DARS Progress Report: Methodology, Metrics & Limitations

    Last updated: November 4, 2026

    Digital Accountability & Recovery System (DARS) — Simpatica.AI

    A companion document to the DARS Progress Report. It describes how the report is produced, what each measure actually counts, what the underlying data can and cannot see, and the boundaries within which the report should be interpreted. Written for clinicians and researchers evaluating the instrument.

    Scope note: this document describes the DARS Progress Report as currently built. A separate "DARS Therapeutic Methodology" paper describes the broader clinical vision (assessment-driven risk profiles, staged Block/Filter/No-Filter recovery plans, coach supervision); features in that paper that are not yet implemented are noted as roadmap here, not described as current behavior.

    1. Purpose and intended use

    The DARS Progress Report turns objective, device-level DNS activity into a structured summary — a "Digital Mirror" — that a therapist and client can review together when working on behavioral-addiction recovery (for example, compulsive use of adult content or high-risk social platforms).

    It is a conversation aid and a source of objective behavioral signal to complement self-report — not a diagnostic instrument, not a validated risk score, and not a standalone measure of recovery. It is meant to be read alongside clinical judgment, the client's own account, and the therapeutic relationship. It does not produce a diagnosis or a treatment recommendation that should be acted on without a clinician's interpretation.

    The report does display a coarse Executive Baseline status (green / yellow / red) and "key factors." This is a non-validated heuristic built from fixed internal thresholds on the measures below, provided as a quick orientation for the reviewing clinician — not a clinical severity rating. Its thresholds have not been validated against any outcome.

    The report's value and its limits come from one fact: it sees network-level behavior with precision and sees nothing of intent, emotion, or content directly. Everything below follows from that.

    2. Where the data comes from, and what is collected

    DARS derives its measures from DNS activity logs provided through Control D, the DNS filtering service that enforces the client's content boundaries across their enrolled devices. When a device looks up a domain, Control D records the lookup and whether its filters allowed, blocked, or redirected it. DARS ingests these logs on a schedule (roughly every 15 minutes).

    What is actually stored. Each log record contains the full queried hostname (e.g. example.com), the timestamp, the device, the filter outcome and which filter triggered it, the DNS record type, and network-identity fields: the device's source IP address, ISP/ASN, and approximate city/country, plus the resolved answer addresses. DARS then adds its own derived category label to each domain (see §3).

    What is not collected. DARS does not capture page content, the path or query string of a URL, keystrokes, screenshots, messages, or search terms. A DNS log shows that a hostname was looked up, not what was done there. (We note this precisely because "DNS only" is sometimes read as "nothing identifying" — the hostname plus source IP and city is itself identifying, and is stored.)

    How categories are assigned (important). The Inner / Middle / Outer labels are not supplied by Control D. DARS assigns a single recovery-zone label to each domain using a generative-AI classifier (Google Gemini), globally across all clients, and caches it. These labels are not individually reviewed by a clinician, are not tailored per client, and a domain keeps its label until it is re-classified. Category-mapping error therefore propagates into every distribution and percentage in the report.

    3. The measures, and how to read them

    These are Simpatica-defined behavioral proxies built for this product. They are not validated psychometric instruments and do not map to any established clinical scale. Each is defined below as the system actually computes it, with its specific weakness.

    Unit of analysis: the session

    Most measures are computed over sessions, not individual requests. DARS groups a device's logs into a session whenever the gap between lookups is under ~2 minutes, and labels the whole session by a single "primary" domain within it. A session is thus a coarse proxy for a block of attention, not a measure of time spent; many sessions are a single instant, and background or cached lookups may not appear at all.

    Three Circles distribution (Inner / Middle / Outer-Work / Outer-Rest)

    The share of sessions whose primary domain is labelled off-limits (Inner), risky-but-permitted such as dating/social (Middle), or productive/restful (Outer "Green" zones). Framework adapted from the recovery model used by Sex Addicts Anonymous; DARS applies it to AI-assigned domain categories rather than to behaviors the individual defines. Percentages are rounded to whole numbers, so a genuine Inner share below ~0.5% displays as 0%. Because a blocked off-limits lookup rarely becomes a session's primary domain, Inner-Circle exposure is better read from blocked events than from this pie — a 0% Inner slice does not by itself mean zero off-limits attempts.

    Probing Velocity

    The per-day count of blocked DNS lookups (deduplicated) over the evaluation period — i.e. how often the client's own filters blocked a request, across all filter categories, divided by the number of days in the window. The report also shows the immediately preceding equal-length window and a pooled multi-window baseline for trend context.

    A rising Probing Velocity is described as potentially consistent with increased seeking or craving pressure, but this is an inference the data cannot confirm. A blocked lookup is not a confirmed intentional attempt: it may be a background request from an app, an ad network, or an embedded tracker, and in practice a large share of blocks come from a single broad filter (e.g. social). Read the trend as a prompt for a clinician to explore, not as a measurement of intent.

    Context Transitions

    The count of times, within the window, that a client moved from a work/productive session into a risk-category (Middle) session on the same device within a short interval (roughly 2–5 minutes). It is offered as a hypothesis-generating signal for stress- or fatigue-triggered seeking. It is a timing coincidence between two categorized sessions; attributing it to cognitive fatigue is a discussion point, not a finding. A companion cross-device signal flags a riskier category appearing on a second device shortly after activity on another.

    Executive Baseline status and "key factors"

    A green/yellow/red label derived from fixed thresholds on the above (presence of Inner-Circle activity, probing-velocity change, context transitions, streak direction). The thresholds are invented defaults, not validated cutoffs. Treat the status as a quick visual orientation, not a clinical rating.

    Streak

    The number of consecutive recent days (among days that have session data) with no session whose primary domain is labelled Inner-Circle. Days with no activity neither extend nor break it, and it is bounded by how far back the recent session data reaches.

    A note on the numbers. Rates like "events per day" are only meaningful against a reference. Where possible the report compares a metric to the same client's own prior periods; there is no population norm or clinical threshold for these proxies. Short windows can swing percentages wildly from small absolute changes, and a few high-activity days can dominate an average. Read every value with the window length and the daily distribution in view.

    4. What the data cannot see (blind spots)

    Because the signal is DNS-level and device-scoped, whole categories of activity are invisible:

    • Unmonitored devices and networks. Activity on a device that is not enrolled, or on a network path outside the configured DNS, is not captured.
    • Encrypted DNS, VPNs, alternative resolvers. A device set to bypass the filtering DNS (DNS-over-HTTPS to another provider, a VPN, a changed resolver) can circumvent capture. Absence of activity in the data is not proof of absence of activity. DARS does not currently detect such evasion.
    • Content and context. DARS sees that a hostname was requested, not what was viewed, typed, searched, or intended, nor the client's emotional state.
    • Category-mapping error. Labels are AI-assigned, global, and unreviewed; a domain may be mis- or un-categorized, and distributions inherit that error.
    • Shared devices. If a monitored device is used by more than one person, activity is not attributed to individuals.
    • Timing. Different panels bucket time slightly differently; time-of-day readings for a client outside the configured timezone can be offset.

    The practical consequence: DARS is strong evidence of what it observed and weak evidence about what it did not. A clean report is consistent with genuine stability and with undetected activity through an unmonitored path. Clinicians should weight it accordingly.

    5. Correlation, not causation

    The report uses hedged, observational language — "may suggest," "is consistent with," "warrants monitoring" — and asks readers to hold the same posture. The measures are correlational behavioral signals. None establishes that a pattern was caused by craving, fatigue, or stress. Those are clinical hypotheses the data can help a therapist raise and investigate, not conclusions the data supports on its own.

    6. Validation status and research collaboration

    We are candid about where these measures stand. They have not been validated against established clinical measures, and we make no claim that they predict relapse, recovery, or any clinical outcome. They are behavioral proxies we believe are clinically meaningful and worth studying — which is different from claiming they are proven. The category labels are AI-generated and unreviewed; the status thresholds are unvalidated defaults.

    This is where we are seeking research partnership. We would welcome collaboration with addiction-research faculty to:

    • validate or refute these proxies against established self-report and clinical measures;
    • characterize their stability and sensitivity to real behavioral change;
    • test whether any carry predictive signal for clinically meaningful outcomes;
    • assess the AI category-assignment for accuracy and bias; and
    • refine or retire constructs that do not hold up.

    We would rather present honest proxies and invite scrutiny than overstate what we have. Findings that contradict our framing are useful to us.

    7. Privacy, consent, and data handling

    • Consent. Device monitoring is undertaken with the client's consent as part of enrolling in the DARS program; it is a chosen component of recovery, not covert surveillance. Today consent is handled through program enrolment; a dedicated, signed therapist–client monitoring-consent record captured in the application is on the roadmap (§8) and not yet a stored artifact.
    • What is collected. DNS lookups with hostnames, timestamps, device identity, filter outcomes, and network-identity fields (source IP, ISP/ASN, approximate city/country) and resolved addresses; plus DARS-derived category labels. Not content, URL paths/queries, keystrokes, search terms, or screens (§2).
    • AI processing. The report's narrative paragraphs are generated by a large language model (Google Gemini via Vertex AI). Aggregated metrics and dates are sent for all tiers; the opening orientation also includes the client's profile objective and device types. Domain names are sent to a second model for category classification. These are third-party processing steps; teams with heightened confidentiality requirements should factor that in.
    • Access. A client's data is visible to the client and to the coaches/therapists and administrators within their organization, through authenticated application routes.
    • Privacy controls. The privacy policy and data/monitoring controls are documented at simpatica.ai/privacy.
    • Retention / deletion. Logs are retained without a fixed expiry today; deletion is available to administrators. Control D's own retention is configured in the Control D account.

    8. Known limitations summary

    Limitations, in brief:

    1. Measures are proprietary behavioral proxies, not validated clinical instruments; the status thresholds are unvalidated defaults.
    2. Signal is correlational; it does not establish intent or causation.
    3. Significant blind spots (unmonitored devices, VPN/encrypted-DNS bypass, no content/search, no evasion detection): a clean report is not proof of abstinence.
    4. Category labels are AI-assigned, global, and unreviewed, and propagate into distributions.
    5. The unit of analysis is a coarse "session" proxy; counts are sessions, not requests or time.
    6. Metrics lack population norms; read relative to the client's own baseline and window length.
    7. Short evaluation windows can overstate volatility.

    This document accompanies the DARS Progress Report and is versioned with the product. For questions about methodology or research collaboration, contact the Simpatica team.