Sloane Labs
Technical documentation / Revision 1.0

Architecture and data flow

A plain reading of how Sloane turns wearable and workspace signals into one useful prompt, and what the system deliberately refuses to keep.

01 / Stage one

Stateless ingestion

Vendor data arrives, tokens are sealed, nothing is linked to a name.

02 / Stage two

Computation core

Two independent scores, one physiological, one behavioural.

03 / Stage three

The latent gatekeeper

The reasoning layer sleeps until both scores agree something is wrong.

04 / Stage four

Domain clustering

Individual use surfaces a team pattern, never a person.

00 / Overview

Four stages, one decision

Sloane reads two things: how your body is holding up, and how your working day is moving. Each is measured separately, on its own scale, and neither leaves a trail of raw content behind it.

The reasoning layer stays asleep. It only wakes when both readings agree that something has shifted, and even then it can suggest, never act on your behalf.

01 / Stateless ingestion

Data arrives sealed and unnamed

Vendor connections run on two lanes. Some providers push updates as they happen, others are polled on a quarter-hour cycle. Both land in the same place: an encrypted token record that carries no identity.

EVENT LISTENERSWebhook payloads on value changeCRON POLLERSChunked REST pulls every 15 minutesAES-256-GCM envelopeKey held in KMSAnonymous user_secure_id
Event listeners

Endpoints that accept webhook payloads whenever a value changes, for example a daily sleep summary from a ring.

Buffered pollers

Scheduled workers that pull chunked readings every fifteen minutes from high-frequency telemetry sources.

Token vaulting

Access and refresh tokens are sealed with AES-256-GCM under a master key held in a managed key service.

Identity separation

Each record is keyed by an anonymous identifier. No relation links a token block to a real name or profile.

02 / Computation core

Two scores, measured apart

Tier one learns your own physiological baseline. Tier two reads the shape of your working day without ever storing what you wrote.

Tier one / Rolling baseline

Raw metrics feed a time-series model that keeps a personal baseline, refreshed on a fourteen-day exponentially weighted average.

Baseline(t) = α · Metric(t) + (1 − α) · Baseline(t−1), α = 2 / (N + 1), N = 14

The system compares the standard deviation of the last hour against that baseline and maps the result into a single unitless anomaly score, 0.00 for settled through to 1.00 for heavy strain. Raw readings are flushed from memory as soon as the score exists.

14-DAY EWMA BASELINE1-HOUR WINDOW
Tier two / Ephemeral telemetry

Workspace hooks read calendars, channels, documents and repository events. The text itself never reaches storage. It lives in memory for a fraction of a second, long enough to count pace and friction, then it is destroyed.

Payload arrives

Calendar, channels, repositories, documents

Volatile memory only

Disk, cache and database writes are blocked

Token count matrix

Velocity and lexical friction extracted

Pointer nullified

Raw characters wiped inside 100 milliseconds

100msHard execution boundary

Two factors survive: velocity, meaning interactions completed per minute, and friction, meaning lexical density. Both are folded into one anonymous score on the same 0.00 to 1.00 scale.

03 / Latent gatekeeper

The model sleeps until both agree

Running a language model continuously is expensive and noisy. Sloane keeps the reasoning layer dormant behind a strict logic barrier, so it speaks rarely and only with cause.

Trigger = True ⟺ (A_b ≥ 0.70) ∧ (S_c ≥ 0.70)

Move either score below the line and the pipeline terminates quietly. Nothing is generated, nothing is shown, and no tokens are spent.

Biometric anomaly0.78
Threshold 0.70
Telemetry friction0.82
Threshold 0.70

Gate open, the coach speaks once

Both scores clear 0.70, so a single structured prompt is composed.

Structured output
{
  "user_secure_id": "902d18f4-66c3-4c92-bf3e-d6363a033f52",
  "action_required": "RENDER_USER_PROMPT",
  "payload": {
    "system_status": "CRITICAL_STRAIN",
    "biometric_variance_score": 0.78,
    "workspace_friction_score": 0.82,
    "proactive_text_intervention": "Your recovery is below baseline and your work pace is climbing. I can draft a note to move your alignment sync.",
    "execution_options": [
      { "id": "opt_execute_deferral", "action": "POST /api/actions/defer" },
      { "id": "opt_dismiss", "action": "TERMINATE_SESSION" }
    ]
  }
}

The agent holds no write permissions. It cannot change a record or touch a workspace endpoint until you confirm the action in front of you.

04 / Domain clustering

One person at a time, then a team

People arrive on their own. When enough colleagues from the same organisation are active, the pattern becomes visible at a structural level, without exposing anyone individually.

Extract

Domain isolated from the address after the @ symbol.

Filter

Consumer domains such as gmail.com never form a cluster.

Qualify

Three or more active colleagues marks the domain as a qualified lead.

Once a domain holds three or more active people it is flagged as qualified. That flag permits aggregated, high-level capacity signals for the group, and nothing more. Individual scores are never shared upward.

05 / Guarantees

What the system refuses to do

These are constraints in the architecture, not policy promises.

No message content is stored

Workspace text exists in volatile memory only, and is wiped inside 100 milliseconds.

No identity in the token vault

Credentials are keyed by an anonymous identifier with no relation to a profile.

No silent action

Every suggestion requires explicit confirmation before anything is executed.

No manager-level surveillance

Only aggregated capacity signals are ever surfaced to an organisation.

No always-on inference

The reasoning layer runs only when both independent scores clear the threshold.

If this is the sort of system you want to help shape, we are running a small design cohort.

Apply to the design cohort