When Should the AI Speak?

The contradiction arrives at 14:40 — a new paper undermining the assumption in the code currently open in the editor. The system knows: the match is high-confidence, the evidence is attached, the resolution is safe at BRIEF, the novelty is real. It also knows the reader is presenting in a meeting until 15:30. Delivering now means a buzzing pocket, a glanced phone, a derailed sentence — the information consumed at the worst possible moment, its value converted into damage. Delivering at 15:35 means the reader finishes the meeting, opens the editor, and meets the contradiction exactly when action is possible.

Everything upstream is settled. Qualification (16), novelty (15), resolution (17), safety (18) — done. The only remaining question is the chapter’s entire subject:

Given information that qualifies, is novel enough, has a safe resolution, and deserves some attention, when should it enter the person’s foreground?

qualified information + safe resolution
      ↓
delivery policy
      ↓
NOW / QUEUE / DIGEST / SIDECAR

One terminological lock first, because two chapters now border the same word: Chapter 17’s NOTHING means this item does not deserve attention at all. Chapter 19 never reuses NOTHING for delay — its vocabulary is NOW (foreground now), QUEUE (hold for the next natural break), DIGEST (batch into a scheduled roundup), SIDECAR/AVAILABLE (present without demanding). An item here has already earned attention; timing decides its moment, never its worth. Timing is therefore not another relevance judgement: a critical contradiction can be DIGEST (terrible moment, non-expiring value) while a modest API rename can be NOW (the reader is blocked on exactly that call this minute).

Two costs, six properties

The policy balances two failures, neither acceptable as doctrine:

INTERRUPT TOO EARLY → context switching, distraction, lost work
DELIVER TOO LATE    → missed decisions, duplicated work, expired action

“Always proactive” and “never interrupt” both fail — the first taxes every delivery, the second lets expiring value rot. The decision runs on six separated properties, because one importance score cannot express their combinations:

RELEVANCE · NOVELTY · RELATION · RESOLUTION (all inherited)
URGENCY — how quickly does its value decay?
INTERRUPTION COST — what does foregrounding cost right now?

High value with low urgency (an independent replication of a used paper) digests. Modest value with high urgency (the argument renamed today in the API being implemented) interrupts. High importance at a terrible moment (the contradiction during the presentation) queues. The combinations are the policy; any single-score router collapses them. One more separation constrains all three: urgency is not evidential confidence. An item may be relevant, novel, and urgent while its evidence is unconfirmed — high urgency must never promote an unverified claim into NOW as though established. The uncertainty travels with the item (“potential breaking change detected; evidence not yet confirmed”), and timing and epistemic status remain separate dimensions through every route.

State for the decision follows the least-inference doctrine: use the least inferential state sufficient to make the timing decision. Observable or declared context only — current task, active document, meeting topic, explicit focus mode, deadline proximity, mid-action fragility, whether the candidate affects current work, whether it expires, whether the action is reversible. Never inferred stress, overload, or boredom. ProMemAssist (Pu et al., UIST 2025, DOI 10.1145/3746059.3747770, verified via arXiv 2507.21378: smart-glasses working-memory modelling from multimodal signals, timing predictor balancing assistance value against interruption cost, 12 participants on cognitively demanding tasks, more selective assistance and higher engagement than an LLM baseline) is used exactly as what it is — a mechanism precedent for value-vs-cost timing with a small wearable-task study — not as evidence for a general radar timing policy, and its working-memory inference is discussed, not inherited. Krisam et al. (SIGHCI 2025, verified via AIS eLibrary: nine interviews, design-science requirements, LLM topic-relevance prototype for refocusing multitasking meeting participants) is design evidence for refocus-on-relevance alerting — not a demonstrated distraction reduction. Lam et al.’s Just-In-Time Objectives (CHI 2026, DOI 10.1145/3772318.3790713; reported 66–86% preference wins over generic outputs with objectives exposed as editable first-class objects — full read pending, numbers taken as reported) feeds the policy differently than its name suggests: current-objective modelling as input (objective + relation + urgency + cost → route), never as proof that JIT output should interrupt.

Inspectable policy, correctable interruptions

Timing policy is user-set and user-questionable, because a hidden engagement optimiser would maximise foregrounding and call it relevance:

Focus mode: interrupt only for
[x] current-task blockers · [x] expiring information
[ ] contradictions · [ ] new implementations
Everything else: morning digest

and after every interruption, an account:

Why did you interrupt me?
Because: affects your open task; expires before
your next digest; high match confidence.
[send these later] [never interrupt for this relation]

Corrections — snoozes, “tell me now,” “not for this” — are training signal for the policy and first-class measures in EXP-19, which freezes the entire upstream stack (source through preservation result) and varies only delivery: A immediate, B fixed batching, C user-pull only, D task-aware (objective + urgency), E cost-aware task timing (+ observable interruption cost). Scored separately: timely useful deliveries, harmful interruptions, important-but-late items, opens/actions, time-to-action, task disruption, and every manual correction type — never click-through rate, since an immediately read notification may have been a terrible interruption. Six hard cases required: expiring breakage, important-but-non-expiring support, current-task contradiction, future-task relevance, repetitive-but-new updates, fragile-task moments. Failure criteria: E ties B/D (cost modelling adds nothing); corrections cluster unusably (policy unlearnable); pull-only ties all (proactivity unnecessary); interruption harm unmeasured by opens (validates the anti-CTR rule). What timing success would not justify: any upstream claim — relevance, novelty, safety all arrive certified, and timing takes them on trust.

What this chapter earned

Delivery timing modelled separately from everything it routes: qualified, novel, safely resolved information enters NOW/QUEUE/DIGEST/SIDECAR by urgency against observable interruption cost under least-inferential state, with user-set policy and per-interruption accounts. The book’s information pipeline is now complete end to end — and every stage reduces a different waste. Part II can assemble its Radar.

References

• Pu, K. et al. (2025). ProMemAssist. UIST 2025 (DOI 10.1145/3746059.3747770; arXiv:2507.21378). Verified via arXiv: WM modelling + timing predictor; 12 participants; selective assistance + engagement vs LLM baseline. Used as value-vs-cost precedent with small-study fence; WM inference not inherited.
• Krisam, C. et al. (2025). From FOMO to Focus. SIGHCI 2025 Proceedings. Verified via AIS eLibrary: 9 interviews, design requirements, prototype. Used as design evidence; no distraction-reduction claim taken.
• Lam et al. (CHI 2026). Just-In-Time Objectives. DOI 10.1145/3772318.3790713. Numbers as reported (66–86% wins, editable objectives); full read pending. Used as current-objective input, not interruption evidence.

Proposed experiment EXP-19: delivery timing under full freeze

Status: PROPOSED. Per the design above (A–E, frozen stack, six hard cases, separated harm/benefit measures, corrections as signal, anti-CTR rule).