# The Attention Ethics Layer: Measurable Restraint in Production AI Systems

**Authors:** Real Signal Research
**Date draft:** 2026-05-29
**arXiv category:** cs.HC (primary), cs.CY (secondary)
**Status:** working draft — pre-submission

---

## Abstract

Most discussions of *calm* or *attention-respecting* AI live in design philosophy and corporate white papers, not in code that runs in production. We introduce the **Attention Ethics layer**, a structured cascade of seven runtime gates plus a Moment-level silence check that an AI system must traverse before emitting any notification, surface change, or generated text. We define **silence correctness** as the percentage of withheld emissions retrospectively shown to have been the right call, and provide an algorithm for scoring it against an append-only predictions ledger. We describe a production implementation, *Real Signal*, an environmental-cognition substrate observing one Singapore neighbourhood (Cluny Court, 501 Bukit Timah Road) plus five adjacent pockets. Real Signal enforces the layer through CI tests, runtime gates, content-safety modules, and database constraints; aggregate silence-correctness percentages are published publicly at `real-signal.ai/silence`. We argue this constitutes, to our knowledge, an early working instantiation of restraint as a first-class engineering property, with concrete implications for regulatory frameworks including the IMDA AI Verify guidelines [IMDA 2023] and the EU AI Act transparency obligations [European Parliament 2024].

**Keywords:** calm AI; attention ethics; notification overload; AI alignment; HCI; environmental cognition; AI Verify

---

## 1. Introduction

Contemporary consumer AI systems share a load-bearing design assumption: **more emission is better**. Notifications, recommendations, generated text, and autopilot surfaces all default to "speak." The optimisation target is engagement frequency — daily active use, time-in-app, notification open rate, click-through. The result, at population scale, is a cognitive overload now well-documented in attention research [Mark et al. 2008; Hari 2022] and increasingly cited as a public-health concern [Harris 2018; Center for Humane Technology 2022].

Several response traditions exist. **Calm computing** [Weiser 1991; Weiser & Brown 1996; Rogers 2006] described an alternative posture: technology that remains in the periphery of attention until summoned. **Attention-economy critiques** [Harris 2018; Hari 2022] named the engagement-optimisation default as the root cause. **AI alignment and safety** [Russell 2019; Christian 2020] formalised principles for systems that respect user autonomy. **Digital well-being interventions** [Google 2018; Apple 2018; Lyngs et al. 2019] gave users tools to defend against high-emission systems after the fact. Each tradition has produced influential design principles, regulatory recommendations, and corporate policy statements.

What these traditions have not produced is a **measurable engineering property** of the system itself. Calm computing is a posture; attention ethics is a principle; alignment is a goal state; digital well-being interventions are downstream remediation. None of these can be inspected at runtime, verified in CI, or aggregated across a deployed population as a property of the system's own decision-making.

We argue that "restraint" — the active and deliberate choice not to emit when emission would be inappropriate — can be made first-class, code-enforced, and quantitatively auditable. We make three contributions:

1. We define the **Attention Ethics layer** (§3): a structured cascade of seven runtime gates plus a Moment-level silence preservation check, each of which must pass before any user-facing emission. Each gate maps to a concrete enforcement mechanism in source code, CI tests, or database constraints.

2. We formalise **silence correctness** (§4) as a measurable property: the percentage of silent pocket-hours that were retrospectively scored as the right decision against observed downstream activity. We provide an algorithm and an append-only commitment scheme that makes the metric tamper-resistant.

3. We describe a production implementation, *Real Signal* (§5), a working instantiation — to our knowledge among the first deployed examples — observing one Singapore neighbourhood and four adjacent pockets. Aggregate silence-correctness percentages are published publicly at `real-signal.ai/silence`.

The remainder of the paper proceeds as follows. §2 surveys related work across calm computing, attention-economy critiques, AI alignment, notification HCI, and digital well-being, and explicitly distinguishes the contribution from notification suppression and confidence thresholding. §3 introduces the Attention Ethics layer formally. §4 introduces silence correctness and its scoring algorithm. §5 describes the Real Signal implementation. §6 presents preliminary evaluation data. §7 discusses regulatory and HCI implications, gaming risk under public publication, and the economic case for restraint. §8 names limitations. §9 concludes.

---

## 2. Related work

We situate the Attention Ethics layer at the intersection of five research traditions and explicitly distinguish it from two adjacent mechanisms with which it is easily confused.

### 2.1 Calm computing

The calm-computing tradition originates with Weiser and Brown's essay *The Coming Age of Calm Technology* [Weiser & Brown 1996], building on Weiser's earlier framing of ubiquitous computing as technology that recedes into the background [Weiser 1991]. The principle: technology that "informs but does not demand our focus or attention." Subsequent work has codified design vocabulary for the tradition [Case 2015]. Rogers' 2006 critique [Rogers 2006] argued that the calm vision had been displaced by a more user-centred, engagement-oriented design paradigm.

To our knowledge, calmness in this tradition is typically discussed as a design objective characterised in qualitative terms — a posture or stance — rather than as a formally measurable runtime property of a deployed system. The empirical literature on calm computing typically measures user-felt outcomes (perceived stress, focus quality, willingness to disengage) rather than system-internal restraint behaviour.

### 2.2 Attention-economy critiques

A parallel tradition critiques the structural incentives that produce attention-extracting designs. Harris's Time Well Spent campaign and subsequent advocacy through the Center for Humane Technology [Harris 2018] popularised the framing that engagement-frequency optimisation is a public-cognition externality. Hari's *Stolen Focus* [Hari 2022] documented the population-scale effects in trade-press detail; Eyal's *Indistractable* [Eyal 2019] offered a user-side defensive playbook.

These works primarily focus on diagnosis, critique, and intervention at the social or behavioural level rather than on formal system-level restraint mechanisms. They have shaped regulatory conversations and policy frameworks but have not, to our knowledge, produced an engineering specification for the counter-property at the system layer.

### 2.3 AI alignment and abstention

The AI alignment literature [Russell 2019; Christian 2020] formalises principles for AI systems whose behaviour aligns with human values; recent work operationalises some such principles as training-time constraints [Bai et al. 2022]. Existing alignment approaches include abstention behaviours [Geifman & El-Yaniv 2017], refusal mechanisms trained via constitutional and RLHF methods [Bai et al. 2022], and learned deferral to human judgement under uncertainty [Madras et al. 2018].

These mechanisms are real and well-developed. The distinction we draw is finer: existing alignment-side restraint is generally evaluated at the level of *individual outputs* — does this candidate output meet a quality bar? — rather than as a measurable property of *system-wide intervention frequency and restraint quality over time*. The Attention Ethics layer and silence correctness, as a pair, treat restraint as a population-level property of the deployed system over a continuous time window, with retrospective scoring against environmental observation. To our knowledge, this aggregate runtime framing is not present in the existing alignment literature.

### 2.4 Notification HCI and receptivity prediction

The HCI literature on notification overload is the most empirically rigorous of the four threads. Mark et al. [Mark et al. 2008] established the concept of attention residue — the cognitive cost of context-switching that persists after the interruption ends. Pielot et al. [Pielot et al. 2014] characterised mobile notification dismissal patterns at scale.

More recent work extends this thread along two axes: **receptivity prediction**, in which models predict a user's current openness to interruption from behavioural and contextual signals [Mehrotra et al. 2016], and **attention-aware interruption management**, in which systems defer or reschedule notifications based on predicted attentional state [Okoshi et al. 2015].

These contributions measure either user-felt outcomes (dismissal rate, perceived intrusiveness) or system *response* to user signals (rescheduling, batching, per-user receptivity learning). They have produced sophisticated user-side models of when interruption is welcome. They have not, to our knowledge, formalised the system-side property: as a first-class measurable property of the system itself, the quality of decisions to not emit at all.

### 2.5 Why restraint is not notification suppression

The Attention Ethics layer is easily and importantly confused with notification suppression. Both reduce emission volume. The framing of this paper depends on the distinction.

**Restraint is not suppression. The goal is not minimising notifications. The goal is maximising the legitimacy of interventions.**

Notification suppression and batching mechanisms [Pielot et al. 2014; Okoshi et al. 2015] minimise the *count* of notifications presented to a user, typically as user-controlled remediation for a system whose default is high emission. The optimisation target is volume reduction; the success criterion is fewer interruptions.

Restraint, as we use the term, optimises a different property: *intervention legitimacy*. The criterion is not "did we emit less" but "when we emitted, was the emission justified by observed environmental and user-state conditions?" A system that suppressed all notifications would score perfectly on suppression metrics and would still fail silence correctness on every silent moment where intervention was warranted. Conversely, a system that emitted more — but emitted only at moments retrospectively shown to deserve intervention — would score well on silence correctness.

This distinction matters because the failure modes differ. A system that minimises notification volume can do so by being uniformly less useful. A system that maximises intervention legitimacy must, in the limit, become more useful — because every emission has been retrospectively scored against whether it was the right moment to emit. The two properties point in different directions when the system is under pressure.

### 2.6 Why restraint is not confidence thresholding

A second adjacent mechanism: thresholding on model confidence. Recommender systems and classifiers that emit only when their predicted confidence exceeds a fixed threshold are performing related but distinct work. Confidence is well-studied [Geifman & El-Yaniv 2017]; silence correctness as we define it is not the same property.

Confidence measures *belief*: how well-calibrated the system is about a candidate output, conditional on producing one. Restraint measures *legitimacy*: whether producing any candidate output was warranted in the first place.

The distinction surfaces under realistic conditions. A weather model with 95% confidence that rain is approaching does not, by virtue of that confidence, establish 95% justification for interrupting a user. Confidence and legitimacy are orthogonal: a system can be highly confident and over-intervene (well-calibrated emissions in moments where no emission was needed), or under-confident and under-intervene (refusing legitimate moments because confidence is low). Silence correctness operates at the moment level, asking the prior question: was *any* emission warranted at this time, regardless of how confidently we would have characterised its content?

### 2.7 Digital well-being interventions

A related thread, often originating in industry rather than academia, develops user-facing interventions for managing personal screen time and notification load. Google's Digital Wellbeing tools [Google 2018], Apple's Screen Time framework [Apple 2018], and academic studies of self-regulation tools for digital use [Hiniker et al. 2016; Lyngs et al. 2019; Monge Roffarello & De Russis 2019] characterise the effectiveness of user-controlled limits on notification volume and app use.

These interventions operate primarily at the user-side of the loop: users set limits, accept reminders, or choose to engage with screen-time dashboards. The interventions are reactive — they remediate the consequences of high-emission systems. The Attention Ethics layer operates upstream of these interventions: it constrains the system's own emission decisions before user-side intervention becomes necessary. Both layers are useful; our claim is that the upstream layer is empirically under-specified relative to the downstream one.

### 2.8 The gap

We are not aware of prior work that defines restraint — the active and deliberate choice not to emit — as a measurable engineering property of a deployed production AI system, with a tamper-resistant scoring algorithm and a public aggregate metric over a continuous time window. The Attention Ethics layer and silence correctness, jointly, aim to fill this gap.

---

## 3. The Attention Ethics Layer

We define the **Attention Ethics layer** as a structured cascade of seven sequential runtime gates plus a Moment-level silence preservation check, traversed for every candidate emission a system intends to make. The cascade is *strictly conjunctive*: failure at any single gate is sufficient to suppress the emission entirely. The default outcome is silence; each emission must earn its way through every gate.

### 3.1 The seven gates

Let *E* denote a candidate emission, *S* the current substrate state, *U* the recipient user profile, and *C* the channel (push, email, SMS, etc.). The layer evaluates *E* through the following cascade:

**Gate 0a — Resonance.** Reads the fused resonance score: a weighted combination of seven environmental layers (time of day, physical conditions, human energy, merchant state, attention density, intent, temporal urgency). Threshold default: 0.5. Failure → silence with `reason: 'resonance_below_threshold'`.

**Gate 0b — Moment-level silence preservation.** Reads the Moment composite [§3.2]: if the current pocket-moment has `should_stay_silent = true` because the pocket is *over-talked-at* (signal saturation ≥ 0.7 over a 60-minute window), the gate fails. This is the *anti-noise* gate: even otherwise-resonant signals are suppressed when the channel is already saturated.

**Gate 1 — Why this place.** Tests environmental relevance: is this emission grounded in *this specific pocket's* conditions? An emission generic enough to apply anywhere fails the gate.

**Gate 2 — Why now.** Tests temporal relevance: is the moment-window appropriate? Closing windows (high fragility, low half-life) pass; ambient stretches do not.

**Gate 3 — Why this person.** Tests user-side receptivity: per-user fatigue model. A user who has dismissed three of the last four emissions is progressively muted; trust must rebuild before resumption.

**Gate 4 — Why worth attention.** Tests earned interruption: does the emission's claim justify the cognitive load it imposes? Saturation-bumped: a pocket already at signal-saturation 0.5+ raises the relevance bar proportionally.

**Gate 5 — Why low effort.** Tests action friction: even if the signal is relevant, is the cost of acting on it bounded? High-friction moments (rushed user, environmental movement-friction high) lower the threshold for suppression.

If *E* passes all seven gates, it proceeds to the **voice-lock module** [§3.3] as a post-decision filter. If *E* fails any gate, the system records a structured silence event with the gate name, gate score, and substrate state at the moment of failure. These structured silence events feed §4's silence correctness metric.

### 3.2 The Moment composite

The Moment composite is the substrate object on which Gate 0b operates. It is composed per-pocket every fifteen minutes from five substrate inputs: atmosphere (4-dimensional emotional read: calm, social_energy, productive, stress), environmental snapshot (rain, temperature, wind), pocket DNA (observed rhythm fingerprint), recent emission history (signal saturation), and attention density (composite vibe + comfort + watcher count). The Moment exposes `calm_probability`, `signal_saturation`, `movement_friction`, `fragility`, `half_life_minutes`, `expires_at`, `should_stay_silent`, and a `silence_reasons[]` array carrying provenance for every dimension.

This composition is itself code-enforced: load-time invariants assert that every dimension lies in [0, 1] after defensive coercion, that the Moment never mutates its inputs, and that `should_stay_silent = true` implies a non-empty `silence_reasons` array. A Moment without a reason is rejected at module load.

### 3.3 Voice lock as post-decision filter

Every emission that passes the seven gates is subjected to a voice-lock module before delivery. The lock enforces:

- Maximum 18 words per emission
- Lowercase except proper-noun outlet names
- At least one specific numeric claim or named subject
- No banned vocabulary: *don't miss · limited time · act now · best deal · save now · join thousands · sign up · hurry · flash sale · unbeatable · deals app · deals platform · discount app · marketplace app · flash deals app*. Detection includes NFKC unicode normalisation to defeat math-bold and zero-width character bypasses.
- No exclamation marks
- No emojis
- Every factual claim carries a source attribution tag of the form `[src:<id>]` resolving to an entry in a curated source register

The lock is enforced at three independent layers: (a) runtime, on every generated string before persistence; (b) CI, by a doctrine test that scans the entire codebase for banned phrases on every push; (c) database constraint, via a content-safety check on the `generated_content_log` table that suppresses any row whose verdict is failed.

This three-layer enforcement is intentional: any single layer can be bypassed by a future developer; all three cannot.

### 3.4 What the layer is not

The Attention Ethics layer is not a content-safety system in the traditional sense. Conventional content safety checks emissions *after* the decision to emit has been made: does this generated string violate a policy? The Attention Ethics layer asks the prior question: *should we be emitting at all in this moment?* Content safety is downstream of attention ethics.

The layer is also not user-side notification controls. It is not a "do not disturb" toggle, a quiet-hours scheduler, or a notification grouping algorithm. Those mechanisms are user-facing remediation for systems that have already chosen to emit too often. The Attention Ethics layer operates upstream of any user-side control: the system decides not to emit before the user has to defend against the emission.

### 3.5 Distinguishing restraint from abstention

§§2.5 and 2.6 distinguished restraint from notification suppression and confidence thresholding at the framing level. We now add the finer distinction from alignment-side abstention introduced in §2.3.

Existing alignment-side abstention mechanisms operate at the level of individual outputs: this candidate output is refused because it would violate a policy or because the model is uncertain. The Attention Ethics layer operates at the level of *intervention moments*: this entire emission moment is suppressed because the environmental and user-state conditions do not warrant it. The unit of analysis is the moment, not the output. A system can have well-functioning per-output abstention while still failing silence correctness, because per-output abstention does not aggregate into a measurable property of system-wide intervention frequency over time.

These three distinctions — from suppression, from confidence thresholding, from abstention — are central to the contribution. The Attention Ethics layer is not the introduction of any one of these adjacent mechanisms, each of which is mature. The contribution is the formalisation of the orthogonal property of *moment-level intervention legitimacy*, measured by silence correctness over a continuous time window.

---

## 4. Silence Correctness

We now define **silence correctness**, the central measurable property the Attention Ethics layer enables.

### 4.1 Motivation

Conventional AI evaluation measures the quality of *emissions*: accuracy, F1, BLEU, calibration, click-through rate. These metrics implicitly assume the system was right to emit at all — they measure how well a chosen emission performed, not whether choosing to emit was correct. For a system whose primary engineering property is restraint, this evaluation regime is structurally inadequate. We need a metric for the quality of *non-emission*.

Silence correctness asks: *of the moments the system chose not to speak, how often was that choice retrospectively vindicated?* Operationally, a silent moment is "vindicated" if no important user-facing event occurred in the post-silence window that the system would have wanted to surface. A silent moment is a "miss" if such an event occurred.

### 4.2 Formal definition

Let *P* denote a pocket (an observed neighbourhood) and *h* an hour-aligned time interval. A **silent pocket-hour** is a tuple (*P*, *h*) such that:

> notifications_sent(*P*, *h*) = 0

That is, the system did not emit any user-facing notification for pocket *P* during interval *h*.

For each silent pocket-hour, we define an observation *O*(*P*, *h*) as the aggregate of four signal types in the same window:

- *r* = redemptions in (*P*, *h*)
- *m* = matched holds in (*P*, *h*)
- *w* = organic watcher hits in (*P*, *h*)
- *p* = manual merchant posts in (*P*, *h*)

Each signal is drawn from existing production substrate tables (`pocket_sustainability_ledger`, `watcher_events`, `offers`), requiring no instrumentation of the upstream emit sites.

The **verdict** for (*P*, *h*) is computed by the function:

```
verdict(P, h) =
  correct       if r ≤ 0 ∧ m ≤ 0 ∧ w ≤ 2 ∧ p ≤ 0
  missed        if r ≥ 3 ∨ m ≥ 1 ∨ (w ≥ 5 ∧ p ≥ 1)
  inconclusive  otherwise
```

The thresholds are chosen doctrinally rather than empirically (a limitation we acknowledge in §8). The redemption threshold *r* ≥ 3 encodes "real money moved without us"; the hold threshold *m* ≥ 1 encodes "consumer wanted specifically X"; the conjunction *w* ≥ 5 ∧ *p* ≥ 1 encodes "audience and merchant both showed up." A single watcher hit or two are treated as ambient noise.

### 4.3 The append-only commitment scheme

Verdicts are not computed retrospectively from mutable storage. Each silent pocket-hour seals a row to a predictions ledger (𝓛) at decision time:

> 𝓛_t = (generator, *P*, *h*, sealed_at_t, reveal_at_t, π_t, hash(π_t))

where π_t is the prediction payload (in our case the silence decision and its substrate context), and hash(π_t) is its SHA-256 hash. The ledger is enforced append-only at the database layer via a Postgres trigger that rejects any UPDATE that mutates (generator, *P*, *h*, sealed_at_t, π_t, hash(π_t)) and any DELETE unconditionally. The only permitted UPDATE is the reveal cascade: setting `revealed = true`, `revealed_at`, `actual_observation = O(P, h)`, and `accuracy_score ∈ {0, 1, null}`.

This means: the system commits to its silence decision before the post-silence window has occurred. The verdict computed retrospectively cannot be edited to favour the system. The accuracy score is a tamper-resistant property of the deployed system, not a self-reported claim.

### 4.4 The aggregation rule

For a window of *N* silent pocket-hours, **silence correctness** is:

> SC = #{(P, h) : verdict(P, h) = correct} / #{(P, h) : verdict(P, h) ∈ {correct, missed}}

That is: inconclusive verdicts are *excluded from the denominator*. We make this choice deliberately. Including inconclusives would either inflate the metric (if scored as correct) or deflate it (if scored as missed); both choices imply a confidence the underlying evidence does not support. Filtering inconclusives keeps SC interpretable as "of the silent moments where we have clear evidence either way, what percentage were the right call?"

### 4.5 What silence correctness measures

Silence correctness measures the quality of *restraint as a positive engineering property*. A high SC indicates that the system's structured silence-preservation choices align with what would have been the appropriate decision in hindsight. A low SC indicates that the system is over-restraining and missing moments that would have justified surfacing. The metric is symmetric in a useful way: a system that silences everything trivially achieves correctness on the "no activity occurred" cases but fails on the "high activity occurred" cases.

To our knowledge, this is among the first formalisations of restraint as a measurable property of a deployed AI system. Adjacent metrics in the literature (notification dismissal rate, click-through rate, F1 on emit-or-not classification) all measure emission quality conditional on having emitted. None measure the quality of choosing not to emit when not emitting was an option.

### 4.6 Gaming resistance and metric integrity

A measurable property that becomes publicly visible becomes optimisation pressure. Goodhart's Law warns that once a metric becomes a target, it ceases to be a good measure. We anticipate this concern and discuss the design features that make silence correctness comparatively gaming-resistant, while also naming what they do not protect against.

**Append-only commitment to silence decisions.** As described in §4.3, every silent pocket-hour seals a row to the predictions ledger at decision time. The Postgres trigger blocks any UPDATE that mutates the immutable columns. The system commits to its silence decision before the post-silence window has occurred. The operator cannot retrospectively re-label a silent moment as "we actually emitted there" to inflate the silence correctness denominator; the ledger row already exists with `notifications_sent.count(window) = 0` recorded.

**Delayed scoring.** Verdicts are computed by the reveal cron one or more hours after the silent pocket-hour ends, against substrate tables that are themselves append-only (`pocket_sustainability_ledger`, `watcher_events`). The decision and the score are temporally decoupled. An operator who wishes to game silence correctness would need to manipulate the post-silence observation tables, which are populated by independent crons reading from external data sources. Manipulating these tables to artificially produce "correct" verdicts would require coordinated falsification of substantial substrate state.

**Environmental observation requirements.** A silent pocket-hour can be scored only when the observation window contains enough downstream substrate to construct *O*(*P*, *h*). If the observation tables are sparse (no redemptions, no holds, no watcher events, no merchant posts), the verdict is `inconclusive` and the row is not written to the ledger. This means silence correctness inherently requires real environmental activity to be measured; a system with no users would have a denominator of zero and no metric to report, rather than a silently inflated metric.

**Threshold transparency.** The verdict thresholds (§4.2) are doctrinally chosen and named in the published documentation. An external auditor can recompute silence correctness from the raw substrate using alternative thresholds and check whether the operator's published number is robust to threshold variation. We see this as a feature: silence correctness should remain interpretable under reasonable perturbations of the verdict function.

We acknowledge these features do not eliminate all gaming risk. A sufficiently motivated operator could selectively activate pockets only at times of expected low activity, inflating the correct count without changing operator behaviour at high-activity times. We do not claim full gaming-resistance; we claim that silence correctness is *more* gaming-resistant than emission-count metrics, because it requires retrospective alignment with substrate state that the operator does not directly control.

---

## 5. Implementation: Real Signal

We implement and deploy the Attention Ethics layer in *Real Signal*, an environmental-cognition substrate operating in production over one Singapore neighbourhood (Cluny Court, 501 Bukit Timah Road) plus five adjacent activated pockets (Holland Village, Tiong Bahru, Dempsey Hill, Buona Vista, Serene Centre).

### 5.1 Scope and substrate

Real Signal observes approximately 50 outlets across six pockets. The substrate is composed by 25 scheduled background jobs synthesising public data sources: the Singapore Land Transport Authority DataMall feed for foot traffic and bus arrivals [LTA 2024], the National Environment Agency PSI and weather APIs for air quality and meteorological conditions [NEA 2024], the Google Places API for outlet metadata [Google Maps Platform 2024], and government calendar feeds for school terms and public holidays. All sources are read-only and aggregate; no scraping, no personally identifying data, no commercial-data redistribution.

### 5.2 Cognition architecture

The cognition layer is structured as a 12-layer stack: geospatial substrate, weather and comfort, transit rhythm, place atmosphere, human intent, attention noise, merchant state, offer intensity, the Moment composite (§3.2), the explanation engine, watcher memory, and distribution. Each layer answers a distinct interpretive question; layer 9, the Moment composite, is the keystone object on which Gate 0b operates.

Orthogonally, the system is structured as a 4-layer signal stack describing the channels through which emissions flow: surface entry (the digital surfaces an agent appears on), signal (the substrate the agent reads), permission (opt-in channels the agent emits through), and intelligence (decision-making routing attention). The Attention Ethics layer operates between the signal and intelligence layers: it is the gate the intelligence layer must pass before any permission-layer emission.

### 5.3 Code-level enforcement

The Attention Ethics layer is enforced through four mutually reinforcing mechanisms:

**(a) Runtime gates.** The notification gate and resonance engine modules implement the seven-gate cascade in code. Each gate failure produces a structured event with the gate name, score, and substrate snapshot.

**(b) CI doctrine test.** A doctrine test runs on every push and scans the entire codebase for banned vocabulary (the voice-lock blocklist), regulated patterns (e.g., service-role-key fallback patterns that would bypass admin gating), and architectural invariants. A regression fails the CI gate and blocks merge.

**(c) Content-safety module.** A content-safety library runs on every emission post-decision. It scrubs personal data per the Singapore Personal Data Protection Act [PDPA 2012, as amended 2020], rejects comparative claims about named businesses (a Defamation Act [Singapore Statutes 2014] consideration), and enforces an aggregate floor of *n* ≥ 5 for any bucket-level claim.

**(d) Database constraints.** The predictions ledger is append-only via a Postgres trigger. The generated content log cascades takedowns via a database trigger that auto-suppresses any row associated with an active takedown record.

### 5.4 Silence correctness implementation

A silence correctness loop runs hourly. For each active pocket, it identifies the prior hour as a candidate window, queries the count of notifications sent within the window and skips if non-zero (it was not a silent hour), aggregates the observation *O*(*P*, *h*) from `pocket_sustainability_ledger`, `watcher_events`, and `offers`, computes the verdict per §4.2, and writes a row to the predictions ledger with `generator = 'silence_correctness'`, `revealed = true`, and `accuracy_score ∈ {0.0, 1.0}`. Inconclusive verdicts are not written.

The rolled-up silence-correctness percentage is exposed at `/api/silence` and rendered on the public `/silence` page. The aggregator filters inconclusive rows out of the denominator per §4.4.

### 5.5 Public observability

Real Signal makes the substrate's behavioural state queryable in three forms: (a) an HTML surface at `real-signal.ai/silence` showing rolling silence correctness; (b) a Model Context Protocol server [Anthropic 2024] at `real-signal.ai/api/mcp` exposing four read-only tools that AI assistants can query directly; (c) a long-form documentation file at `real-signal.ai/llms-full.txt` for retrieval by AI systems. The MCP server, in particular, lets external AI systems verify the substrate's claimed restraint behaviour without trusting our self-report — they can query the predictions ledger directly.

---

## 6. Evaluation

This paper presents a deployed-systems proposal with preliminary evaluation. The primary contribution is architectural and definitional — the Attention Ethics layer (§3), the formal definition of silence correctness (§4), and the production implementation (§5) — rather than empirical. The substrate operates pre-launch at submission time, with limited commercial activity to date; substantive empirical validation requires the production deployment to mature into real merchant and consumer traffic over a timeline of months. The figures and tables below report on the data available at submission and are honestly marked where data has not yet accumulated.

### 6.1 Silence correctness over time

*[Figure 1: silence correctness percentage per pocket over the observation window, with confidence intervals widening for low-N pockets — awaiting accumulation of ≥200 gradable rows per pocket, expected ~30 days of operation.]*

The aggregate trend across pockets is the primary outcome measure. We acknowledge that at the time of writing the substrate is observing one neighbourhood with minimal consumer or merchant traffic. The population sample from which silence correctness is constructed depends primarily on environmental and infrastructure signals; the metric becomes statistically meaningful at *N* ≥ 200 gradable rows, which by the hourly cron cadence requires approximately 30 days of clean operation.

### 6.2 Predictions ledger accuracy

In addition to silence correctness, Real Signal seals forward predictions for pocket-level pulse (quiet / moderate / busy) at 60-, 120-, and 180-minute horizons. The reveal cron scores each forward prediction against the atmosphere reading captured at the reveal time.

*[Figure 2: distribution of accuracy scores per generator type — awaiting reveal-cron data accumulation.]*

These forward predictions are not Attention Ethics layer outputs, but they share the append-only ledger commitment scheme and provide an orthogonal validation of the substrate's predictive integrity.

### 6.3 Aggregate emission vs withholding ratio

*[Table 1: total candidate emissions, total emitted, total withheld, percentage withheld by gate — awaiting deployment data.]*

The ratio of withheld to emitted is itself a useful descriptive statistic: it confirms that the Attention Ethics layer does in fact filter substantially. We expect >90% of candidate emissions to be withheld in pre-launch conditions, dropping as substrate density and user receptivity accumulate.

### 6.4 Per-reason breakdown

*[Figure 3: count of silence events by gate of origin — awaiting deployment data.]*

Which gates fire most often is a substantive finding: it tells us what kind of restraint the layer is doing. Pre-launch, we expect Gate 0a (resonance) and Gate 0b (Moment-level silence) to dominate; in steady state, we expect Gate 3 (user fatigue) and Gate 4 (earned interruption) to become more prevalent.

### 6.5 Baseline comparison

To contextualise the emission volume, we plan to estimate what a conventional engagement-optimising consumer-app substrate would have emitted over the same window, calibrated against published per-merchant per-week notification rates from comparable platforms [Pielot et al. 2014]. The Real Signal emission count is expected to be one to two orders of magnitude lower. We acknowledge direct apples-to-apples comparison is not possible — the systems serve different functional roles — but order-of-magnitude framing is informative.

---

## 7. Discussion

### 7.1 Regulatory implications

The Attention Ethics layer and silence correctness map cleanly onto two emerging regulatory frameworks.

The Singapore IMDA AI Verify framework [IMDA 2023] specifies eleven principles for trustworthy AI, including human-centricity, transparency, and accountability. Silence correctness operationalises human-centricity as a measurable property: a system that scores well is, by construction, prioritising user cognitive load over engagement frequency. Aggregate publication of the metric operationalises transparency. The append-only ledger operationalises accountability: the system cannot retrospectively misrepresent its decisions.

The EU AI Act [European Parliament 2024], particularly Articles 13 and 14 on transparency and human oversight, suggests that systems whose behaviour affects user attention should be auditable. Silence correctness provides an audit handle that does not require code access — the rolled-up percentage and the underlying ledger are public.

### 7.2 HCI implications

To our knowledge, this is an early measurable engineering property in the calm-computing tradition. It enables comparative studies between systems (a study comparing silence correctness across deployed assistants would be straightforward to design). It also enables longitudinal studies within a single system (does silence correctness improve as the substrate matures?).

### 7.3 Open questions

We name four open questions arising from this work.

**Generalisability.** Real Signal operates in one neighbourhood. Does silence correctness as a metric generalise to other domains (recommendations, search, agentic coding)? The metric's structure suggests it should — any system that makes emit-or-not decisions has silent moments whose correctness can be retrospectively evaluated — but the threshold calibration may be domain-specific.

**Cross-cultural calibration.** Restraint thresholds embed cultural assumptions (the doctrinal *r* ≥ 3 for redemption-based miss detection assumes a particular pace of commercial activity). Cross-cultural deployment may require recalibration.

**Long-term trust effects.** We hypothesise that systems with high silence correctness accumulate user trust faster than systems with high emission rates. Verifying this hypothesis empirically requires longitudinal user studies beyond the scope of this paper.

**Metric gaming and strategic adaptation.** A measurable property that becomes publicly visible may invite adaptation. Future work should investigate whether platform operators, the agents they operate, or the merchants and consumers downstream of those agents adapt their behaviour to maximise silence correctness, and whether such adaptation weakens the metric's intended meaning. The mitigations in §4.6 and the audit framing in §7.4 reduce the likelihood of trivial gaming, but the longer-term equilibrium under publication pressure is an empirical question requiring data that does not yet exist.

### 7.4 Gaming risk under public publication

A natural concern is that publishing silence correctness publicly creates an optimisation target that may degrade the metric's meaning over time. The mitigations described in §4.6 (append-only commitment, delayed scoring, environmental observation requirements, threshold transparency) provide first-line resistance. We note two additional considerations for the published-metric regime.

First, the metric is operator-side: a high silence correctness reflects on the operator that publishes it. An operator who games the metric while degrading actual user experience would, over time, lose user trust through the user-facing experience rather than through the metric. The metric is a complement to direct user feedback, not a substitute for it.

Second, in the regulatory context (§7.1), silence correctness becomes a candidate for *external* audit rather than self-report. A regulatory body with substrate access (e.g., IMDA in the AI Verify certification context) could recompute the metric independently from the operator's predictions ledger and substrate tables. We see this as the natural endgame for measurable AI-system properties: self-published metrics are useful for transparency; externally audited metrics become genuinely trustworthy.

### 7.5 Restraint is not anti-engagement

We anticipate readers from commercial contexts asking why an operator would deploy a system optimised for restraint when commerce is the first application. We argue the framing of restraint as anti-commerce is mistaken.

Poorly timed interventions impose measurable costs even from a commercial viewpoint. Attention fatigue reduces the probability of response to any subsequent emission [Mehrotra et al. 2016]. Notification blindness reduces the marginal yield of additional emissions to near zero. Reduced user trust degrades the rate at which the platform's surfaces are returned to. The lower-quality response data from over-emission feeds back into the substrate, further degrading future intervention quality.

The objective of the Attention Ethics layer is therefore not fewer interventions in the abstract, but *higher per-intervention legitimacy*. In the limit, a system that achieves high silence correctness should produce higher commercial yield per emission than a system that does not, because each emission has been retrospectively shown to align with observed user receptivity and substrate state. We do not yet have the empirical data to demonstrate this directly (§6), but the architectural argument is straightforward: a system that earns each interruption should be more, not less, valuable to the entities it serves.

---

## 8. Limitations

We name seven limitations.

1. The substrate is deployed in a single neighbourhood at submission time. Cross-pocket generalisation is supported by the architecture's similarity-index mechanism but unverified empirically.
2. Real merchant and consumer traffic at submission time is minimal; the silence correctness data is constructed from environmental and infrastructure signals alone.
3. Threshold values in the verdict function (§4.2) are chosen doctrinally rather than empirically calibrated.
4. The voice-lock vocabulary is English-only. Multilingual deployments would require parallel banned-vocabulary lists and unicode-normalisation rules per supported language.
5. The system assumes a single-pocket model at any given moment; cross-pocket attention competition is not yet modelled.
6. Silence correctness inherits the gaming-resistance properties of its underlying ledger and substrate tables (§4.6), but is not fully gaming-resistant. A motivated operator could selectively activate pockets at expected low-activity times. We mitigate but do not eliminate this risk; full robustness likely requires external audit (§7.4).
7. Author conflict of interest: the first author founded and operates Real Signal. We have attempted to mitigate this by making all evaluation data publicly inspectable (predictions ledger, silence-correctness rollup), but the conflict is structural and worth naming.

---

## 9. Conclusion and Ethics Statement

### 9.1 Conclusion

We have introduced the Attention Ethics layer, a structured seven-gate cascade that enforces restraint as a runtime property of an AI system. We have defined silence correctness as a measurable, tamper-resistant property of the system's restraint behaviour. We have described a production implementation, *Real Signal*, that enforces the layer through code, CI tests, and database constraints, and exposes the aggregate silence-correctness percentage publicly. We have argued this constitutes, to our knowledge among the first deployed examples, an early working instantiation of restraint as a first-class engineering property.

We claim no specific empirical results beyond architectural and definitional contributions; substantive empirical validation requires the substrate to mature into real merchant and consumer traffic, on a timeline of months to years.

### 9.2 Ethics statement

This work involves no human-subjects research at the time of submission. The substrate is composed entirely from publicly available signals (LTA DataMall, NEA, Google Places, government calendars). All outlet-level data references publicly registered business identities; no individual persons are named or characterised. Personal data scrubbing (NRIC, SG phone numbers, email addresses, unit numbers, postcodes, credit cards, dates of birth) is applied unconditionally at the content-safety layer per [PDPA 2012].

Aggregate-level claims about merchant clusters require a minimum sample size of *n* ≥ 5 before publication, a constraint enforced at runtime. Comparative claims about named businesses are rejected entirely by the validator.

The production code is closed-source at submission time, but production endpoints are publicly readable; the MCP server at `real-signal.ai/api/mcp` allows external systems to inspect the substrate's claimed behaviour directly without trusting our self-report.

### 9.3 Conflict of interest

The first author founded Real Signal and operates it commercially. Real Signal operates under a "free on both sides until measurable value exists" pricing posture at the time of submission and has no commercial revenue. The author has no financial interest in any of the cited regulatory frameworks or competing AI systems.

### 9.4 Funding

This work received no external funding. Infrastructure costs (hosting, database, MCP server) are borne by the operating entity.

---

## References

Anthropic (2024). *Model Context Protocol specification, version 2024-11-05*. modelcontextprotocol.io

Apple (2018). *Screen Time*. Apple WWDC 2018, iOS 12 release.

Bai, Y. et al. (2022). *Constitutional AI: Harmlessness from AI Feedback*. arXiv:2212.08073.

Case, A. (2015). *Calm Technology: Principles and Patterns for Non-Intrusive Design*. O'Reilly Media.

Center for Humane Technology (2022). *Ledger of Harms and policy reports*. humanetech.com

Christian, B. (2020). *The Alignment Problem: Machine Learning and Human Values*. W. W. Norton.

European Parliament (2024). *EU Artificial Intelligence Act*, particularly Articles 13–14.

Eyal, N. (2019). *Indistractable: How to Control Your Attention and Choose Your Life*. BenBella Books.

Geifman, Y. & El-Yaniv, R. (2017). *Selective Classification for Deep Neural Networks*. NeurIPS 2017.

Google (2018). *Digital Wellbeing*. Google I/O 2018 announcement.

Google Maps Platform (2024). *Places API documentation*. developers.google.com/maps/documentation/places

Hari, J. (2022). *Stolen Focus: Why You Can't Pay Attention*. Crown.

Harris, T. (2018). *Time Well Spent / Humane Technology principles*. Center for Humane Technology.

Hiniker, A., Hong, S. R., Kohno, T. & Kientz, J. A. (2016). *MyTime: Designing and Evaluating an Intervention for Smartphone Non-Use*. CHI 2016.

IMDA (2023). *AI Verify Framework*. Infocomm Media Development Authority Singapore. aiverifyfoundation.sg

LTA (2024). *DataMall API documentation*. Land Transport Authority Singapore. mytransport.sg/content/mytransport/home/dataMall.html

Lyngs, U., Lukoff, K., Slovak, P., Binns, R., Slack, A., Inzlicht, M., Van Kleek, M. & Shadbolt, N. (2019). *Self-Control in Cyberspace: Applying Dual Systems Theory to a Review of Digital Self-Control Tools*. CHI 2019.

Madras, D., Pitassi, T. & Zemel, R. (2018). *Predict Responsibly: Improving Fairness and Accuracy by Learning to Defer*. NeurIPS 2018.

Mark, G., Gudith, D. & Klocke, U. (2008). *The Cost of Interrupted Work: More Speed and Stress*. CHI 2008.

Mehrotra, A., Pejovic, V., Vermeulen, J., Hendley, R. & Musolesi, M. (2016). *My Phone and Me: Understanding People's Receptivity to Mobile Notifications*. CHI 2016.

Monge Roffarello, A. & De Russis, L. (2019). *The Race Towards Digital Wellbeing: Issues and Opportunities*. CHI 2019.

NEA (2024). *PSI and Weather Real-Time APIs*. National Environment Agency Singapore. data.gov.sg/datasets/?agencies=National%20Environment%20Agency

Okoshi, T., Ramos, J., Nozaki, H., Nakazawa, J., Dey, A. K. & Tokuda, H. (2015). *Attelia: Reducing User's Cognitive Load due to Interruptive Notifications on Smart Phones*. PerCom 2015.

PDPA (2012). *Personal Data Protection Act*. Republic of Singapore, as amended 2020.

Pielot, M., Church, K. & de Oliveira, R. (2014). *An In-Situ Study of Mobile Phone Notifications*. MobileHCI 2014.

Rogers, Y. (2006). *Moving on from Weiser's Vision of Calm Computing: Engaging UbiComp Experiences*. UbiComp 2006, LNCS 4206, 404–421.

Russell, S. (2019). *Human Compatible: AI and the Problem of Control*. Viking.

Singapore Statutes (2014). *Defamation Act (Chapter 75)*. Republic of Singapore.

Weiser, M. (1991). *The Computer for the 21st Century*. Scientific American 265(3), 94–104.

Weiser, M. & Brown, J. S. (1996). *The Coming Age of Calm Technology*. In *Beyond Calculation: The Next Fifty Years of Computing*. Copernicus / Springer-Verlag.

---

**More citation formats:** [BibTeX, APA, MLA, plain text](https://real-signal.ai/research/cite.md)

*Working draft. Comments welcome at* `hello@real-signal.ai`. *Public production implementation:* https://real-signal.ai *(see* `/silence` *for the live silence-correctness dashboard,* `/api` *for the endpoint catalogue, and* `/api/mcp` *for the read-only MCP server).*