A secondary screener for oral reading fluency.

FluencyScope measures oral reading fluency: a student reads a leveled passage aloud for sixty seconds in the browser, and the reading is scored for words correct per minute (WCPM), accuracy, and prosody, with a short comprehension check that qualifies the fluency score. It is a secondary screener, designed to complement district-mandated instruments such as Acadience — never to replace them.

It is deliberately not: tutoring or practice software, curriculum, a standalone comprehension assessment, a student-facing product with accounts or logins, or an analytics platform that groups or ranks students across classrooms. Results are advisory screening data for the teacher — not a diagnosis, a grade, or a placement decision.

Word-level scoring is deterministic. Speech is transcribed by an automatic speech recognition engine, then aligned to the passage and classified (correct, substitution, omission, self-correction, mispronunciation) by transparent, rule-based algorithms. No language model judges whether a word was read correctly. Self-corrections count as correct, following Hasbrouck–Tindal convention.

Scores are reported against the Hasbrouck–Tindal (2017) norms — public domain and widely recognized — as At / Below / Well Below Benchmark bands for beginning, middle, and end of year. Whole-class screening uses a median-of-three-passages protocol (Forms A/B/C per level), mirroring Acadience administration so bands are interpretable in district-familiar terms. Passages are leveled (Lexile 520–1185) with below-grade routing for students reading below grade level.

AI is used only where judgment is explicitly advisory and labeled as such: a prosody rating on the familiar 1–4 scale, a short written observation for the teacher, and grading of open-response comprehension answers. The model receives structured data and text only — never raw student audio.

FluencyScope is built against the Chancellor’s AI guidance (March 2026) as a design constraint, not a compliance afterthought:

• Advisory-only framing. Every AI-generated output carries a visible advisory disclaimer; no output is framed as a determination about a student.
• Prominent teacher override. Teachers can approve, reject, or reclassify any flagged word, adjust prosody, and edit summaries; metrics recompute from teacher corrections, and the override is the record.
• Public explainability. A plain-language page describing exactly where deterministic rules end and AI begins is public at /explainability, linked from every report.
• Voice treated as potentially biometric. Student audio is accessible only to authenticated teachers of the student’s school; students never need accounts, and the student flow collects a first name and initial only as entered by the student.
• No model training on student data. No student audio, transcript, or result trains any model — ours or a vendor’s. Teacher corrections are used only as categories to refine rules and prompts, never as training data.
• Data retention policy with automatic audio deletion: in development— reports are already designed to remain fully usable after audio removal.

• Multi-tenant isolation is enforced in the database itself (Postgres row-level security keyed to each school) — a school’s data is invisible to every other school by construction, not by application logic.
• Students access assessments through an expiring shared link; no student accounts, emails, or passwords exist in the system.
• Reports are teacher-only: anonymous access to session results is disabled at the database-policy level.
• Processing runs on Vercel and Supabase infrastructure with transcription by Deepgram and advisory text by Anthropic; all vendor processing is inference-only, with no training on student data.

We hold the position that a fluency screener’s largest risk is silent: ASR errors masquerading as student reading errors, disproportionately for students whose speech differs from the engine’s training distribution. Our stance is to validate before wider deployment and to publish the method:

• Transcription accuracy (WER) validation against hand transcriptions, with explicit coverage of AAVE, Spanish-influenced English, and newcomer accents — the acceptance test is that dialect and accent do not score as reading errors. in progress
• Matched-student comparison — the same students assessed manually and via FluencyScope, coordinated through the piloting school as data custodian. planned
• ERMA review (NYC DOE research and data-privacy approval) as the deployment gate for wider use. in preparation

Licensed per school, per year — a school-level decision rather than a district procurement. The student flow is built for real-classroom constraints: standard NYC-spec Chromebooks, flaky WiFi (recordings are kept and retried until upload succeeds — a student never re-reads because of a network failure), and no installation beyond a browser.

Piloting now with NYC middle schools. For a demonstration with your school’s own passages and benchmarks, contact us through your school’s pilot lead, or start with the product overview on the main page.