Documentation

SLP Instruction Manual

Everything speech-language pathologists need to know about GestaltPath — from clinical foundations and AI classification to workflow recommendations and data privacy.

Version 1.0 April 2026 16 sections

1. What Is GestaltPath?

GestaltPath is a mobile application designed to coach parents and caregivers of gestalt language processors between therapy sessions. It does not replace speech-language therapy. It does not interact with the child directly. Instead, it equips the adults in a child's life with tools to:

  • Log their child's utterances quickly (by typing or speaking)
  • Classify each utterance into an NLA stage using AI
  • Receive personalized, stage-appropriate coaching prompts daily
  • Practice responsive language modeling during live interactive sessions
  • Track their child's language development over time

The app is grounded in Marge Blanc's Natural Language Acquisition (NLA) framework and adheres to the clinical principle of modeling language one stage ahead of the child's current production level.

2. Who This App Is For

GestaltPath serves two audiences:

Primary users — Parents and caregivers who want to support their child's language development at home, at mealtimes, during play, and across daily routines.

Clinical partners — Speech-language pathologists who can:

  • Recommend the app to families on their caseload
  • Review exported data to supplement clinical observations
  • Use progress reports to inform treatment planning
  • Validate or override AI classifications flagged for review

The app supports multiple user roles during setup: Parent, Caregiver, SLP, and Teacher. SLPs may create their own accounts to explore the app or to directly support families.

3. Clinical Foundation

GestaltPath implements Marge Blanc's six-stage NLA framework as its classification backbone. Every AI classification, coaching prompt, and progress metric maps to these stages.

The app enforces the "model one stage ahead" rule throughout:

  • Coaching prompts target the child's current stage + 1
  • Live Practice Mode coaching hints suggest language at the next stage
  • Suggested phrases after practice sessions are calibrated one stage ahead
Benchmark
The AI classification system was benchmarked against 50 expert-coded utterances across 16 language models. The production system uses a three-tier cascade (detailed in Section 11) that achieved 100% accuracy on the benchmark at the final tier.

4. Getting Started

4.1 Account Setup

  1. Download and open GestaltPath. A brief splash screen appears.
  2. Welcome carousel — Four introductory screens explain the app's core value: "Every phrase has meaning," "Understand their language stages," "Get daily coaching tips," "Track progress over time."
  3. Create an account — Enter your full name, phone number, and email address. Alternatively, sign in with Apple.
  4. Verify your phone — A 6-digit code is sent via SMS. Enter it to confirm your identity.

4.2 Profile Configuration

After authentication, two additional steps are required:

  1. Select your role — Choose from Parent, Caregiver, SLP, or Teacher. Each role includes a brief description.
  2. Select your relationship — Options are filtered by role. For SLPs, available options include "Speech-Language Pathologist." For parents, options include biological parent, adoptive parent, foster parent, grandparent, and others.

4.3 Adding a Child

Before using the app, you must add at least one child profile. The only required field is Name (up to 50 characters). Optional but clinically valuable fields include:

FieldNotes
Date of birthUsed to calculate age for display
PhotoDisplayed on home screen and child switcher
DiagnosesMulti-select: Autism, Apraxia, Language Delay, Hearing Loss, Down Syndrome, Other, None
Therapy statusIn therapy, Seeking therapy, Not in therapy, Discharged
Communication methodsMulti-select: Verbal, AAC device, Signs, Pictures, Gestures
Favorite mediaComma-separated list (e.g., "Bluey, Frozen, Wheels on the Bus") — helps contextualize gestalt sources
Estimated NLA stageInitial stage estimate (1–6) — informs the AI cascade routing
NotesFree text (up to 1,000 characters)
SLP Recommendation
When helping a family set up their child's profile, provide an initial NLA stage estimate. This optimizes the AI classification pipeline by routing utterances through the appropriate model tier from the start.

5. Quick Language Logger

The Language Logger is the central data collection tool. It is accessed via the + button (center tab) in the bottom navigation bar.

5.1 Manual Text Entry

  1. Tap the + tab to open the Logger.
  2. Type the child's utterance in the text field (up to 500 characters).
  3. Optionally add context notes (up to 1,000 characters) describing the situation.
  4. Select a routine context: Mealtime, Play, Bedtime, Bath, Car, Outside, School, Therapy, or Other.
  5. Tap "Log & Get AI Suggestion" to save and classify.

5.2 Voice Recording

  1. Tap the red microphone button (large, 80px circle) to begin recording.
  2. A real-time waveform visualization (20 animated bars) confirms audio capture.
  3. A duration timer displays elapsed recording time.
  4. Tap the button again to stop recording.
  5. The audio is sent to OpenAI Whisper for transcription. A spinner indicates processing.
  6. The transcribed text auto-fills the utterance field. You can edit the transcription before submitting.
  7. Proceed with context notes and routine selection, then tap "Log & Get AI Suggestion."
Important
Audio is processed via API and not stored beyond the transcription step. No audio recordings are retained on-device or in the cloud.

5.3 Understanding AI Classification Results

After submission, the Classification Result screen displays:

  • Stage badge — Large, color-coded indicator showing the predicted NLA stage (1–6)
  • Stage name and description — Brief explanation of what the stage represents
  • Confidence indicator — Color-coded: >70% High 40–70% Medium <40% Low
  • AI reasoning — One-sentence explanation of why this stage was selected
  • Linguistic markers — Specific patterns identified (e.g., "echoed media phrase," "novel word combination")
  • Classification tier — Which model in the cascade resolved the classification (T1, T2, or T3)

5.4 Reviewing and Overriding AI Suggestions

Parents have three options after viewing a classification:

ActionWhat Happens
AcceptThe AI's suggested stage is saved as the final classification
AdjustA stage picker opens; the parent selects a different stage (1–6). The override is recorded alongside the AI suggestion
SkipNo stage is assigned. The utterance is saved without classification

All three actions are recorded. When an SLP reviews exported data, they can see the AI's original suggestion/confidence, whether the parent accepted/adjusted/skipped, and whether the utterance was flagged for review.

6. Live Practice Mode

Live Practice Mode transforms the app into a real-time coaching companion during parent-child interactions. It is accessed via the waveform icon (rightmost tab).

6.1 Starting a Session

  1. Tap the Practice tab.
  2. If this is your first session, a consent screen explains that the microphone will be active. Accept to continue.
  3. Select an activity context from 13 options: Play, Mealtime, Bedtime, Bath, Outside, Reading, Car, Dressing, Music, Therapy, Screen Time, Snack, or Other.
  4. Tap "Start Session".

6.2 During a Session

The session screen displays:

  • Session timer — Elapsed duration with a red recording indicator dot
  • Pause/Play/Stop controls — Pause without ending, resume, or stop entirely
  • Live utterance feed — As speech is detected, each utterance appears with transcribed text, a stage badge, timestamp, and coaching hint
  • Speaker toggle — Tap any utterance to toggle between "Child" and "Me" (caregiver)
  • Coaching hint banner — A yellow-highlighted area at the top with the current coaching suggestion
How VAD Works
The app uses Voice Activity Detection to automatically segment speech. It listens continuously, detects when someone begins speaking, records until 1.5 seconds of silence, then processes that segment. Segments between 0.5 and 15 seconds are captured.

Offline support: If the device loses connectivity, an orange banner reads "Offline — transcription unavailable." The session continues recording locally and will process when connectivity returns.

Session maximum duration: 30 minutes.

6.3 Post-Session Summary

When you stop a session, a summary screen presents session duration, total utterance count, stage distribution, all utterances with classifications, and AI-generated coaching phrases with target stages and explanations.

Background Reclassification
During live practice, only Tier 1 (the fastest model) is used for speed. After the session ends, utterances are automatically reclassified through the full cascade for improved accuracy.

7. Gestalt Library

The Gestalt Library is a running collection of the child's known gestalts — scripted phrases, echoed chunks, and recurring language patterns. It is accessed via the books icon tab.

7.1 Browsing and Searching

  • Search bar — Full-text search across gestalt text, source, and meaning
  • Stage filter — Show only gestalts at a specific NLA stage (1–6)
  • Source filter — Filter by TV/Movie, Song, Book, Caregiver, Peer, or Other
  • Child switcher — Switch between children if multiple profiles exist

Each gestalt card displays the text in quotes, a color-coded stage pill, the interpreted meaning (prefixed with "→"), and tags for source category and communicative function.

7.2 Adding a Gestalt

FieldLimitNotes
Gestalt text300 charsThe phrase as the child says it (required)
Source200 charsWhere it came from (e.g., "Frozen," "Dad at bedtime")
Source categoryPickerTV/Movie, Song, Book, Caregiver, Peer, Other
Meaning500 charsWhat the child means when they say it
Communicative functionPickerRequesting, Protesting, Greeting, Expressing Emotion, Commenting, Labeling, Answering, Initiating, Narrating, Other
NLA StageAI or manualTap "Classify with AI" or select manually

7.3 Managing Entries

  • Tap a gestalt card to view its full detail screen
  • Long-press to access the context menu: Mark Active/Inactive or Delete

8. Daily Coaching Prompts

The app generates up to 3 personalized coaching prompts per day per child. These appear on the Home screen below the recent utterances section. Each coaching prompt includes:

  • Prompt text — A specific, actionable suggestion (e.g., "During mealtime, try narrating your actions: 'I'm pouring the milk. Pour, pour, pour.'")
  • Explanation — Why this strategy is effective for the child's current stage
  • Target stage — The NLA stage being modeled (always one stage ahead)
  • Example phrases — Concrete language examples the parent can use
  • Routine context — When to use this prompt

Parents can rate each prompt from 1 to 5 stars. Prompts are generated based on the child's current primary NLA stage, recent utterance history, and daily routine contexts.

9. Progress Dashboard

The Progress Dashboard provides visual analytics on the child's language development. It is accessed via the chart icon tab.

9.1 Stage Trends Chart

A time range picker lets you select 7 Days, 30 Days, 90 Days, or All Time. The chart displays daily utterance counts (blue bars), a rolling average trend line (green), and an overall average (dotted gray). This answers: "Is the family logging consistently?"

9.2 Stage Distribution

A horizontal bar breakdown shows the percentage of utterances at each NLA stage, using the app's ocean-to-shore color gradient. This answers: "What is the child's current stage profile, and is there evidence of progression?"

9.3 Milestones

MilestoneIconTrigger
First utterance loggedStarFirst entry saved
First Stage [N] utteranceTrophyFirst time a new stage is detected
50 / 100 / 500 utterancesFireCumulative logging milestones

10. Settings and Data Management

Settings are accessible from the gear icon on the Home screen.

Children

View all child profiles, add additional children (subject to plan limits), or edit existing profiles.

Appearance

Auto (follows system), Light, or Dark.

Plan Management

PlanChild LimitNotes
Trial10 childrenTime-limited; expiration date shown
Pro10 childrenPaid subscription
ExpertUnlimitedPaid subscription

Data Export

Tap "Export All Data" to generate a comprehensive JSON file containing all utterances, gestalt library entries, coaching prompts with ratings, and practice session summaries. This file can be shared via email, AirDrop, or saved to Files.

For SLPs
The JSON export is the primary mechanism for sharing data between families and their SLP. Encourage families to export and share regularly.

Account Management

  • Sign Out — Logs out of the current session
  • Delete Account & All Data — Permanently removes the account and all associated data. Requires double confirmation. Irreversible.

11. Understanding the AI Classification System

11.1 Three-Tier Cascade Architecture

GestaltPath uses a cascading AI classification system designed to balance speed, accuracy, and cost:

Utterance submitted │ Tier 1: Google Gemini 3 Flash (Handles Stages 1-3, >75% confidence) │ If Stage 4+ or confidence < 75% ↓ Tier 2: xAI Grok 4.1 Fast (Handles Stages 4-5, >70% confidence) │ If Stage 6 or confidence < 70% ↓ Tier 3: Anthropic Claude Opus 4.6 (Final arbiter, all stages)
Why This Matters Clinically
Stages 1–3 have more distinct linguistic markers (echolalia, chunk mixing, single-word isolation) and are reliably classified by faster models. Stages 4–6 require nuanced judgment about whether language is truly self-generated versus recombined from gestalts — these are escalated to more capable models.

In benchmark testing: Tier 1 achieved 100% accuracy on Stages 1–3. Tier 3 achieved 100% accuracy across all stages. The cascade resolves approximately 60–70% of utterances at Tier 1, 20–25% at Tier 2, and 5–10% at Tier 3.

11.2 Confidence Levels and What They Mean

LevelThresholdDisplayAction
High>70%Single stage suggestion (green)Parent accepts or adjusts
Medium40–70%Two candidate stages (orange)Parent selects between the two
Low<40%Flagged for SLP review (red)Parent should consult with SLP

11.3 SLP Review Flags

When AI confidence falls below 40%, the utterance is automatically flagged. This typically occurs with:

  • Ambiguous Stage 2 vs. Stage 4 utterances — The distinction between recombined gestalt chunks and novel word combinations is clinically subtle
  • Context-dependent classifications — The same surface form can represent different stages depending on whether it is echoed or self-generated
  • Transitional utterances — Language produced at the boundary between two stages
Review Checklist
When reviewing flagged utterances, check: the AI's reasoning, the alternative stage proposed, the linguistic markers identified, and the parent's context notes.

12. NLA Stage Reference

This reference is embedded in the app's AI system and used for all classifications.

S1
Stage 1 — Whole Gestalts (Echolalia)
Whole, unmodified chunks from external sources
Key markers: Exact or near-exact echoes; multi-word phrases as single meaning units; preserved intonation; recognizable source.
"To infinity and beyond!" → meaning: "I'm excited!"
Modeling guidance: Model mitigated gestalts (Stage 2) — mix and match parts of the child's known gestalts.
S2
Stage 2 — Mitigated Gestalts (Mixing & Matching)
Chunks broken apart and recombined from different sources
Key markers: Chunks from different sources combined; partial gestalts mixed; the combination is new; may sound choppy or telegraphic.
"Go + want cookie" → parts of two scripts combined
Modeling guidance: Model single words extracted from known gestalts (Stage 3).
S3
Stage 3 — Single Words Isolated from Gestalts
Individual words extracted and used standalone
Key markers: Single words used meaningfully; traceable to known gestalts; one word at a time; referential or labeling/requesting use.
"park" → extracted from "Let's go to the park!"
Modeling guidance: Model novel two-word combinations (Stage 4).
S4
Stage 4 — Novel Two+ Word Combinations
Self-generated, original phrases — not from prior gestalts
Key markers: Two or more words combined in a novel way; not traceable to a known gestalt; emerging word order; short but original.
"want park" → not from a script
Modeling guidance: Model complex sentences with developing grammar (Stage 5).
S5
Stage 5 — Complex Original Sentences
Longer, grammatically complex self-generated sentences
Key markers: Full sentences with developing grammar; question forms; negation; adjectives/adverbs; clausal complexity; some grammatical errors.
"I want to go outside now"
Modeling guidance: Model advanced language with conditionals and abstract reasoning (Stage 6).
S6
Stage 6 — Advanced Self-Generated Language
Complex, grammatically sophisticated language
Key markers: Conditional/hypothetical language; past/future reference; embedded clauses; mental state verbs; narrative and reasoning; near-typical grammar.
"If it stops raining, maybe we could go to the park after lunch?"
Modeling guidance: Continue supporting flexible, complex language use.

13. Clinical Workflow Recommendations

13.1 Recommending GestaltPath to Families

  1. Set the right expectations. GestaltPath is a coaching and tracking tool, not a diagnostic instrument or replacement for therapy.
  2. Help with initial setup. Provide an initial NLA stage estimate when the family adds their child.
  3. Fill in the clinical fields. Encourage families to complete the diagnoses, therapy status, and communication methods fields.
  4. Populate the Gestalt Library early. Work with the family to enter the child's known gestalts with their sources and meanings.
  5. Explain the "model one stage ahead" principle. A five-minute explanation of the NLA stages during a session can significantly improve engagement.

13.2 Using Exported Data in Treatment Planning

The JSON export (Settings → Export All Data) contains the utterance log, gestalt library, coaching interactions, and practice sessions. Recommended uses:

  • Review the utterance log to identify patterns (e.g., consistent Stage 2 during play but Stage 1 during mealtimes)
  • Use stage distribution over time to document progress for insurance or school reporting
  • Identify utterances where the parent adjusted the AI's suggestion — these highlight areas worth discussing
  • Review flagged utterances together to build the parent's NLA staging skills

13.3 Interpreting Progress Reports

  • Daily counts — Inconsistent logging may indicate the family needs encouragement. Consistent logging (even low volume) is more valuable than sporadic bursts.
  • Rolling average trend — An upward trend typically correlates with increased parental engagement.
  • Stage distribution shifts — Look for gradual movement toward higher stages over weeks/months. Sudden jumps may indicate breakthroughs or changed staging accuracy.
  • Milestones — "First Stage [N] utterance" milestones are clinically significant. Cross-reference the date with therapy sessions.

13.4 When AI Flags for SLP Review

ScenarioWhat to Look For
Stage 2 vs. Stage 4 confusionIs the utterance composed of recombined gestalt chunks (Stage 2) or genuinely novel combinations (Stage 4)? Check the Gestalt Library for known component parts.
Context-dependent stagingThe same surface form (e.g., "want juice") could be Stage 1 (echoed), Stage 3 (isolated words), or Stage 4 (novel combination). Parent context notes are critical.
Transitional languageUtterances at stage boundaries often have low confidence. These are often the most clinically interesting data points — they may signal the child is beginning to transition.
Recommendation
Establish a routine (weekly or biweekly) where the family shares flagged utterances with you. This can be done asynchronously via data export or during sessions.

14. Privacy, COPPA Compliance, and Data Security

GestaltPath was designed with pediatric data privacy as a primary concern:

RequirementImplementation
Data encryption at restAES-256
Data encryption in transitTLS 1.3
Audio storageAudio is never stored. Streamed to transcription API, transcribed, and discarded.
COPPA complianceParental consent required during onboarding. Data collected from parents about their child — not from the child directly.
Data minimizationOnly utterance text, context notes, and classifications stored. No biometric, location, or behavioral tracking.
Right to deletionParents can export all data and permanently delete their account at any time.
Live Practice ModeMicrophone only active during explicitly started sessions. Persistent recording indicator visible. Separate consent flow required.
Row-Level SecurityAll database tables enforce RLS. Parents can only access their own children's data.

15. Troubleshooting

IssueSolution
AI classification failsCheck internet connectivity. Tap "Retry." The utterance is saved locally regardless.
Poor voice transcriptionReduce background noise. Speak clearly and at a moderate pace. Edit the transcription before submitting.
Live Practice misses utterancesEnsure microphone is unobstructed. App requires ≥0.5s of speech followed by 1.5s silence. Overlapping speech may not segment correctly.
Sync error bannerData is stored locally first. Tap "Retry" or dismiss — data syncs automatically when connectivity improves.
"Offline" during Live PracticeDevice lost internet. Session continues; transcription resumes when connectivity returns.
Cannot add more childrenTrial and Pro plans support up to 10 children. Upgrade to Expert for unlimited.
"SLP Review" flag appearsAI confidence was below 40%. Review with a qualified SLP. See Section 11.3.

16. Glossary

TermDefinition
GestaltA whole chunk of language acquired as a single unit, typically echoed from an external source
Gestalt Language Processor (GLP)A child who acquires language by learning whole phrases first, then gradually breaking them apart into smaller units
NLANatural Language Acquisition — Marge Blanc's 6-stage framework for how gestalt language processors develop language
Mitigated gestaltA Stage 2 utterance where parts of different gestalts are mixed and recombined
EcholaliaRepetition of words or phrases heard from others; in GLP, these are Stage 1 gestalts that carry communicative intent
VADVoice Activity Detection — technology that automatically detects when speech begins and ends
Cascade / Three-tier cascadeThe AI classification pipeline that routes utterances through progressively more capable models based on complexity and confidence
Confidence scoreA 0–100% rating indicating how certain the AI is about its stage classification
SLP review flagAn automatic marker on utterances with <40% AI confidence, indicating the need for professional review
Modeling (language modeling)The practice of demonstrating language at the appropriate developmental level for the child to hear and produce
Tier 1 / Tier 2 / Tier 3The three AI models in the cascade: Gemini 3 Flash (fast, Stages 1–3), Grok 4.1 (mid, Stages 4–5), Claude Opus 4.6 (comprehensive, all stages)
GestaltPath is not a medical device and does not provide clinical diagnoses. All AI classifications are suggestions intended to support — not replace — professional speech-language pathology services. Families should always work with a qualified SLP for clinical assessment and treatment planning.