Questions, evidence, and the next iteration
Living engineering notebook
Questions become evidence
The Learning Lab exists to turn practical voice-engineering questions into inspectable experiments, working implementations, reproducible evidence, and the next better question.
01 · Why the lab exists
The repository describes itself as an executable, local-first learning, diagnostic, and solution-design environment.
README.md#why-this-existsThe current product map preserves prototype, evidence, and production-readiness boundaries module by module.
chatgpt-project-pack/01_CURRENT_PRODUCT_MAP.md02 · Recursive learning loop
Name the user, engineering, or evidence gap.
Inspect the repository, project boundaries, and current primary sources.
Implement the smallest useful, understandable experiment.
Capture behavior, timing, lifecycle events, and failure signals.
Exercise deterministic, browser, security, and bounded live boundaries.
Record evidence, limitations, provenance, and what remains unknown.
Commit, review, and deploy only the evidence-supported change.
Use observed evidence to choose the next hypothesis.
Question again returns the evidence to the start of the loop.
03 · Current architecture
A person supplies the question, priorities, constraints, and review decision.
Repository verifiedThe implementation partner inspects evidence and changes the scoped working tree.
Repository verifiedLocal changes remain inspectable before they become repository history.
Repository verifiedA reviewed snapshot records a logical implementation or evidence change.
Repository verifiedCanonical source control and review history for the current architecture.
Repository verifiedDeployment infrastructure; a preview or production result still requires independent verification.
Repository verifiedThe deployed experience exposes only the reviewed runtime and configured provider boundaries.
Repository verifiedTests, observations, provider events, and human review reveal what happened.
Repository verifiedEvidence narrows the next question instead of silently promoting an assumption.
Repository verifiedThe repository remote targets omnineuralengine/ONE-voice-lab on GitHub.
git:remote:originThe cross-device handoff records the GitHub-to-Vercel preview workflow and its review boundary.
docs/CROSS_DEVICE_HANDOFF.md#vercel-continuityEntire is represented only as a parallel context layer; no checkpoint identifier is present in tracked evidence.
Lab Evolution architecture boundary04 · Evidence philosophy
Implementation, tests, or tracked repository history support the claim. This is not production certification.
chatgpt-project-pack/01_CURRENT_PRODUCT_MAP.md#evidence-labelsCurrent official Deepgram documentation supports the API claim. This does not prove account entitlement.
chatgpt-project-pack/05_DEEPGRAM_CAPABILITY_MAP.mdAvailable evidence is incomplete, so the claim remains explicitly provisional.
chatgpt-project-pack/01_CURRENT_PRODUCT_MAP.md#evidence-labelsThe direction is incomplete, disabled, or future-facing and must not be presented as working behavior.
chatgpt-project-pack/01_CURRENT_PRODUCT_MAP.md#evidence-labels05 · Current module maturity
Implemented · Prototype
Remove visitor account and API-key friction while keeping execution explicit, bounded, and server-funded.
Implemented · Review
Orient visitors to Lab status, workflows, privacy boundaries, and explicit learning recipes.
Implemented · Review
Check server-side Deepgram configuration without exposing the permanent key or account metadata.
Implemented · Review
Make a guarded prerecorded URL request inspectable from input through transcript and evidence.
Implemented · Review
Let a user inspect a local audio file before an explicit, validated transcription request.
Implemented · Prototype
Teach microphone consent, short-lived authentication, realtime events, failure handling, and cleanup as one lifecycle.
Implemented · Review
Generate and inspect bounded existing /v1/speak synthesis without conflating it with Flux TTS.
Implemented · Prototype
Compare voice-agent-first /v2/speak synthesis through explicit batch generation, A/B playback, and sanitized evidence.
Implemented · Prototype
Explore consent, disclosure, opt-out, fallback, and narrowly gated synthetic-voice patterns.
Implemented · Review
Provide reviewed sample audio with provenance for repeatable learning flows.
Implemented · Review
Verified Nova-3 language exploration and safe handoffs.
Implemented · Review
Make transcript privacy choices and downstream policy consequences inspectable.
Implemented · Review
Synthetic audio inspection and signal learning tools.
Implemented · Review
Registry-backed request inspection and explicitly guarded execution.
Implemented · Prototype
Connect isolated voice APIs to inspectable client architectures, failure experiments, and readiness questions.
Implemented · Review
Turn voice-system concepts into deterministic exercises, debugging paths, and handoffs.
Implemented · Prototype
Compare STT, TTS, realtime, and round-trip evidence with explicit provenance and lifecycle inspection.
Implemented · Review
Provide editable, reviewable integration recipes with credential placeholders and explicit handoffs.
Implemented · Prototype
Show how questions become experiments, implementations, evidence, delivery, and the next iteration.
Implemented · Review
Paste or structure a live technical problem and produce a deterministic field brief.
Experimental · Prototype
Typed Flux /v2 conversational-turn observability with deterministic replay; live provider validation remains manual.
Implemented · Review
Collaborative architecture discovery, diagnostics, and handoff.
Implemented · Review
Deterministic discovery, recommendation, and POC planning.
Implemented · Review
Compile an approved case into validated customer materials.
Implemented · Review
Mobile/PWA navigation and guarded call-side tools.
Implemented · Review
Optional safe semantic keyboard companion.
Implemented · Review
Versioned evidence, decision, question, risk, validation, and action intelligence.
Implemented · Review
Redacted payload and code inspection without execution.
Implemented · Review
Deterministic SDK generation and version diagnosis.
Partial · Prototype
Registry and release evidence are present; a standalone radar UI is not complete.
Implemented · Review
Truthful provider, configuration, capability, and adapter visibility.
Implemented · Review
Private provider-neutral fixture and protected-live TTS comparison with versioned evidence.
Experimental · Prototype
Typed deterministic Voice AI failure replay with explicit execution, simulated provenance, local usage facts, and bounded scorecards.
Partial · Prototype
Public infrastructure and a safe empty state are implemented; no public early-access experiment is currently configured.
Implemented · Review
Server-only official documentation retrieval with curated fallback.
Experimental · Prototype
Gateway-backed explanation, critique, routing, red-team, and POC proposals layered above deterministic Lab evidence.
Experimental · Prototype
Ephemeral per-session AI request metadata without raw prompts or generated content.
Planned · Prototype
Voice capture for problem summaries is not yet implemented.
06 · Evolution timeline
The first repository commit established the local-first control room, core voice workflows, diagnostics, learning paths, and evidence-aware handoff material.
Git commitf6017b4c206c357088cd1424b97a5c0c6d6f9b26
The baseline README records 101 unit tests and 33 focused browser tests passed.
git:f6017b4c206c357088cd1424b97a5c0c6d6f9b26:README.md#current-verification-statusThe root commit records the initial Deepgram Applied Voice Lab baseline.
git:f6017b4c206c357088cd1424b97a5c0c6d6f9b26The baseline README records the local-first architecture, module map, evidence labels, and verification boundary.
git:f6017b4c206c357088cd1424b97a5c0c6d6f9b26:README.mdThe repository added progressive discovery, deterministic recommendations, editable architecture, failure diagnosis, rehearsal, and privacy-aware handoff behavior.
Git commit82e2cb00b1939db91e1e89ae8fc1fbabfb2cea5e
The implementation commit and module documentation record the interview-oriented Architecture Studio expansion.
git:82e2cb00b1939db91e1e89ae8fc1fbabfb2cea5eThe module document describes its deterministic recommendation, diagnostics, handoff, tests, and limitations.
docs/architecture-studio.mdRepository documentation recorded a Vercel preview where the Architecture Studio and health route returned HTTP 200 while hosted cross-device sync remained disabled.
Git commite9edaf28a0066435dfed653e408b604e7ea5a960
The handoff records the preview URL, HTTP results, local-demo health mode, and deployment boundaries.
git:e9edaf28a0066435dfed653e408b604e7ea5a960:docs/CROSS_DEVICE_HANDOFF.mdA deterministic discovery-to-recommendation-to-POC workflow became a dedicated studio with browser and domain coverage.
Git commitdd8a5563927889a72c0ffc72fce20f6fdc27b517
The commit adds the Pre-Sales route, domain engine, components, documentation, and tests.
git:dd8a5563927889a72c0ffc72fce20f6fdc27b517The lab gained a touch-oriented launcher, bounded offline shell, demo checkpoints, and allowlisted local preferences.
Git commit8725eeadde0e5f6b992efb92dfa0016d9d3785b7
The feature commit adds the PWA shell, offline assets, Pocket components, tests, and documentation.
git:8725eeadde0e5f6b992efb92dfa0016d9d3785b7A concise registry-backed assistant added safe examples and handoffs from Pocket into full lab modules.
Git commit79ee31d442a2a40e79a7c06c0ad63c2550896086
The commit adds Pocket API data, components, documentation, and deterministic coverage.
git:79ee31d442a2a40e79a7c06c0ad63c2550896086The release added the Live Solution Case Graph, question and decision tools, Workbench, SDK Doctor, Deliverables Studio, Control Deck, and expanded reviewer evidence.
Git commit2df61df0c4428639fe869dd3781c27c47b76d52d
The implementation commit records the evidence-grounded Live Solution Lab expansion.
git:2df61df0c4428639fe869dd3781c27c47b76d52dThe release changelog lists the Case Graph, ledger, question copilot, deliverables, diagnostics, and review expansion.
git:2df61df0c4428639fe869dd3781c27c47b76d52d:CHANGELOG.mdThe repository added typed Flux /v2/listen turn-event inspection, deterministic replay, local metrics, scorecards, Mermaid handoffs, and Vercel analytics.
Git commitfb7a46961277883b9695bec5a5b96579b731579e
The implementation commit records the Flux Observatory and analytics addition.
git:fb7a46961277883b9695bec5a5b96579b731579eThe module document distinguishes deterministic replay, live-provider mode, and the missing retained live validation record.
docs/FLUX_CONVERSATION_OBSERVATORY.mdThe current working tree adds public Open Lab controls, shared ONE primitives, a sanitized flight recorder, and a registry-driven Flux batch synthesis studio while preserving existing Aura TTS.
Git commit25bcf498ca04cf49be2d1bd2d6188cee6cef9d7a
The implementation evidence index maps the Open Lab, ONE, Flux, and security changes to their source and test files.
chatgpt-project-pack/07_IMPLEMENTATION_EVIDENCE.mdThe Git commit records the Open Lab, ONE visual system, and Flux TTS implementation snapshot.
git:25bcf498ca04cf49be2d1bd2d6188cee6cef9d7aThe capability map records the official Flux /v2/speak contract and dated voice-catalog verification boundary.
chatgpt-project-pack/05_DEEPGRAM_CAPABILITY_MAP.mdThe readiness snapshot records passing mocked coverage and the unsuccessful bounded live authorization smoke without promoting it to provider proof.
chatgpt-project-pack/08_TEST_AND_READINESS_STATUS.mdThe repository adds a structured recursive-learning timeline, explicit GitHub-to-Vercel delivery architecture, an experimental Entire context boundary, and evidence-backed evolution affordances across implemented modules.
Git commit0e6dd24fbeb23de154423961da7edf13442c420a
Final local validation passed 440 npm tests, 76 default browser tests with 6 intentional project guards, and 29 Open Lab tests.
chatgpt-project-pack/08_TEST_AND_READINESS_STATUS.mdThe implementation commit records the Lab Evolution registry, UI, module affordances, responsive visual baselines, and focused tests.
git:0e6dd24fbeb23de154423961da7edf13442c420aThe module document records the evidence labels, canonical delivery path, experimental Entire boundary, and current limitations.
docs/LAB_EVOLUTION.mdThe canonical site adopts the supplied ONE mark, a concise Vercel hostname with continuity redirect, Vercel performance telemetry, and privacy-bounded Supabase product-interest events.
The deployment and analytics document defines the canonical URL, event allowlist, RLS boundary, and data that is deliberately not retained.
docs/ANALYTICS_AND_DEPLOYMENT.mdThe schema and application route restrict analytics to coarse product surfaces and the three registered providers.
supabase/migrations/20260825190232_viewer_analytics.sql07 · Current experiments
The batch implementation and mocked boundary pass, but the bounded Cole and Jack provider smoke returned provider_authorization_failed and no audio.
Next hypothesisA correctly authorized and entitled server credential will permit one bounded Cole request and one bounded Jack request without changing the public credential architecture.
The readiness snapshot records exactly two requests, no retry, safe request IDs, and no successful audio proof.
chatgpt-project-pack/08_TEST_AND_READINESS_STATUS.md#live-provider-statusStreaming remains disabled until the deployed JWT authentication, binary audio, lifecycle, and cleanup path is proven.
Next hypothesisA bounded deployed browser smoke can validate authentication, first audio, lifecycle events, Stop, and complete cleanup before the control is enabled.
Official documentation describes the endpoint, bearer JWT pattern, raw audio, and lifecycle, but documentation is not deployment proof.
chatgpt-project-pack/05_DEEPGRAM_CAPABILITY_MAP.mdThe one-browser local fallback is implemented; optional hosted cross-device behavior still requires configured infrastructure and deployment testing.
Next hypothesisA non-production two-device soak can expose reconnect, expiry, deletion, and authorization gaps without changing the local-first fallback.
Architecture documentation records the local fallback and the unverified shared-session boundary.
docs/architecture-studio.md#known-limitationsEntire is modeled as a parallel observational context layer, not source control and not deployment infrastructure.
Next hypothesisA redaction-aware checkpoint reference may improve implementation reproducibility while GitHub remains canonical and Vercel behavior remains unchanged.
No tracked checkpoint identifier or deployment responsibility is claimed in the current registry.
Lab Evolution architecture boundary08 · Known limitations
The Lab is community-built learning and prototype software, not an official Deepgram product or production-certified service.
The limitation is explicit in the current product map.
chatgpt-project-pack/01_CURRENT_PRODUCT_MAP.mdMocked route and browser evidence passed, but the bounded live smoke returned authorization failures and no audio.
The exact bounded smoke and its non-proof boundary are recorded.
chatgpt-project-pack/08_TEST_AND_READINESS_STATUS.md#live-provider-statusThe shared integration remains disabled until /public/brand/omni-neural-engine-mark.svg is supplied.
The missing path and disabled behavior are documented and tested.
docs/KNOWN_LIMITATIONS.mdSDK Release & Regression Radar is partial, and Speak the Problem remains planned and unavailable.
Both boundaries remain explicit in the product map and capability registry.
chatgpt-project-pack/01_CURRENT_PRODUCT_MAP.md09 · Next hypotheses
Resolve the current authorization boundary, then verify Cole, Jack, playback, download, content type, request ID, and redaction without retries.
The repository does not yet distinguish credential authorization from model entitlement.
chatgpt-project-pack/08_TEST_AND_READINESS_STATUS.md#live-provider-statusValidate the documented browser authentication and raw-audio lifecycle in a deployed environment before exposing a working Streaming control.
The implementation intentionally leaves streaming disabled pending deployment proof.
docs/KNOWN_LIMITATIONS.mdUse non-production configuration for the documented two-device reconnect, expiry, deletion, and authorization test.
The next test is recorded in the Architecture Studio backlog.
docs/architecture-studio.md#post-interview-backlogTest whether a sanitized Entire checkpoint reference improves reproducibility while commits remain canonical in GitHub and deployments continue through Vercel.
No checkpoint is currently asserted; this is a bounded next hypothesis only.
Lab Evolution architecture boundary