Technical Documentation
Structured under the field’s five quality criteria. Validity follows Pellegrino’s three-part structure. Each heading has either evidence or a stated gap.
Ontology-based construct graph. 30-construct synthetic DAG with fragility assignment (durable/fragile/stubborn). Clinical packs: 19 CLIN rules, 9 LabPath rules.
Five-level hierarchy: unaided (w=0.95) → open question → hint → worked example → direct instruction (w=0.08). Weight ordering verified in patent golden numerics.
Distractor-targeted generation with misconception-bound traps (CLIN-01). Cascade DAG for misconception propagation. Resolution durability tracking.
Wrightyard validator harness: 28 executable rules, severity staging, fail-closed on crash. Case Source Compiler with trace manifests.
Validity · Reliability · Comparability
Construct graph grounds measurement in domain structure. Depth levels (procedural → conceptual → transfer → integration) map to cognitive complexity.
Scaffold-conditioned weight links measurement to instruction. tierBlockedBy drives next-action recommendations. Policy module maps depth to allowed scaffolds.
Tempered evidence update verified against patent golden numerics (0.5032, 0.8087). CML Rasch calibration with self-test harness (4 designs, all converge, correlation >0.97). 36-cell SIM-001 audit with pre-registered predictions. GAP: No field trial data yet. Inferential validity against human outcomes not established.
Deterministic replay: same seed → same hash. Certificate tamper detection via SHA-256 chain. Self-test harness proves recovery at >95% coverage.
CRN (common random numbers) in SIM-001: base uniforms drawn once, mapped through cell-specific quantile functions. Seeded randomness ensures cross-cell comparability.
Four tiers (instructional → reporting → placement → credential) with ascending evidence requirements. tierBlockedBy tells consumers what stream is short.
4-stream mastery gate catches fluent-but-flawed (misconceptions), memorizers (no transfer), crammers (no persistence). Each mode fails exactly one stream.
Distinguishes helped from unaided performance. Side-by-side comparison demo: same correct rate, 71-point transfer gap. Help is context, not penalty.
36-cell factorial with pre-registered predictions including negative results. Engine loses on state recovery (pre-registered). Wins on calibration and auditability.
Evidence chain rendered in plain language. Scaffold context as fact, never penalty. No numeric weights or thresholds. Family variant with zero internal vocabulary.
Full evidence ledger with context tags. P(mastery) with depth distribution. Certification card with 4-stream gate checks. Countersign flow.
Interactive demos (sandbox, gate, compare). Workshop mode. Live calibration API. 8 published studies recovered.
permittedUses/prohibitedUses on every record. Default prohibits educator evaluation. Consent-scope gated exports. Withdrawal cascade.
Keyboard parity declared in Wrightyard rules (CLIN-13) but not yet audited on the YardStick product UI itself. WCAG audit not performed.
Standalone repo, independent Vercel + Render deployment. 353 engine tests, 54 work-order tests, all passing independently.
CML Rasch + PCM with Newton-Raphson convergence. N=500 factorial (36 cells) running. Privacy export with sufficient-statistic mode.
23 endpoints. Health, events, state, evidence, certifications, calibration, audit, wrightyard. All verified on production.
Current deployment is single-tenant. Multi-district isolation not implemented. Consent-scope enforcement is per-record but not per-tenant.
U.S. Patent Application No. 19/709,832 (filed June 16, 2026). Controlled Multi-Agent Artificial Intelligence Inquiry System with Coauthorship Integrity Verification, Escalating Reasoning Pressure Protocol, Adaptive Misconception-Targeted Distractor Generation, and AI-Resistant Behavioral Evidence Capture. Inventor: Kumar Sumbhav Srivastava.