Voxarah Capture

Voxarah is a literacy evidence system for parents of struggling readers. For this hackathon we built a capture layer on top of its existing deterministic engine, on a separate deployment so the live site is untouched. A standard screener scores a reading by literal word match. That math penalizes documented dialect pronunciations as if they were decoding errors. Voxarah separates the two. Its deterministic rules engine cross-references what a reader actually said against published phonological rules, excuses documented dialect variation by name, and never excuses a real decoding error. Meta tooling works on the material side only. Passage Studio uses Muse Spark 1.3 on the Meta Model API to generate grade-level passages. Every generated passage then goes through Voxarah's own verifier before it can be used. The verifier can refuse. Child voice and child data never leave the device: transcription runs on-device, and no child recording or transcript is sent to any server. Project name: Voxarah Capture. Company / product: Voxarah. Company incorporated in the United States; bootstrapped. Contact email: support@voxarah.com. Live app / demo URL: https://accent-aware-reading-lab.lovable.app WHAT A JUDGE CAN VERIFY LIVE: 1. The test bench. On The Last Test, a dialect reading scores 76% on a literal word-match check and 95% through Voxarah, with the one remaining flag a real error ("toast" for "test"). On the General American control, both checks agree at 76%, so the gap opens only where sourced rules exist. On a struggling reader sample, both checks agree at 86% and nothing is excused, because the errors were real decoding errors. 2. Passage Studio. Unlock the teacher area with the passcode (see judge instructions below), choose Grade 2, and press Generate. The record that opens shows the model name (muse-spark-1.3), the verified sounds, the try count, the time, and the engine version. Decodability is labeled on screen as a pattern heuristic, not a validated readability measure. Press "Add to test bench" and run the Dialect reader: in the recorded run the literal check scored 74% and Voxarah scored 100%. 3. The refusal. In Passage Studio, after the passcode, press "Demonstrate refusal." The checker is set to demand one more target word than a Grade 2 passage can hold under its 60-word cap. The model tries, the verifier rejects it with diagnostics, the diagnostics are fed back for a second try, and the second try is rejected too. The refusal card shows both tries' real reasons and the model name. Nothing from this flow is saved or added to the test bench. MEASURED COSTS (from the live usage log, 2026-10-30): Studio generation, passage passes on first try: about 3.4 cents per press. Studio generation with a retry: about 6 cents per press. Demonstrate refusal press: about 4.5 cents. The $10 daily ceiling covers roughly 150 presses. PRIVACY: A child's voice, recording, transcript, alignment, and scoring stay on the device in the browser. No exceptions. Meta tooling processes passage and curriculum material only, never a child's voice or data. The demo uses the entrant's own voice only. No real child data appears anywhere in the submission. HONESTY LINES: Decodability is a pattern heuristic, not a validated readability measure. The bench scores are prototype word-match scores for the specific readings shown, not DIBELS or Acadience results, and not a general claim about screeners. Voxarah does not diagnose and does not score children. It is an evidence and record system. Whisper transcription is probabilistic and can normalize speech. That is why the deterministic rules layer downstream is the judge. PRIVATE TESTING NOTES / JUDGE INSTRUCTIONS: Open the live build: https://accent-aware-reading-lab.lovable.app Teacher passcode for Passage Studio: (1101) Suggested order: test bench first (Dialect reader, then the General American control, then the struggling reader), then Passage Studio generation, then Demonstrate refusal. Live voice capture is optional. The scored, deterministic path is the sample readers and the typed transcript path.