The Personal AI

At twenty past nine, Priya begins an architecture review. The channel has a standing policy she declared, observations from earlier reviews, an inference about derivations, and the means to apply temporary instructions and carry boundaries to selected services. It could use all of them.

Priya says: I have ten minutes. In this review view, show material that bears on the decision. Keep the decisive equations and assumptions visible; fold the supporting steps, with the full calculation available.

Before using the accumulated state, the system has a question to answer: what does it add to that request?

Chapter 27 made consequential beliefs inspectable and correctable. That does not establish that more of them should exist. The closing question of Part III is therefore:

How much personal state does this communication actually require, and what must an additional item justify?

more personal state ≠ better communication

The title The Personal AI names the temptation as well as the possibility. A system may seem to understand Priya because several narrow mechanisms work together. Their composition does not establish a need for a comprehensive account of her.

Six available inputs, not six necessities

For this constructed session, six kinds of input are available:

Standing policy
  Architecture reviews: preserve technical detail;
  expand derivations when the decision depends on them.

Retained observation
  Priya expanded derivations in earlier architecture reviews;
  the retained records include those sessions' conditions.

Inference
  Expanded derivations may be useful in similar reviews;
  provisional, with scope limited to the supporting evidence.

Session overlay
  Derived from today's explicit request: keep decisive material
  visible and fold only supporting steps for this review.

Friction contract
  For this review, remind her before opening another paper;
  opening it immediately remains available.

Portable boundary
  A separately accepted temporary rule for the selected news
  and video recommendation feeds; search remains available.

The ten minutes are current context. The instruction about what to fold supplies permission for the presentation change. The overlay is the bounded application derived from those inputs, not another name for the time limit.

Priya explicitly accepts the reminder for this review. She separately authorises holding new consensus-research recommendations out of the selected feeds, with direct opening and an easy local override preserved. Both rules end when the review or session ends, or at half past nine, whichever comes first. Their permission is not inherited from an expired session or inferred from her wish to finish.

The portable rule also carries the temporary-learning limit: behaviour produced under it cannot silently become durable preference evidence.

These records are available to the relevant mechanisms. They need not all contribute to a single rendering, and they should not all be displayed continuously. Inspection follows their effects, as the previous chapter required.

What contributes to the view?

Three items arrive: a library release mentioned in the architecture document, a new consensus-paper recommendation in a selected feed, and a colleague’s note about the review containing a calculation Priya needs to check.

The note bears directly on the current decision. A candidate view puts the decisive equation and its failure assumptions in the main explanation, with supporting steps folded and the full calculation available. This follows from the source’s preservation obligations and Priya’s current instruction. Her standing policy is compatible with it: any derivation the decision actually depends on must remain visible.

The ten-minute condition does not by itself settle which material is dispensable. If a supposedly supporting step turns out to determine the conclusion, the channel must keep it or expose that the requested brief view cannot meet the preservation obligation. An overlay cannot resolve that conflict by relabelling necessary material as optional.

The library release is relevant to the document but does not change the calculation under review. Under Priya’s instruction for this view, it remains available outside the main explanation. That task-relative selection needs no conclusion about her general interest in libraries or her tolerance for distraction.

The consensus recommendation encounters a different mechanism. Suppose the video feed supports the accepted temporary filter but the news service offers only a broader research block. The adapter can report the supported application and the unsupported destination; it cannot apply the broader block or claim that all incoming sources are protected. A recommendation can therefore still appear in the news service. The reading view’s task selection does not prove enforcement elsewhere.

If Priya chooses to open another paper, the reminder applies within its accepted scope, with immediate override. Neither the unfinished review nor an ignored reminder authorises stronger intervention.

The proposed explanation of the calculation’s presentation is short:

Why this view?

Source obligation:  decisive equation and assumptions kept visible
Current request:    only supporting steps folded for this review
Standing policy:    decision-relevant technical detail preserved
Past observations:  not needed for this presentation choice
Preference inference: not used

That last distinction is the point of the construction. State the system holds is not necessarily state the communication needs.

Try subtracting before adding

Set aside the earlier observation and the inference about derivations. The stated rationale for this candidate view remains: the current request, the source obligations and the applicable declared policy already supply it. In this worked construction, those two extra records justify no additional presentation choice.

The operation needs a name:

not consulted ≠ disabled ≠ deleted

An item not consulted for this decision can remain eligible for a later permitted use. Disabling a predictor blocks its use within a selected scope without deleting separately authorised evidence. Deleting a retained item ends that retention; it does not by itself remove every derived effect. The subtraction here excludes optional conditioning from the proposed comparison. It does not delete history or disable binding constraints.

This is not a measured ablation or proof that a renderer would produce identical outputs without them. It identifies a redundancy in the argument for using them here. If a system claims they improve this view, it needs to say what improvement they contribute rather than point to the fact that they were available.

The conclusion is also local. An observation unused today might have a legitimate continuity role in another task. Chapter 23’s retention decision remains separate from current use. A record should not be erased merely because this view did not need it, nor retained indefinitely because some possible future use can be imagined.

Additional state might help where Priya has supplied less detail, where a permitted scoped inference supports an appropriate optional view, or where a declared standing requirement saves her from repeating it. Those are reasons to investigate a benefit, not to assume one. Requiring her to restate an accessibility requirement at every interaction would not become good design merely by reducing the number of retained items.

The useful discipline is to ask what fails without an item. Which communication decision needs it? Could the source, current task or a narrower declared rule supply what is missing? What evidence supports its scope, what use is permitted, and how long does that justification last? The answer should identify a contribution, not a theory of Priya.

Separation is necessary to the proposal, not a safety proof

The information still needs distinct origins and roles:

declared ≠ observed ≠ inferred
durable history ≠ current context ≠ session overlay
goal ≠ intervention authority
policy ≠ interpretation of policy
supported scope ≠ applied scope

A combined description such as an expert who likes detail and dislikes distraction loses those distinctions. It cannot tell the channel whether an equation is decisive, whether the current request permits folding it, or whether a receiving service confirmed the boundary’s expiry. It also gives Priya no precise target for correction.

But separate namespaces cannot guarantee correct composition. A wrongly scoped inference can remain neatly labelled. An adapter can misread a boundary while preserving its origin field. Two valid instructions can conflict in the same situation. Separation makes the relevant questions representable; it does not answer them automatically.

Consider another constructed review. Priya asks for an externally shareable view on her phone. The inputs are accurate, within their supported scopes and legitimately available:

Declared:  terse first views
Observed:  she normally expands derivations herself
Inferred:  familiarity with consensus basics in this review context
Current:   viewing on mobile
Boundary:  do not expose internal project history

The composer combines brevity, recoverable detail, assumed familiarity and limited screen space into an aggressively compressed explanation. It correctly excludes internal project history. But it also drops a shareable timing qualification that determines whether the recommendation applies. No input changed namespace or gained a broader scope. The combined representation still failed its preservation obligation.

local validity ≠ compositional validity

Valid inputs do not establish that the resulting communication preserves what this decision requires. The example illustrates that distinction; it does not measure the failure rate of a composed system.

Nor has the book earned a universal precedence ordering. Preservation and accessibility constrain the view. Within those constraints, applicability and authority must be established for the particular instructions. In this session, the current request specifies which supporting material may be folded; the overlay applies that permission. If it conflicts with an applicable standing requirement and no accepted priority resolves the conflict, the channel should expose the choice rather than promote temporary state into authority.

Correction retains the same limits. Priya can reopen the supporting steps for this view without creating a standing preference. A change to a standing rule needs a distinct declared decision. A corrected inference must reach its dependent uses; a changed portable boundary must reach its enforcement points. Neither repairing a record nor keeping records separate proves that the effects have been repaired.

The claim that this organisation produces more local correction and fewer silent conflict resolutions remains untested. The construction specifies the intended discipline without establishing its comparative success.

More state creates more obligations

An additional belief is not only another opportunity to adapt. It needs a permitted origin, a supported scope, a retention justification, current applicability and a limit on use. If it affects communication, its contribution must be inspectable and its correction must reach the affected state.

Combination adds another question. Two individually scoped items might be used together to infer something neither supports alone. Turning Priya’s repeated derivation expansions and a quiet-feed rule into she dislikes distraction creates a new inference. It cannot borrow authority from the fact that both inputs were legitimately retained. Chapter 27’s rule about scope expansion also applies to what composition produces.

This is why counting records cannot define a minimum. Six richly consequential records may be more demanding than twenty narrow ones. Removing origin, scope or expiry metadata merely to make an item smaller can make the boundary harder to inspect or end without reducing what it does. The amount of state matters together with what it reveals, what it changes and what repairing it requires.

The chapter therefore proposes a burden of justification for additions. A possible benefit does not authorise collection. Authorised collection does not establish benefit. A richer view that Priya happens to accept does not prove that every retained input was necessary or that its influence should persist.

available state ≠ necessary state
authorised state ≠ beneficial state
unused now ≠ useless forever

None of these distinctions establishes that fewer records always communicate better. They prevent the opposite assumption from becoming the architecture’s default.

A path without optional personal conditioning

Priya should be able to ask for communication without optional prediction, inferred preferences, historical adaptation or other non-required learned personal state. Those inputs are excluded from conditioning the view, not necessarily deleted. That path still has a task and constraints:

source
+ explicit current request and conditions
+ preservation and accessibility obligations
+ binding user policy
+ privacy / data-use boundaries and authority restrictions
        ↓
communication without optional personal conditioning

A non-personalised path is not an ungoverned path.

This is a usable alternative, not a degraded mode defined by the absence of a profile. It can still respond to the explicit ten-minute request, preserve the decisive qualification and make the full calculation available.

Excluding optional personal conditioning must also reach its effects in the current view. A cached summary shaped by an inferred preference still carries that influence even if the belief is no longer supplied as an input. The channel must replace the affected representation or establish that the influence has been removed; otherwise it must report the limit. Removing an input from a list is not proof that the communication no longer depends on it.

stop using ≠ invalidate derivatives ≠ recompute current state ≠ undo past effects

Even repairing the current view cannot recreate the history in which the predictor was never active. Its earlier contributions to cached representations, subscription ordering or dependent beliefs need their own accounting, with unresolved effects still reported as unresolved.

The scope of the switch must be honest. Disabling optional personal conditioning does not silently revoke Priya’s protective boundary, discard a standing accessibility requirement or authorise learning from behaviour that her temporary rule excluded. Those obligations may require retained policy and authority records. The alternative uses less optional personal state; it is not a claim that the system holds or uses no state about her at all.

If she also wants an existing rule revoked, that is a separate decision whose effects must be confirmed. Changing how a view is conditioned cannot be reported as deletion of history, cessation of all future inference or removal of every distributed control.

The alternative can also support inspection. Priya may compare a view with optional personal conditioning and one without it, keeping the current task and binding constraints fixed. As in Chapter 27, a newly generated difference is not automatically attributable to the removed state. The comparison is a proposed aid; demonstrating a communication benefit requires a separate account of what changed and why.

What should survive this session?

At half past nine, the temporary presentation and friction permissions end if they have not already ended with the review. The portable rule’s end must reach the applied controls; unconfirmed removal remains visible as unconfirmed.

The session also cannot survive indirectly as a new trait. Priya’s use of a folded calculation during an explicitly brief review is evidence about that conditioned interaction, not a standing preference for short mathematics. Any legitimately retained observation carries the temporary condition’s provenance. No record may silently promote the behaviour into the policy the overlay helped cause.

The standing requirement remains unless she revises it. The earlier history retains its own justified lifetime. Unused inferences do not become more authoritative merely because they have sat in storage through another successful-looking session.

This is continuity with a limit. The system should preserve what has a legitimate continuing role and shed what belonged to now, without making today’s construction an excuse to accumulate another layer tomorrow.

How much is enough?

Part III has proposed increasingly capable mediation: continuity, temporary adaptation, authorised friction, portable boundaries and inspection of consequential beliefs. Together they also create increasingly demanding obligations. A fluent result does not show that the accumulated state was necessary, that its scope was correct, or that its effects ended when promised.

The constructed review suggests that a current request and a few applicable constraints can sometimes supply the argument for a view without a prediction about the person. It establishes no measured sufficiency, optimal record count or benefit from reducing state. Whether richer state improves communication, and under which conditions, remains an open empirical question.

The next part examines the stronger temptation: if a system can anticipate Priya’s communication choices, perhaps it can predict more of her behaviour. That claim needs its own target and evidence. It does not follow from this session, and prediction would still not confer permission to speak or act for her.

personal state helps communication
        ≠
personal state predicts the person
        ≠
system reproduces the person

Part IV must say what is being measured when the target moves from improving an interaction to predicting a person’s behaviour. Part III ends with a narrower burden of justification: How little of you does good communication actually require?