RUNBOOK Press Resources
Quick facts, description copy, real screenshots from the current build, downloadable brand assets, colors, typography, and contact information for press and media inquiries.
App Name
RUNBOOK
Platform
iOS — iPhone 13+
Category
Reference / Utilities
Price
Free — three packs included
Developer
Justin Tormey
Status
Pre-launch · TestFlight coming soon
Language Model
Qwen3-4B-Instruct-2507 · Apache 2.0
Embeddings
nomic-embed-text-v1.5
Inference
100% on-device. No cloud API, no account, no analytics in the app.
Refusal Gate
Similarity 0.65 — below it the model is never called
Packs at Launch
3 free — 65 documents · 330 chunks
Accessibility Modes
Night Mode (red-on-black) · Hands-Free (on-device voice in, spoken answers out, screen awake)
App Description Copy
One line
An offline iOS knowledge app that only answers from the packs you've downloaded — real sources, cited by name, or it says so and stops.
Tagline
Zero bars. Full answers.
Short — 50 words
RUNBOOK answers emergency and first-aid questions entirely on your phone, with no signal and no account. It reads only the knowledge packs you've installed — Army field manuals, Ready.gov, NOAA/NWS — cites the document by name, and refuses when the answer isn't there.
Three paragraphs
RUNBOOK is an offline iOS knowledge app with a built-in constraint: it only answers from knowledge packs you've downloaded. If a question doesn't match pack content above a similarity score of 0.65, the language model is never invoked at all — RUNBOOK says "not in your packs" and points at the nearest documents instead of guessing.
There is one screen and one way to ask. Type a question and RUNBOOK answers in plain prose; describe an emergency in progress and the same prompt leads with the most critical action in terse numbered steps. No mode to pick, nothing to configure under pressure. Night Mode filters the whole interface to red on black so you keep your dark-adapted vision, and Hands-Free lets you ask out loud, keeps the screen awake, and reads the answer back — on-device speech recognition in, on-device speech synthesis out, radios off the whole time. A built-in reader lets you browse the full source documents inside any installed pack, not just the answers drawn from them.
RUNBOOK launches with three free packs — Basic First Aid, Emergency Basics (US), and Weather Lite — built from U.S. Army field manuals, the U.S. Fire Administration, FEMA/Ready.gov, and NOAA/NWS. The model that ships, Qwen3-4B-Instruct-2507, is the one that survived a 40-question graded safety battery that disqualified four others. No account, no cloud sync, no internet required after download.
Four screens from the current build
These are real captures from the shipping build — not mockups, renders, or composites. Nothing has been retouched, and no text has been substituted. Free to reproduce in coverage of RUNBOOK.
Brand Assets
The app icon file above is the exact asset that ships in the app bundle — the |RB mark: a green bar, white RB, solid black. Please don't recolor, rotate, or add effects to any of these assets, and don't place the wordmark on a background that isn't #0B0D0B or #F2F4F2.
Brand Colors
Typography
Display / UI (sans, human voice)
-apple-system, BlinkMacSystemFont
San Francisco on Apple devices. System UI fallback elsewhere. Used for prose — the parts a person wrote.
Monospace (machine voice: sources, versions, status)
SF Mono, Cascadia Code, Menlo
Used for citations, document titles, pack versions, status labels, and the RUNBOOK wordmark. If a machine asserted it, it's set in mono.
Four models were disqualified before one shipped
Every candidate model ran the same 40-question graded safety battery — in-pack questions, must-refuse questions, emergency-voiced questions, and adversarial questions — hand-graded against the source documents. These are the failures that ended each candidacy.
- Phi-4 Mini — disqualified Blended heat-injury steps into a hypothermia answer — the opposite treatment for the opposite emergency — under cross-topic retrieval pressure.
- Granite 4.0 Micro — disqualified Asked what to do for shock, it returned the heatstroke cooling protocol, then cited the shock document as its source for a procedure that document does not contain.
- Gemma 3 4B QAT — disqualified Asked about a choking infant, it gave the adult procedure, with no acknowledgment that infants need a different technique entirely.
- Gemma 4 E4B — never graded Ruled out on footprint before the battery: it would not fit in 8 GB alongside the embedding model and the app.
- Qwen3-4B-Instruct-2507 — shipped Correct shock-versus-heat discrimination. Respected victim-type boundaries — it refused the infant-choking question rather than improvising an adult procedure. Zero fabricated citations across the battery. Apache 2.0, and it runs entirely on the phone.
The full roster and the battery itself live in the repo (eval/rag_eval.py). See the Trust Story on the home page for the long version.
Press Contact
For interview requests, review access, or asset requests, reach out directly.
Key Angles
- An offline iOS knowledge app that disqualified four AI models on a graded safety battery — one gave heatstroke cooling for a shock question and cited the shock document for it; another gave a choking infant the adult procedure — before shipping the one that didn't fail.
- On-device inference, fully offline after pack download — no cloud API, no account, no subscription, no analytics in the app.
- Every answer names its source document — Army field manuals, U.S. Fire Administration, FEMA/Ready.gov, NOAA/NWS — and a built-in reader lets you open that document in full.
- A hard similarity gate at 0.65 means the app refuses to guess — below the threshold the model is never called at all. "Not in your packs" is a designed outcome, not an error state.