Product brief

Status: pre-release (v0.3 in review; v0.2.0 is the latest source release). Built and tested in CI on the iOS 26 simulator; not yet on the App Store and without users. Numbers in this document come from the repository's own eval and CI, and say so; targets are marked as targets.

v0.3 at a glance

v0.3 (pull request "PocketBrains v0.3: honest generalization, tool trimming, RAG eval, docs site") is a measurement release. It makes the generalization number trustworthy before improving it, prepares the Foundation Models brain for a smaller tool budget, and gives "ask your notes" an eval large enough to mean something.

v0.3 metrics (from CI, deterministic brain and trimmer only)

"Before" is the v0.3 branch with the new splits committed and the router unchanged. Not measured: anything about Foundation Models or MLX, and whether trimming helps the model; those need a device (roadmap).

v0.3 trade-offs

• A lower headline number on purpose. The new held-out split scored 30.2% before any work, below v0.2's 35.3% on its own held-out split. Reporting the new split as the headline makes the product look weaker this release and is the only way the next number means anything.
• Constructions, not sentences. The router gained frames for general constructions ("X is done", "can X slip to Y", "Idea: …") rather than rules for each dev miss. Dev reached 100%; the held-out split moved 9.4 points. The remaining 32 held-out misses are left to the model tiers rather than chased with rules that would only fit this corpus.
• Trimming ships off. Recall@8 is 96.3%, so roughly 1 corpus case in 27 would be offered a set without the right tool, and the trimmed session drops multi-turn context. Without on-device evidence that the smaller schema helps the model, the safer default is the full catalog.
• "Push the invoice to Monday" is a reschedule. The v0.1 expectation (updateTask) predates rescheduleTask; the expectation was corrected and the reason recorded, rather than keeping a known-wrong test or bending the router to it.

v0.2 at a glance

v0.2 (pull request "PocketBrains v2: multi-step agent, cited note Q&A, system integrations, smarter tools") moves the agent from "one sentence, one tool call" to "one request, a checked plan", and makes every action reversible.

v0.2 metrics (from CI, deterministic brain only)

Not measured: anything about Foundation Models or MLX (planning accuracy, grounded-answer faithfulness, latency). Those need hardware and are the first roadmap item.

v0.2 trade-offs

• Opt-in system data (EventKit) versus the privacy story. Calendar read and Reminders write are the first time PocketBrains touches data it did not create. Both are off by default, explained on a dedicated screen before the iOS prompt, and requested as separate permissions. The app still makes no network calls. But if the user's calendars or reminder lists sync through iCloud (or accounts such as Google or Exchange added to iOS), reminders that PocketBrains exports leave the device through the system's sync, under the user's account settings, and calendar events it reads may have come from those services. The Integrations screen says so. Calendar data is read per request and never stored in PocketBrains' database; only the identifiers of exported reminders are kept, so an export can be undone.
• Full-access calendar permission. iOS 17+ offers write-only calendar access, but reading agenda context requires full access. PocketBrains asks for the least that delivers the feature and never writes events.
• Undo instead of confirmation. The agent acts first and makes every action reversible, rather than asking "are you sure?" on each step. That keeps capture fast; the journal is the safety net. There are still no delete tools.
• Plans stop on first failure and keep completed steps. Completed steps stay applied (and are undoable as a group) rather than rolling back automatically; a step that ran but failed its postcondition is reverted before stopping. Auto-rollback would hide partial progress the user may want.
• Tool count versus a ~3B model. 22 to 24 tools is a lot of schema for a small context window; integration tools are offered only after opt-in. (v0.3 added per-request trimming behind a switch.)
• Widget reads a snapshot, not the database. Moving the SwiftData store into the App Group would be a data migration with real risk; a small JSON snapshot written after each task change is enough for a glanceable widget.
• Held-out honesty over a higher number. The deterministic router was tuned on dev paraphrases only; the held-out split shows what that buys off-grammar (35.3%), and the rest is the model tiers' job.

Problem

People who run their own work (independent consultants, founders, product and project leads, students with heavy course loads) keep their tasks, project status and notes in three or four apps. Capturing is easy; connecting is not. "What's blocking the website?" means opening the task app, finding the project, reading dependencies, then searching notes for context. The cloud assistants that can answer that question in one sentence need the user's entire work life on someone else's servers, which many of these users either cannot accept (client confidentiality, NDAs, regulated work) or do not want.

On-device language models in iOS 26 change what is possible: a capable model that runs locally, with native tool calling. The gap is a product that uses it to act on real structured data, works when the model is not available, and makes the privacy guarantee verifiable instead of a marketing line.

Target users

Primary persona for v1: an independent professional on a recent iPhone who captures on the move and plans at a desk, and who will not upload client material to a chatbot.

Jobs to be done

1. Capture without filing. "Remind me to send the invoice Friday" becomes a dated task in the right project, from one sentence.
2. Know where things stand. "What's blocking the website redesign?" returns blockers, what they wait on, and the next milestone.
3. Start the day oriented. A morning brief and an agenda of overdue, due, blocked and this week's work.
4. Find what I already know. Semantic and keyword search across notes; "summarize my notes from yesterday".
5. Connect ideas to work. Link a note to a project; new notes are linked automatically to the projects and tasks they mention.
6. Trust it. Know, and be able to check, that nothing leaves the phone.

Scope

In (built on main)

• One conversational thread over a local SwiftData store, with a 24-tool agent (v0.1 shipped 13; v0.2 added planning, undo, cited note Q&A, recurrence, scheduling and opt-in integrations).
• Three-tier brain: Apple Foundation Models, optional MLX Qwen3 build, and a deterministic intent router that always works.
• Spaces: Today, Projects and a Knowledge constellation, sharing one write path with the agent.
• Morning brief and evening reflection, due-date notifications, Siri and Shortcuts intents for capture, full JSON export.
• Deep Glass design system with Metal shaders and Reduce Motion support.
• Privacy manifest declaring no tracking and no collected data.

Out (deliberately)

• Accounts, sync, collaboration and any server component.
• Cloud model fallback. If no on-device model is usable, the deterministic router answers; the app never sends a prompt off the device.
• Email ingestion. (Calendar read shipped in v0.2 as an explicit opt-in; see "v0.2 trade-offs".)
• Non-English input for the deterministic router.

Requirements

• Every agent action goes through a tool and is shown as a card; the model may not claim an action it did not perform (AgentVoice instructions).
• One write path (DataServices) for agent, views and intents.
• A turn never dead-ends: model failure falls back to the deterministic router; a failed turn offers retry.
• Works in Airplane Mode with full functionality of the active brain.
• Store failures never silently drop data: the broken store is backed up before falling back.

Success metrics and evals

The accuracy floor in CI is set just below the measured canonical score so a routing regression fails the build. Paraphrase results are reported but not gated, because improving them with more rules would overfit the corpus; that is what the language-model tiers are for.

Minimum viable quality

Thresholds for a first TestFlight build, using the metrics above and consistent with the v1 targets. "Current" is CI run 37723402205, deterministic brain only. Foundation Models rows cannot be scored until issue #5 lands, so no v1 build ships before they are.

The "Do not ship" line for the held-out split protects the measure: it fails only if a change makes the floor worse than before v0.3, not if the number stops rising. "Ship" on the best available brain is the gap the model tiers have to close.

Cost at 1x and 10x usage

Estimates. They use only this repository's own assumptions: inference runs on the user's device, there is no backend, no account system and no per-request API (see "Local versus cloud"), CI runs on GitHub-hosted runners, and App Store distribution needs an Apple Developer Program membership ($99 a year, docs/LAUNCH.md). "1x" is a first TestFlight cohort; "10x" is ten times the users and requests. The repository makes no assumption about absolute user counts, so none is invented here.

The cost a cloud assistant would scale with engagement (tokens per request times requests) is zero here at any usage level. What grows with usage is support and App Store review work, which the repository does not estimate, and the on-device cost each user carries, which issue #5's device runs should measure (time to first token, and ideally energy per request).

Local versus cloud: trade-offs

The decision follows from the target user: for someone under NDA, a stronger cloud model they cannot use has no value. The capability gap is narrowed by keeping the model's job small (pick a tool, fill arguments, phrase a short reply) while deterministic code does the data work.

Key decisions

Risks

Roadmap

Now

• Run the 243-utterance corpus (including both held-out splits and the compound requests) against the Foundation Models brain on device; publish tool and plan accuracy, and time to first token (issue #5).
• Validate tool trimming on device, trimming on versus off, then choose its default (issue #6).
• Measure cited-answer faithfulness of model-written answers on a small hand-labelled set (issue #7).
• Fix the O(n²) blocker lookup flagged in docs/QA_REPORT.md (issue #10).

Next

• Dynamic Type mapping and a full VoiceOver pass (issue #8).
• Localization of UI strings; multilingual routing via the model tiers.
• Interactive widget (complete a task from the widget through App Intents, issue #9).

Later

• iCloud private-database sync as an explicit opt-in, with the trade-off stated in the privacy story.
• TestFlight and App Store release, subject to the licensing and naming decisions recorded in the pull request that introduced this document.