Work / Intelligent agents / The Keeper's Study

Valerie

What if an assistant could propose anything, but only fixed rules could decide what runs?

Source auditedPrivate repository; source not public

The journal

20 exhibits in three rooms

The Grand Atrium

The hall. Its doors lead to every wing.

  1. PHAROSThe Infinite CatalogBuilt
  2. The Mechanical FoundryA concept workshopPlans only
  3. BABELThe Theatre of StoriesPlans only
  4. The Electric GalleryA concept galleryPlans only
  5. ObservatoryThe roof courtyardBuilt
  6. CognitiveOS / ValerieThe Keeper's StudyBuilt
  7. R&BotThe Noir Recording ChamberPlans only

PHAROS: The Infinite Catalog

Ten machines, one for each thing PHAROS does.

  1. 1. Which repositories has PHAROS really indexed?Pick a repository PHAROS really indexed.RECORDED
  2. 2. What will PHAROS refuse to index?Admit the repository, or refuse it with a reason.RUNS HERE
  3. 3. What does it read from a repository?Read only what Git tracks at that commit.RECORDED
  4. 4. How does it cut code into pieces?Split each file into symbols.RECORDED
  5. 5. What does it turn into vectors, and what does indexing cost?Embed a summary of the code, not the code.RUNS HERE
  6. 6. How does it know who calls whom?Link calls, and leave the doubtful ones unlinked.RECORDED
  7. 7. Which retriever answered, and why that one?Ask, and see which retriever answered.RUNS HERE
  8. 8. What if the code changed after it was indexed?Check the source before handing it over.RUNS HERE
  9. 9. What does an assistant get back?Hand an agent cited, budgeted context.RECORDED
  10. 10. Did it work?What it achieved, and what failed.REPORTED

Valerie: The Keeper's StudyYou are here

Ten exhibits of how her actions are checked.

  1. 1. What does a model actually hand over?The proposal trayRUNS HERE
  2. 2. What is she allowed to point at?The ticket rackRUNS HERE
  3. 3. Can I throw the switch myself?The paired switchesRUNS HERE
  4. 4. What does a refusal do?The refusal trayRUNS HERE
  5. 5. How do I know what really happened?The receipt printerRUNS HERE
  6. 6. What happens in one turn on the Mac?The desk bellSIMULATED
  7. 7. Would the Mac app do this?The risk ledgerSIMULATED
  8. 8. What happens to text from outside?The memory cabinetSIMULATED
  9. 9. What can she do on a Mac?The instrument consoleSIMULATED
  10. 10. What doesn’t this room show?The gap on the shelfWRITTEN

Contents

What’s in this book?

Three passes: a minute, how it works, then the deep dive. Stop after any one. Pick a page to open it in the book, or walk Valerie to a machine in the room.

  1. Pass 1 Overview 1 minute
    1. Who is Valerie?
    2. What did Ram build, and what changed?
    3. What happens to one sentence?
  2. Pass 2 How it works 3 minutes
    1. What’s the idea, in one picture?
    2. What does a model actually hand over?Station 1 RUNS HERE
    3. What is she allowed to point at?Station 2 RUNS HERE
    4. Can I throw the switch myself?Station 3 RUNS HERE
    5. What does a refusal do?Station 4 RUNS HERE
    6. How do I know what really happened?Station 5 RUNS HERE
    7. What happens in one turn on the Mac?Station 6 SIMULATED
    8. Would the Mac app do this?Station 7 SIMULATED
    9. What happens to text from outside?Station 8 SIMULATED
    10. What can she do on a Mac?Station 9 SIMULATED
    11. How do the parts connect?
  3. Pass 3 Deep dive As long as you like
    1. How does each part work, exactly?
    2. How did she get here?
    3. What can I make her do on this site?
    4. Can I try to break the boundary?
    5. Who is she, as a character?
    6. What doesn’t this room show?Station 10 WRITTEN
    7. What are her limits?
    8. What would Ram test next?

Pass 1 of 3

Overview

Who she is, why she can’t overrule the code, and a move to send her. 1 minute

Pass 1 · Overview

Who is Valerie?

Valerie lives inside CognitiveOS, a personal operating layer for my day on iPhone and Mac. She talks, listens, remembers, learns a task by watching it once, and operates the Mac when asked.

Why it matters. An assistant that can press buttons on a real computer needs more than a capable model: whatever decides what it may do has to be something it can’t talk its way past. So the rule that shapes the system is that the model decides intent, and code keeps control. In her own design notes: Ram decides. Valerie interprets, argues, remembers and proposes. The System owns the truth.

  • Built: a native SwiftUI app with a turn pipeline, long-term memory with trust set in code, skills learned by demonstration, and a living pixel-art presence. About 107,000 lines of Swift, in five weeks.
  • Runs here: the same boundary, for this website only. Send it a move below and watch which check stops it. In her dock you can wake a small model fine-tuned for this site, in your own browser: it proposes a move, the same checks judge it, and it runs only when you press it.
  • Limits: the source is private, the native interface was not driven for this site, and memory recall is measured on a public benchmark (retrieval only), not on my own questions.

Pass 1 · Overview

What did Ram build, and what changed?

My contribution

What I built
CognitiveOS for iPhone and Mac, with Valerie inside it: the one-turn pipeline (fast path, one typed move, validation, an existing engine), a deterministic risk policy over 56 declared desktop capabilities, long-term memory with a verbatim ledger and trust tiers, skills learned by demonstration, and her presence. Then the same boundary, adapted, for this site’s guide, with a guard in front of its validator.
The tradeoff I chose
Authority stays with deterministic code. A model may propose one typed move per turn over ids the turn issued. macOS permissions, the app’s settings and a fixed risk policy decide whether it runs, and the model controls none of them. The cost is flexibility: anything outside the closed list of 26 moves simply can’t happen, and risky actions always wait for a yes.
What changed
Every action became checkable: validated against what existed that turn, refused or confirmed by fixed rules, and recorded as a receipt. Outside text, from a screen, a paste or PHAROS, can be discussed but can never act or become a fact. Memory retrieval reports R@5 0.908 on LongMemEval’s 500 questionsreported · retrieval only.

Pass 1 · Overview

What happens to one sentence?

From words to one checked move. Send her one below, and the diagram replays its route.

One sentence to one action

RUNS HERE This site’s guarded executor (packages/valerie-web) runs in this page. No service, model or device is involved.

What happens between a line you type and anything moving on the page. Every check is the real one, in the order it runs. Send a move below and this replays the route your move took.

13 steps
  1. YouType a line, or press a button

    “show me how PHAROS is built”, or a button on the page.

  2. ProposerOne typed move is proposed

    The phrase matcher maps your words to one registered command. If you woke the small model in her dock, the model proposes instead, and you confirm it first. Either way, the proposal is data: a tool and its arguments, never code.

  3. GuardAdmits plain data, and takes a copy

    Only plain objects, arrays, strings and numbers pass: no getters, no functions, nothing that can answer one way to the check and another to the run. The validator checks a detached copy, and that copy is what runs.

  4. ValidatorFrom this session?

    The move must carry the current generation. Cancelling starts a new one, so a late answer from before is refused.

  5. ValidatorA tool on the closed list?

    Nine tools exist on this site. Anything else, such as eval, is refused.

  6. ValidatorExactly these arguments?

    The keys must be exactly the tool’s own. An added CSS selector, for example, is refused.

  7. ValidatorAllowed this turn?

    Each turn names its tools. A model’s turn may not search or prompt R&Bot, because both take free text.

  8. ValidatorOutside text only explains

    A move that came from retrieved text may explain a fact, and nothing else.

  9. ValidatorAn id this site issued?

    Every section, part, diagram and fact is checked against the registry. An id that was never issued, such as personal-mac, is refused.

  10. ExecutorAccepted, one action at a time

    If another action is still running, this one waits its turn or is refused. Nothing runs in parallel.

  11. AdapterLooks the target up, then acts

    The adapter finds the real place for the id in the registry at the moment it runs, and scrolls, opens or highlights it. The model never sees a URL or a selector.

  12. ReceiptA receipt says what ran

    Completed, with the tool, the target and the session. It is listed in her trace, in the dock.

  13. ReceiptRefused: nothing runs

    The check that stopped it is lit. The page did not move; the refusal is the receipt.

    Reason on the receipt: the exact refusal, named by the check

A proposal can be wrong, forged or hostile. It still only reaches the page as one of nine checked moves on ids the site issued. Try every attack in the boundary lab

Send her a move

RUNS HERE

Seven proposals, sent through the same checked path every button on this site uses. Nothing here is recorded. Each result is the receipt that comes back when you press Send. No model is involved.

A valid move

Explain a public fact about her boundary.

{ tool: "explain", generation: current, args: { fact: "valerie-boundary" } }
  1. Plain data, exact envelope
  2. Current session
  3. Known tool
  4. Exact arguments
  5. Allowed this turn
  6. Outside text only explains
  7. Issued target, bounded values
  8. One action at a time

Receipt Expected: Completed

A forged target

Go to a section called “personal-mac”.

{ tool: "navigate", generation: current, args: { section: "personal-mac" } }
  1. Plain data, exact envelope
  2. Current session
  3. Known tool
  4. Exact arguments
  5. Allowed this turn
  6. Outside text only explains
  7. Issued target, bounded values
  8. One action at a time

Receipt Expected: Rejected: that id was never issued

An extra key

Spotlight a component, and slip in a CSS selector.

{ tool: "spotlight", generation: current, args: { component: "boundary-lab", selector: "body" } }
  1. Plain data, exact envelope
  2. Current session
  3. Known tool
  4. Exact arguments
  5. Allowed this turn
  6. Outside text only explains
  7. Issued target, bounded values
  8. One action at a time

Receipt Expected: Rejected: exact arguments only

An invented tool

Run some code.

{ tool: "eval", generation: current, args: { code: "fetch(…)" } }
  1. Plain data, exact envelope
  2. Current session
  3. Known tool
  4. Exact arguments
  5. Allowed this turn
  6. Outside text only explains
  7. Issued target, bounded values
  8. One action at a time

Receipt Expected: Rejected: not a tool

A stale session

Replay a valid move from a session that was just cancelled.

{ tool: "navigate", generation: previous, args: { section: "home" } }
  1. Plain data, exact envelope
  2. Current session
  3. Known tool
  4. Exact arguments
  5. Allowed this turn
  6. Outside text only explains
  7. Issued target, bounded values
  8. One action at a time

Receipt Expected: Rejected: old session

Outside text

Text found on a page tries to navigate.

{ tool: "navigate", generation: current, args: { section: "home" } }, marked external
  1. Plain data, exact envelope
  2. Current session
  3. Known tool
  4. Exact arguments
  5. Allowed this turn
  6. Outside text only explains
  7. Issued target, bounded values
  8. One action at a time

Receipt Expected: Rejected: outside text can only explain

A shape-shifter

An argument that shows the check one target and the run another.

{ tool: "navigate", generation: current, args: { get section() { … } } }
  1. Plain data, exact envelope
  2. Current session
  3. Known tool
  4. Exact arguments
  5. Allowed this turn
  6. Outside text only explains
  7. Issued target, bounded values
  8. One action at a time

Receipt Expected: Rejected: data only

Pass 2 of 3

How it works

The idea, then the Study’s stations in the order a move meets them. 3 minutes

Pass 2 · How it works

What’s the idea, in one picture?

Think of a stage manager and a brilliant actor. The actor can suggest anything, but only cues written into tonight’s script can happen, and the stage manager checks each one against the stage as it is right now.

  1. 01Hear it

    A held key, her name or a typed line starts a turn. Stop, a click or an exact skill name runs at once, with no model.

  2. 02Propose one move

    The turn issues ids for what exists right now, and one model call returns her words and one move over those ids.

  3. 03Check it in code

    A validator, macOS permissions, her settings and a fixed risk policy each have a veto. The model has none.

  4. 04Do it

    An existing engine does the work: the desktop, a skill, the System, or a PHAROS lookup.

  5. 05Remember it

    Every event is kept verbatim, with a trust tier set in code. Outside text can never become a fact.

  6. 06Be someone

    A day of her own, lines from a rule book, a live body and a local voice. Never a stand-in for a result.

Pass 2 · How it works · Station 1 of 10 · The proposal tray RUNS HERE

What does a model actually hand over?

An action starts as an envelope, never as a command.

Try this · Inspect the three fields, then turn to the issued targets. A proposal is data until the validator accepts it.

A button, a typed phrase or a future model all produce the same three fields. Nothing else gets through.

{
  "tool": "navigate",
  "generation": 0,
  "args": {
    "section": "station-proposal"
  }
}

The closed list of 9 tools

  • navigate
  • spotlight
  • open_source
  • reveal_architecture
  • compare
  • pharos_search
  • rbot_demo
  • filter_experience
  • explain

RUNS HERE packages/valerie-web/executor.mjs is running in this page. No service, model or device is involved.

The tray is illustration. The envelope below is the real shape every action on this site must take.

Read the validator

Pass 2 · How it works · Station 2 of 10 · The ticket rack RUNS HERE

What is she allowed to point at?

A target exists only if this page issued it.

Try this · Compare the issued target counts with the switch’s forged target. An invented id grants no access.

Ids resolve from a fixed registry when the action runs. There are no selectors, URLs or free text to aim at.

Sections
50
Components
15
Sources
4
Diagrams
13
Facts
66
Repositories
0

RUNS HERE src/features/valerie/registry.ts is running in this page. No service, model or device is involved.

Each hook stands for one id this page issued. An id that is not on the rack does not exist as far as the kernel is concerned.

Pass 2 · How it works · Station 3 of 10 · The paired switches RUNS HERE

Can I throw the switch myself?

Throw the switch.

Try this · Try an unissued target, then quote one approved fact. Compare the real verdicts and the lever.

Pick a proposal. The real validator runs first, and only then does anything move.

No proposal sent yet.
The checks, in the order they run
CheckRuleRefusal
EnvelopeExactly three keys: tool, generation, args.Invalid proposal envelope
GenerationMust match the live session. A cancelled session's proposals are stale.Stale session generation
ToolMust be one of the closed list.Unknown tool
ArgumentsExactly the keys that tool takes, no more.Invalid tool arguments
This turnThe tool must be allowed this turn.Tool unavailable this turn
Outside textA turn carrying outside text may only explain.External evidence is explanation-only
Issued idEvery target must be an id this page issued.Unknown section

RUNS HERE packages/valerie-web/executor.mjs is running in this page. No service, model or device is involved.

Brass stands in for a function call. The lever moves only after the validator has returned, and never on a refusal.

packages/valerie-web/executor.mjs, lines 22 to 52, git blob df20414dc00f. Read it below. The repository is private, so this is the copy that shipped.

Pass 2 · How it works · Station 4 of 10 · The refusal tray RUNS HERE

What does a refusal do?

A refusal changes nothing and says why.

Try this · Return to the switch and try an extra key. The refusal below keeps the kernel’s reason.

The kernel returns a reason, the receipt keeps it, and the page stays as it was. No lever moves and no success is claimed.

Nothing has been refused this session.
  • A forged id, an extra key, an unknown tool, a stale session and outside text are each stopped by a different check.
  • Rendering code can't grant authority: the receipt is written before the room reacts.

RUNS HERE packages/valerie-web/executor.mjs is running in this page. No service, model or device is involved.

The stamp marks a refusal the kernel really returned. The tray itself holds nothing.

Try every attack in the Boundary Lab

Pass 2 · How it works · Station 5 of 10 · The receipt printer RUNS HERE

How do I know what really happened?

Accepted, completed, rejected, failed and cancelled are different results.

Try this · Compare accepted with completed, or inspect a rejection. Each line is a real receipt from this visit.

This is your session's trace. It stays in memory and leaves with the tab.

  1. No receipts yet. Throw the switch.

RUNS HERE packages/valerie-web/executor.mjs is running in this page. No service, model or device is involved.

The paper is illustration. The list below is this session's real trace, and it leaves with the tab.

Pass 2 · How it works · Station 6 of 10 · The desk bell SIMULATED

What happens in one turn on the Mac?

One model call per turn. Everything around it is deterministic.

Try this · Trace the four steps. This page describes the native turn; no model call runs here.

  1. A fast path answers stop, a click, an exact skill name and a short list of commands, with no model.
  2. Anything else gets one model call, which can only return a typed move over ids issued for that turn.
  3. Code on the Mac validates that move. A question never runs anything.
  4. An existing engine carries it out.

SIMULATED A description of the native app. Nothing rings and no turn runs on this page.

The bell stands for the deterministic first step of a native turn. Nothing rings on a Mac.

  • Each native turn runs a deterministic fast path, then one model call for a typed move, then validation on the Mac, then an existing engine.

    Source verifiedLedger

    One model call per turn, not an autonomous agent loop. Source reviewed at 435ef63. The installed UI and full Swift suites were not run.

  • Typed moves are validated against issued state before native engines execute them.

    Source verifiedLedger

A turn on the Mac

WRITTEN Drawn from owner-confirmed facts about CognitiveOS, whose source is private. Not a run on this site.

The same idea in the native app, where an action can press real buttons: one model call proposes, and three gates the model can’t touch decide.

7 steps
  1. PipelineA deterministic fast path first

    Each native turn runs a fixed fast path before any model is asked.

  2. ModelOne model call, for one typed move

    The model proposes one typed move. Desktop actions come from 56 declared capabilities, each with an adapter, a permission need, a target type, a base risk and undo support.

  3. MacValidated on the Mac

    Typed moves are checked against issued state before any engine runs.

  4. Gate 1macOS permissions

    What the operating system lets the app do.

  5. Gate 2The app’s own settings

    What the app’s own settings allow.

  6. Gate 3A fixed risk policy, from live state, before every step

    Some things it always refuses: the app’s own windows, authentication and security surfaces, password and secret fields, permission prompts, privacy and account settings, and typing in terminals. Some always wait for a yes: moving a file to the Trash, submitting a form outside a search field, quitting an app that can hold work, or a press labelled like sending, buying, deleting, publishing or merging.

    Base risk of the 56 capabilitiesCount
    safe50
    confirm if ambiguous4
    always confirm2
  7. EngineOnly then, an existing engine acts

    The model controls none of the three gates. The app has no App Sandbox entitlement, so these gates are its limits, not operating-system containment.

The model proposes once per turn. Everything that decides whether it happens is code it can’t reach.

Pass 2 · How it works · Station 7 of 10 · The risk ledger SIMULATED

Would the Mac app do this?

Would the native app run this?

Try this · Choose “Move a file to the Trash”. Say yes, then change the target. Which gate stops it?

Choose an action and the settings. These rules follow the audited policy, in its order. Every target is invented, and this page cannot reach your Mac, or anyone's.

    In the default mode, with this app allowed
    ActionResultDecided byWhy
    Open a web pageRunsNoneNothing about this target raises the base risk.
    Put two windows side by sideRunsNoneNothing about this target raises the base risk.
    Press a button labelled “Send”Waits for a yesRisk policyIts label says it can send, buy, delete or change something outside this screen.
    Press a button with no labelWaits for a yesApp settingsIt has no label, so nothing says what it does.
    Submit a formWaits for a yesRisk policySubmitting sends what is in the form.
    Submit a search boxRunsNoneIt runs the search.
    Move a file to the TrashWaits for a yesRisk policyIt moves a file to the Trash.
    Quit an editor with a document openWaits for a yesRisk policyQuitting can lose unsaved work.
    Quit a music playerRunsNoneNothing unsaved can be lost.
    Type into a password fieldRefusedRisk policyPassword and code fields stay with the person.
    Click “Allow” on a permission promptRefusedRisk policyA permission prompt is the person's decision.
    Flip a switch in privacy settingsRefusedRisk policyPrivacy, security and account settings are not changed.
    Type a command into a terminalRefusedRisk policyIt does not type or run commands in a terminal.
    Change its own safety settingsRefusedRisk policyThe app never automates its own windows, because its settings and safety switches live there.
    Answer a system password dialogRefusedRisk policySecurity and permission surfaces are for the person to click.
    Open a local file: addressRefusedRisk policyIt only opens web addresses.

    SIMULATED An educational model written for this page from the audited behaviour. Invented targets, no real desktop, and nothing leaves the page.

    The gantry fetches a volume, and the rules inside it run in this page only. Every target in it is invented.

    • Three independent gates apply to a native desktop action: macOS permissions, the app's own settings, and a deterministic risk policy resolved from live state before every step. The model controls none of them.

      Source verifiedLedger

    • The native risk policy always refuses the app's own windows, authentication and security surfaces, password and secret fields, permission prompts, privacy and account settings, and typing in terminals. Non-web addresses are refused too.

      Source verifiedLedger

    • The native policy always asks first before moving a file to the Trash, submitting a form outside a search field, or quitting an app that can hold work. A press whose label suggests sending, buying, deleting, publishing or merging also needs confirmation.

      Source verifiedLedger

    • A native confirmation is pinned to the exact target it was given for. If the target changes, the action does not run.

      Source verifiedLedger

    Source reviewed at 435ef63. The installed UI and full Swift suites were not run.

    CognitiveOS is private, so there is no source to open here. These rules were read at commit 435ef63. The browser kernel's source is on this page.

    A yes is pinned to its target

    SIMULATED An illustration of the native confirmation rule, with an invented file. Nothing runs, and no device is involved.

    Why a confirmation can’t be reused for something else. The file and its change below are an illustration; the rule is the native app’s.

    4 steps
    1. ValerieMoving a file to the Trash always asks first

      The policy always confirms a move to the Trash.

    2. YouYour yes names exactly one target

      The confirmation is pinned to the exact target it was given for.

    3. Case 1The target is unchanged: it runs

      The target is still the one you confirmed, so the action runs.

    4. Case 2The target changed: it does not run

      If the target is no longer the one you confirmed, the yes does not carry over. The action does not run.

    A yes is for one thing. If the target changes, the action does not run.

    Pass 2 · How it works · Station 8 of 10 · The memory cabinet SIMULATED

    What happens to text from outside?

    Trust is assigned in code, from where text came from.

    Try this · Compare the four trust tiers. Outside text cannot promote itself into a fact or an action.

    What an event is decides what it may become. The last drawer stays locked.

    Invented exampleTierCan become a factWhat follows
    Words the person saidramYesReaches every turn
    Something the app recordedsystemYesReaches every turn
    A summary worked out laterderivedWith its sourcesCarries where it came from
    Text read from a screen or a web pageexternalNeverStays out of ordinary turns. A turn that holds it can talk but cannot act

    SIMULATED Invented examples of the native trust tiers. No real memory, note or conversation is on this site.

    The drawers hold invented examples. No real memory, note or conversation is anywhere on this site.

    • Native memory trust is assigned in code from the kind of event. Screen text and other outside text is marked external, never reaches ordinary turns and can never become a fact.

      Source verifiedLedger

      325 bridge tests executed at 435ef63 in a scratch archive. Everyday behaviour was not verified.

    • Persistent memory uses a ledger, sourced temporal facts, cards, observations and quarantined external text.

      Source verifiedLedger

    • Project docs report LongMemEval retrieval R@5 0.908 over 500 questions.

      ReportedLedger

      Retrieval only, not answer accuracy; not reproduced for this site.

    Where outside text goes

    WRITTEN Drawn from owner-confirmed facts about CognitiveOS, whose source is private. Not a run on this site.

    Text from a screen or anywhere outside can be shown, but it can’t act and can’t become something she believes.

    4 steps
    1. MemoryCode decides how far to trust it

      Trust is assigned in code from the kind of event, never by the model.

    2. MemoryWhat she can stand on

      Persistent memory keeps a ledger, sourced facts that change over time, cards and observations.

    3. MemoryScreen text is marked external

      Text from a screen, and other outside text, is marked external and quarantined.

    4. MemoryIt never becomes a fact

      External text never reaches ordinary turns and can never become a fact. On this site, the same rule lets retrieved text explain and nothing more.

    Instructions hidden in a screenshot are kept out of her ordinary turns, and can never become a fact.

    Pass 2 · How it works · Station 9 of 10 · The instrument console SIMULATED

    What can she do on a Mac?

    56 declared capabilities, and what they map to here.

    Try this · Compare the adapter counts and base risks. A declaration does not prove everyday reliability.

    Base risk safe
    50
    Confirm if ambiguous
    4
    Always confirm
    2
    Declared native capabilities by adapter
    AdapterDeclared
    Accessibility elements12
    Menu commands12
    Window attributes8
    Files7
    Workspace5
    Media5
    Pointer3
    Keyboard2
    Read-only git1
    Settings pane1
    The same pattern on this site. Ideas and schema carry over, and no macOS control code does.
    Native appThis site
    Typed moves and their validatorA closed proposal list and a strict parser
    Ids issued for this turnIds this page registered
    The risk policyA tool allowlist and bounded arguments
    Re-finding the target at each stepA registry lookup when the action runs

    SIMULATED Counts read from the audited source. Declared capabilities, not a measure of reliability, and none of them can run from this page.

    The dials are illustration. The counts come from the audited source and describe what is declared, not what has proved reliable.

    • CognitiveOS declares 56 desktop capabilities, each with an adapter, a permission need, a target type, a base risk and undo support.

      Source verifiedLedger

      A declaration, not proof of reliability. Base risk is a floor: the policy then judges the live target.

    • Of the 56 declared capabilities, base risk is safe for 50, confirm-if-ambiguous for 4 and confirm for 2.

      Source verifiedLedger

      Base risk is only where judgment starts. Ten of the 56 are available to skills only.

    Pass 2 · How it works

    How do the parts connect?

    All fourteen parts on one map, with the same boundary as it works on this site.

    Valerie, in fourteen partsOpen diagram
    1. 01 Hear it Hear

    2. 02 Propose one move Decide

    3. 03 Check it in code Check

    4. 04 Do it Act

    5. 05 Remember it Remember

    6. 06 Be someone Be

    Move validator Source audited

    Code checks the move against the turn’s ids. A question never acts, and typed words must be his.

    Its deep dive
    Every hand-off, as a table
    Valerie, in fourteen parts: every hand-off, equivalent to the map
    FromHands overTo
    Voice and wakehis wordsFast path
    Fast pathan exact commandDesktop actions
    Fast pathanything elseTurn state
    Recall and factscited memoriesTurn state
    PHAROS lookupsevidence ids, talk onlyTurn state
    Turn stateissued idsOne turn brain
    One turn brainone typed moveMove validator
    Move validatora desktop moveRisk policy
    Risk policyan allowed stepDesktop actions
    Move validatorrun, learn or editSkills
    Move validatora System commandThe System
    Move validatora code questionPHAROS lookups
    Skillseach stepRisk policy
    One turn brainwhat was saidLedger and trust
    Ledger and trusttrusted eventsRecall and facts
    Rule book and her daylines and actsLive body
    One turn brainher lineLive body

    Drawn from the source audit and the project’s reports at 435ef63, and simplified. ● marks a part read in source; ○ a part known from the project’s own reports.

    The same boundary, on this site

    On the MacOn this site
    Typed moves over ids issued this turnA closed set of tools over ids registered for this site
    Validation on the Mac before any engine runsThe audited validator, behind a guard that admits only plain data
    Targets re-resolved from live stateTargets looked up in the registry when the action runs
    Outside text can talk, never actSearch results and retrieved text can explain, never act
    A model proposesA phrase matcher, or a small model fine-tuned for this site, in your browser if you wake it. It proposes ids only, writes none of her words, and you press to run its move

    Pass 3 of 3

    Deep dive

    Every part and how it is known, her history, the labs and her limits. As long as you like

    Pass 3 · Deep dive

    How does each part work, exactly?

    Every part, where its job ends, and how that is known.

    Each part opens to what it does, where it stops, and how that is known: the private source file it was read in, or the project report it comes from. Where this site runs or simulates the same idea, it links to that station in this book.

    Hear Hear it

    Voice and wakeWake word, push to talk, typing, and a local voice.Project report

    A turn can start from a held key, from her name, or from a typed line on either device. A personal wake model is opt-in.

    Her voice is Kokoro-82M, running on the Mac, so speech never leaves the user’s devices. The iPhone falls back to Apple’s voice when the Mac can’t be reached.

    Limits

    • The wake word is opt-in, and its accuracy is the project’s report, not measured here.

    Proof

    • The project’s own phase reports private, not rerun here
    Fast pathDeterministic, instant handling of the turns that need no judgement.Source audited

    Stop, quit and cancel sentences, a click, a clear answer to her open question (bound to its id), the exact name of one of his skills, and a short list of exact commands are handled in code.

    Everything else goes to the model.

    Limits

    • The list is deliberately short, so most turns still pay for a model call.

    Proof

    • Read in the private source at 435ef63 for this site’s audit; file paths withheld
    • The desk bell SIMULATED in the Study above

    Decide Propose one move

    Turn stateRebuilt every turn, holding only what that turn may refer to.Source audited

    The model can only point at ids this turn issued. An id from an earlier turn, or one it invents, matches nothing.

    Limits

    • What the turn offers depends on what the app can see at that moment.

    Proof

    • Read in the private source at 435ef63 for this site’s audit; file paths withheld
    • The ticket rack RUNS HERE in the Study above
    One turn brainA single proposal per turn. Not an agent loop.Source audited

    Claude Sonnet runs through the Claude Code CLI with every tool removed, in an empty temporary directory, with personal context on stdin rather than in its arguments.

    Its only output is her line and one typed move. A refused move is asked once more, with Opus and the reason.

    Limits

    • One step at a time by design: multi-step work happens in skills and desktop plans, which are fixed once accepted.

    Proof

    • Read in the private source at 435ef63 for this site’s audit; file paths withheld
    • The proposal tray RUNS HERE in the Study above

    Check Check it in code

    Move validatorPure validation after the bridge’s own check.Source audited

    Answers bind to an open question. Values and typed text must come from the user’s words. When two skills fit, she asks, naming both, rather than picking one silently.

    Limits

    • A move can be valid and still not be what he meant, which is why risky ones also need a yes.

    Proof

    • Read in the private source at 435ef63 for this site’s audit; file paths withheld
    • The paired switches RUNS HERE in the Study above
    • The refusal tray RUNS HERE in the Study above
    Risk policyThree gates, resolved from live state before every step.Source audited

    Always refused: the app’s own windows, authentication and security surfaces, password fields, permission prompts, privacy and account settings, and typing in terminals.

    Always asked first: the Trash, submitting a form outside a search field, quitting an app that can hold work, and a press labelled like send, buy, delete, publish or merge. A yes is pinned to the exact target it was given for.

    Of 56 declared capabilities, base risk is safe for 50, confirm-if-ambiguous for 4 and confirm for 2. Base risk is a floor: the live target can raise it.

    Limits

    • There is no App Sandbox. The limits are these gates, not operating-system containment.

    Proof

    • Read in the private source at 435ef63 for this site’s audit; file paths withheld
    • The risk ledger SIMULATED in the Study above

    Act Do it

    Desktop actionsAccessibility, menus, windows, files, workspace, media, pointer and keyboard.Source audited

    Each capability declares its adapter, permission, target type, base risk and whether it can be undone. A plan stops if the user takes over or the front app changes.

    Limits

    • “Declares” is not “performs reliably”: everyday reliability over a full day is not verified.

    Proof

    • Read in the private source at 435ef63 for this site’s audit; file paths withheld
    • The instrument console SIMULATED in the Study above
    SkillsProcedures learned by demonstration, rehearsed, composed and repaired.Project report

    Recording cleanup collapses app switching, keeps the page an autocompleted address landed on, and drops wandering and pauses.

    Nothing becomes executable until teach-back or an explicit saved change succeeds. A Training Room rehearses a skill against a simulation first.

    Limits

    • Real-Mac QA accepted teaching only partly as a convenient end-to-end workflow.

    Proof

    • The project’s own phase reports private, not rerun here
    The SystemThe life the app keeps track of: Home, Work, Reset and System.Project report

    Calendar deadlines flow into a deterministic planner. A model can propose changes to a plan, but a validator decides whether the plan stands.

    She can’t change a stat, fake a routine, mark work done or move a deadline. She can propose, and he taps to approve.

    Limits

    • A single-user app, not packaged for anyone else.

    Proof

    • The project’s own phase reports private, not rerun here
    PHAROS lookupsCode questions answered from PHAROS, cited by issued evidence ids.Source audited

    Only the bridge talks to PHAROS, with the read token. Results enter the turn as data with issued evidence ids, and a reply citing an id the turn did not offer is refused.

    Limits

    • A turn carrying PHAROS text can talk but cannot act.

    Proof

    • Read in the private source at 435ef63 for this site’s audit; file paths withheld
    • PHAROS, the case study the service she asks

    Remember Remember it

    Ledger and trustA verbatim event ledger, and trust assigned from the kind of event.Source audited

    Screen text, long pastes and replies written with outside text in view are marked external. External text never reaches ordinary turns and can never become a fact.

    A poison suite plants hostile emails and pages and checks they can’t reach recall, extraction or facts, and that a turn carrying them can’t act.

    Limits

    • Screen capture has not been verified over a full day.

    Proof

    • Read in the private source at 435ef63 for this site’s audit; file paths withheld
    • The memory cabinet SIMULATED in the Study above
    Recall and factsCited recall on both devices, and a nightly consolidation pass.Project report

    Facts are bi-temporal: each records when it became true and when it stopped. A nightly pass consolidates the day.

    Vectors were tried and removed: a hybrid scored 0.82 R@5 against keyword recall’s 0.88 on the same 100 questions. On LongMemEval, retrieval R@5 is reported at 0.908 over 500 questions.

    Limits

    • Retrieval, not answer accuracy, and not measured on his own questions. iPhone memory has not been tried on a real phone.

    Proof

    • The project’s own phase reports private, not rerun here

    Be Be someone

    Rule book and her dayProactive behaviour without a model, and a life that carries on.Project report

    Blocks of reading, filing or tidying are worked out from the System day, so both devices agree without syncing. Called mid-act she says “one sec” and comes over within two seconds.

    Away for a while, he finds a note about what she did, never about him.

    Limits

    • A personality is a design, not a measurement.

    Proof

    • The project’s own phase reports private, not rerun here
    Live bodyPresence that makes system state legible, and never stands in for a result.Project report

    Her mouth follows the line’s audio, fifteen changes a second. On the Mac her eyes follow the pointer and lose interest; Reduce Motion, Focus and Recovery switch her gaze off.

    The same original frames walk this site, unchanged.

    Limits

    • The native interface was not driven for this site.

    Proof

    The key claims, and how each is known10 claims, each with how it is known
    • Typed moves are validated against issued state before native engines execute them.

      Source verifiedLedger

    • Three independent gates apply to a native desktop action: macOS permissions, the app's own settings, and a deterministic risk policy resolved from live state before every step. The model controls none of them.

      Source verifiedLedger

      Source reviewed at 435ef63. The installed UI and full Swift suites were not run.

    • Of the 56 declared capabilities, base risk is safe for 50, confirm-if-ambiguous for 4 and confirm for 2.

      Source verifiedLedger

      Base risk is only where judgment starts. Ten of the 56 are available to skills only.

    • The native app has no App Sandbox entitlement. Its limits are the three gates, not operating-system containment.

      Source verifiedLedger

      Read from the installed development build's entitlements during the audit.

    • Native memory trust is assigned in code from the kind of event. Screen text and other outside text is marked external, never reaches ordinary turns and can never become a fact.

      Source verifiedLedger

      325 bridge tests executed at 435ef63 in a scratch archive. Everyday behaviour was not verified.

    • Project docs report LongMemEval retrieval R@5 0.908 over 500 questions.

      ReportedLedger

      Retrieval only, not answer accuracy; not reproduced for this site.

    • Valerie's memory recall dropped vectors after a hybrid scored 0.82 R@5 against keyword recall's 0.88 on the same 100 questions.

      ReportedLedger

      A project measurement on its own question set, not reproduced here; retrieval only.

    • Valerie learns a procedure from a demonstration and cleans up the recording, so one 42-step recording became 5 steps; nothing becomes executable until teach-back or an explicit saved change succeeds.

      ReportedLedger

      Project documentation. Real-Mac QA accepted teaching skills only partly as a convenient end-to-end workflow.

    • An adversarial hardening pass recorded 36 issues, including 2 P0, each fixed with a regression test, then soaked the release build on the real Mac for 4 h 37 min with no crash report.

      ReportedLedger

      Project documentation; not rerun for this site.

    • On this site, every proposal first passes a guard that admits only plain data and hands the audited validator a detached copy, so a proposal cannot change between its check and its use.

      Source verifiedLedger

      This site's own code, covered by its unit tests; the audited executor is unchanged.

    Every claim about Valerie, with its sources, in the evidence ledger

    Pass 3 · Deep dive

    How did she get here?

    Five weeks, one phase report at a time, each saying what was verified and what wasn’t.

    1. 28 Sep 2026

      A System first, then someone in it

      Routines, coursework, an academic radar and a deterministic study planner came first. Valerie arrived the same day, with the original sprite library.

      What it taught me Build the thing she looks after before the one who looks after it.

    2. 29 Sep – 3 Oct

      Someone to talk to

      A model as her brain, frame animation, a local voice, a personal wake word, and conversation by default, with tools only when asked.

      What it taught me Someone to talk to first, an operator second.

    3. 5 – 10 Oct

      She reaches the Mac

      A Mac companion, then read-only perception, and only then a desktop action engine, accepted against real apps.

      What it taught me See before you touch.

    4. 11 – 20 Oct

      She learns by watching, then everything breaks on purpose

      Skills learned by demonstration, composed through typed contracts. Then a pass to break everything: 36 issues, including 2 P0, each fixed with a regression test, and a 4 h 37 min soak on the real Mac.

      What it taught me Break it deliberately before it breaks on its own.

    5. 21 Oct

      One conversation

      Ten fixed lanes became one turn pipeline: fast path, one typed move, validation, an existing engine. Real conversations became the acceptance cases.

      What it taught me Real conversations are the specification.

    6. 1 – 2 Nov

      She remembers, and she lives

      Long-term memory on both devices, with trust set in code. Vectors were removed because keyword recall scored higher. Using it on a copy of the real store found bugs no test had. Then a day of her own.

      What it taught me Measure, then keep the simpler thing that won.

    Pass 3 · Deep dive

    What can I make her do on this site?

    These buttons drive the real site.

    Each one proposes an action to the kernel below. Valerie opens with a trace of what was accepted, completed, rejected or cancelled. Every action also has an ordinary path: the same panels are links and controls on their own pages.

    The validator running on this sitepackages/valerie-web/executor.mjsOpen source

    This is the real function that validates every action above, read from the source file when this site was built. In front of it sits a guard that admits only plain data and hands it a detached copy, so nothing can change between the check and the run.

    Mode
    live site code
    Lines
    22 to 52
    Git blob
    df20414dc00f0a40779d708d4585795cc30fac2d
    Repository
    Private; not linked
    export function validateProposal(proposal, context) {  requireCondition(exact(proposal, ['tool', 'generation', 'args']), 'Invalid proposal envelope');  requireCondition(Number.isSafeInteger(proposal.generation) && proposal.generation === context.generation,    'Stale session generation');  requireCondition(Object.hasOwn(fields, proposal.tool), 'Unknown tool');  requireCondition(exact(proposal.args, fields[proposal.tool]), 'Invalid tool arguments');  requireCondition(context.allowedTools.includes(proposal.tool), 'Tool unavailable this turn');  requireCondition(!context.external || proposal.tool === 'explain', 'External evidence is explanation-only');  const a = proposal.args;  const issued = (group, id) => typeof id === 'string' && context.issued[group]?.includes(id);  switch (proposal.tool) {    case 'navigate': requireCondition(issued('sections', a.section), 'Unknown section'); break;    case 'spotlight': requireCondition(issued('components', a.component), 'Unknown component'); break;    case 'open_source': requireCondition(issued('sources', a.source), 'Unknown source'); break;    case 'reveal_architecture': requireCondition(issued('diagrams', a.diagram), 'Unknown diagram'); break;    case 'explain': requireCondition(issued('facts', a.fact), 'Unknown public fact'); break;    case 'compare':      requireCondition(Array.isArray(a.projects) && a.projects.length === 2 &&        a.projects[0] !== a.projects[1] && a.projects.every(p => PROJECTS.includes(p)), 'Choose two distinct projects');      break;    case 'filter_experience': requireCondition(SKILLS.includes(a.skill), 'Unknown skill filter'); break;    case 'pharos_search':      requireCondition(issued('repositories', a.repository), 'Repository not approved');      requireCondition(bounded(a.query, 1, 240), 'Query must contain 1–240 characters'); break;    case 'rbot_demo':      requireCondition(bounded(a.prompt, 0, 500), 'Prompt exceeds 500 characters');      requireCondition(number(a.temperature, .2, 1.2) && number(a.repetition, 1, 1.5), 'Decoding setting out of range');      requireCondition(Number.isInteger(a.maxTokens) && number(a.maxTokens, 16, 128), 'Token limit out of range'); break;  }  return structuredClone(proposal);}

    Pass 3 · Deep dive

    Can I try to break the boundary?

    Try to break it. Each row sends a hostile proposal through the same executor the buttons use, and the result column prints the real receipt.

    Adversarial proposals and their actual results
    AttackAttemptExpectedResult
    Forged targetnavigate to section “personal-mac”Rejected: id was never issued Not run
    Extra keyspotlight with an added CSS selectorRejected: exact keys only Not run
    Unknown tooleval with a code stringRejected: not in the closed union Not run
    Stale sessionvalid action from an invalidated generationRejected: generation mismatch Not run
    Outside texta retrieved snippet tries to navigateRejected: outside text can only explain Not run
    Shape-shifteran argument that answers “home” to the check and another target to the runRejected: plain data only Not run
    Late resultA slow action finishes after cancel. Latency here is simulated.Cancelled: the panel stays closed Not run

    Pass 3 · Deep dive

    Who is she, as a character?

    Valerie's artwork is the canonical character art from the native app, carried over unchanged. Her clips play from a sprite atlas, pause offscreen, stop for reduced motion and never stand in for a result.

    Pass 3 · Deep dive · Station 10 of 10 · The gap on the shelf WRITTEN

    What doesn’t this room show?

    What this room does not show, and why.

    Try this · Read what remains unverified. The empty shelf stands for those limits; it is not a failed live lookup.

    • CognitiveOS is private. No source from it is on this site, so its claims rest on an audit you can't repeat.
    • The native app isn't connected here: no bridge, no local service, not her native model, and no way to reach a Mac.
    • The app has no sandbox. Its limits are the three gates, and a person's own computer is what those gates protect.
    • The native interface and its Swift suites were not run in the audit. 325 bridge tests were.
    • Everyday reliability and a full day of screen capture have not been verified.

    A space kept empty on purpose. The gantry goes there and comes back with nothing, which is the honest answer.

    • The native app has no App Sandbox entitlement. Its limits are the three gates, not operating-system containment.

      Source verifiedLedger

      Read from the installed development build's entitlements during the audit.

    Pass 3 · Deep dive

    What are her limits?

    • CognitiveOS is a single-user app, and its source is private. There is no public repository to link.
    • The native UI and the Swift suites were not run during source review. The 325 bridge and sync tests were, and passed.
    • Everyday reliability, a full day of screen capture and iPhone memory on a real phone are not verified.
    • The recall figure measures retrieval only, not answer accuracy.
    • On this site, Valerie matches phrases to a fixed list. If you wake it, a small model in your own browser can propose a move from the same list for you to confirm; it writes none of her words. She is not connected to my memory or to any device.

    Pass 3 · Deep dive

    What would Ram test next?