Current state
High IQ’s paid notebook is the readable, educational notebook. Optional share cards, mindmaps, audio, images, infographics, and video must never be required for that core experience to finish. Notebook companion execution is currently off. The production registry has zero selected executors, and the provider-evaluation tools are local, synthetic, and unable to make network, storage, analytics, subscription, or persistence calls. NotebookLM is not a product executor and is retained only as historical internal prototype evidence.First-party architecture
TIWIH owns the source projection, evidence rules, recipes, structured briefs, deterministic factual overlays, evaluation corpus, artifact status, storage policy, and mobile experience. A future provider could execute one bounded step, but it cannot own the notebook contract or receive an entire notebook, order, stash, source snapshot, or user profile. The evaluation order intentionally moves from greatest control to greatest uncertainty:- Deterministic TIWIH share cards and structured mindmaps
- Deterministic video composition after evidence and private storage land
- One exact direct image endpoint and one exact direct speech endpoint
- One exact video endpoint, evaluated last
Candidate identity
There is no blanket marketplace approval. A fal.ai or Replicate listing would
need its own exact underlying model/version, license, region, tier, retention,
deletion, and commercial review. There is also no silent fallback: an
unavailable or disallowed endpoint makes the optional artifact unavailable.
Independent hard gates
Every exact endpoint must independently pass commercial rights, regulated-content approval, input training, output rights, retention, deletion, processing region, subprocessors, attribution, moderation, SLA, support, indemnity, and private-storage handling. Each result ispass, fail, or
unknown; unknown blocks evaluation.
Quality, latency, cost, or a successful demo can never compensate for a failed
privacy, safety, rights, storage, or regulated-content gate.
Synthetic-only evaluation
The current harness accepts only bounded synthetic briefs for a share card, mindmap, audio script, image brief, infographic brief, storyboard, or video brief. It rejects private identity fields or values, email addresses, credentials, URLs, clinical or prescriptive copy, Report V1 content, source snapshots, and oversized inputs. A future Report V2 integration may expose only the exact public/shareable projection required by one artifact. It may not send a whole report or source snapshot to a provider. Evaluation packets use versioned, domain-separated SHA-256 identities for content, request, idempotency, and packet data. A future offline receipt can record hashes, usage, cost, retention/delete state, exact provenance, and manual safety review, but no provider artifact URL.Storage and publication
Before any future customer delivery, generated artifacts need private TIWIH-controlled storage, authenticated retrieval, short-lived delivery, deletion, regeneration, retention limits, and auditability. Public provider copies must be copied privately and confirmed deleted. Pending safety review always requires manual publication review.Provider research posture
- OpenAI Images and text-to-speech are provider-generic research subjects; no endpoint or model is approved.
- Luma, Kling, and Runway video research is blocked pending exact identity, written regulated-content confirmation, commercial review, and private storage.
- ElevenLabs remains blocked because its published policy requires prior written authorization for regulated or recreational drug use.
- OpenAI’s current Videos API path is retired from new-integration research because its official documentation says the API is deprecated and scheduled to shut down September 24, 2026.