservative thresholds and hysteresis. These controls improve content consistency and auditability. They do **not** show that a learner understood a concept or that using the game improves learning outcomes. ## Candidate curriculum crosswalks The current mapping is a hypothesis for expert review, not a standards claim. The table below describes the middle-school base catalog. High-school metadata adds quantified criteria, alternatives, test evidence, and tradeoffs. College metadata adds assumptions, repeatable experiments, uncertainty, safety, ethics, and limitations. See `CERTIFICATION_READINESS.md` for the separate K–12, Quality Matters, and ABET boundaries. | Domain | Game concepts | Candidate references | Important boundary | |---|---|---|---| | Science | lever balance, height and potential energy, friction comparison, center of mass | NGSS MS-PS2-2, MS-PS3-2, MS-ETS1-4 | Game actions are physical models; they do not measure force or prove explanations. MS-PS3-2 is treated as relative-energy learning, not a required calculation. | | Technology / CS | ordered processes, binary patterns, debugging, decomposition | CSTA K–12 Level 2 candidate concepts | Broad concept mapping only. Recheck the current standards and CSTA's external alignment-review process before applying. | | Engineering | trusses, beam bridges, foundation pressure, cantilevers | NGSS MS-ETS1-2, MS-ETS1-3, MS-ETS1-4 | The physics and formulas are simplified educational models, not structural or geotechnical analysis. | | Mathematics | rectangular arrays, ratios, right triangles, reflection symmetry | CCSS 6.G.A.1, 6.RP.A.1, 8.G.B.7, 8.G.A.3 | Completing a construction is practice evidence; generalization and mathematical explanation remain unmeasured. | Authoritative references used for the candidate map: - NGSS engineering design: https://www.nextgenscience.org/msets1-engineering-design - NGSS forces and interactions: https://www.nextgenscience.org/topic-arrangement/msforces-and-interactions - NGSS MS-PS3-2: https://www.nextgenscience.org/pe/ms-ps3-2-energy - CSTA K–12 standards: https://csteachers.org/pk12standards/ - CSTA alignment review: https://csteachers.org/k12standards/standards-alignment-review/ - Common Core Mathematics: https://www.thecorestandards.org/Math/ ## Pedagogy basis and limits The model–practice–check–reflect cycle and the adaptive support rules are informed by the IES practice guide on organizing instruction and study and by the Education Endowment Foundation metacognition guidance: - https://ies.ed.gov/ncee/wwc/PracticeGuide/1 - https://educationendowmentfoundation.org.uk/education-evidence/guidance-reports/metacognition These sources informed design decisions. They have not reviewed, endorsed, or certified this product. The game currently infers progress only from observable play. It cannot yet verify a learner's verbal reasoning, misconception model, or ability to transfer the idea outside the game. ## Internal test evidence Automated verification performed for assurance version 2026-07-22.3: - 4,000 deterministic local teaching plans checked for structure, assessment, formula syntax, connected demonstrations, and all 16 concept families. - 2,000 grounded plans checked for concept-level formula validity, curriculum metadata, candidate-only standards wording, success criteria, understandable instruction limits, and network-source restrictions. - 128 simulated learner journeys completed across all 16 concepts and four domains. Each journey included an incorrect attempt, a specific corrective cue, a valid retry, and saved conservative evidence. - 900 additional certification-readiness plans were validated evenly across middle-school, high-school, and college profiles, including disclosure, standards-claim boundaries, user agency, fairness, and provenance checks. - Automatic upward and downward challenge transitions were tested. Input method and action speed were verified to have no effect on those transitions. - Repeated incorrect attempts were tested to ensure the mentor increases support. - Tampered false formulas, untrusted source hosts, and online-source attempts to control lesson content were rejected. - Static accessibility and privacy gates check documented control behavior, live status text, reduced-motion support, Content Security Policy, host restrictions, and the absence of browser profile storage or sensor permissions. - Sapling and mentor integration tests and a production build remain part of the same release gate. Run the gate with: ```text npm test npm run build ``` Automated tests demonstrate software behavior, not educational effectiveness, accessibility conformance, privacy-law compliance, or curriculum alignment. ## Data handling declaration - No player account, name, email, advertising identifier, microphone, camera, precise device location, cookie, or server-side learner profile is requested. - Gameplay observations and adaptive signals remain in the current browser page. - A user who presses Enter downloads those observations in their own `level.dat`; the user controls that file and should avoid posting it publicly. - A user who presses B downloads a non-identifying local concern report. A real certification application still needs a developer-operated receipt workflow. - Online requests contain generic STEM topic terms. They omit credentials and referrer data. Player coordinates, action history, and adaptive signals are excluded. - No advertising or behavioral monetization is included. This data-minimizing design reduces risk but does not replace a deployment privacy notice, data-flow review, retention policy, school authorization, or legal review. A school or service operator must separately evaluate COPPA, FERPA, and any applicable local student-privacy rules. ## Accessibility target WCAG 2.2 is the target, not a conformance claim. Current safeguards include: - keyboard focus and keyboard controls for movement, selection, spawning, reset, save, and tutorial hints; - coarse-pointer controls for movement, camera look, jump, material/shape selection, mining, placing, world options, and tutorial hints; - an assistive-technology description of controls; - live text equivalents for selected material, mentor instructions, and feedback; - reduced-motion handling for interface animation; and - no flashing or timed visual-only success cue. Known gaps include no keyboard mining/building action, the highly visual 3D learning content, English-only instructions, no tested screen-magnifier workflow, no remappable controls, no completed physical mobile-device matrix, and no completed manual audit with multiple screen readers and switch devices. Reference: https://www.w3.org/TR/WCAG22/ ## AI and content risk register | Risk | Current control | Remaining work | |---|---|---| | Incorrect formula | Concept-level local validator and tamper tests | Qualified STEM educator review of every policy and boundary | | Misleading standards claim | Candidate-only metadata and rejection of stronger wording | Formal reviews by the relevant standards organizations or trained reviewers | | Search poisoning or unsuitable text | HTTPS host allowlist; metadata only; online text cannot control lessons | Curated source-page allowlist and periodic content review | | False learner classification | Performance-only stages; no mastery or psychometric claim | Validity, reliability, fairness, and subgroup studies | | Endless failure loop | Specific feedback and support escalation | Human usability study, teacher override, and safeguarding procedure | | Accessibility exclusion | Keyboard and live text equivalents plus automated gates | Independent WCAG 2.2 audit and remediation | | Child/student privacy | No account, identifiers, ads, server profile, or action-data upload | Legal review, public notice, school agreements, deletion/retention procedures | | Physics mistaken for professional analysis | Repeated engineering-simulation disclaimer | Benchmarking and independent engineering validation before any stronger claim | The AI risk workflow should be maintained using NIST AI RMF's Govern, Map, Measure, and Manage functions: https://www.nist.gov/itl/ai-risk-management-framework ## External work required before professional claims 1. Have qualified middle-school, high-school, and college science, mathematics, engineering, computing, accessibility, and learning-design educators review every objective, equation, demonstration, success criterion, and risk. 2. Submit the standards crosswalks to appropriate external alignment-review processes. Until accepted, keep all wording as candidate and unreviewed. 3. Conduct a complete WCAG 2.2 audit with disabled users and multiple assistive technologies, then publish an accessibility conformance report only if earned. 4. Complete child and student privacy impact assessments, deployment notices, data-retention rules, incident procedures, and legal review. 5. Run supervised usability pilots with informed adult consent/assent, safeguarding, stopping rules, and educator observation. 6. Preregister and conduct an independent learning-outcome study using an appropriate comparison and validated measures. Use the current What Works Clearinghouse handbook when planning evidence quality: https://ies.ed.gov/ncee/wwc/Handbooks 7. Audit reliability, construct validity, fairness, and differential performance before interpreting the adaptive evidence as more than game behavior. Only independent reviewers or authorized certification/alignment bodies can award professional status. This build is a stronger auditable prototype and an evidence-collection starting point, not a certified educational system.