1. CapsuleAI: Operational Scope

CapsuleAI is the umbrella name for three connected AI subsystems: (i) the Predictive Risk Engine, which correlates IoT sensor telemetry with historical fungal-growth data; (ii) the 3D Hygrothermal Simulation Engine, which forecasts condensation and colonisation risk on a per-room model; and (iii) the Vision-Based Species Classifier, which scores mold-genus likelihood from inspection imagery.

In addition, CapsuLabs operates a public-site Assistant Chatbot. As detailed in §3A below, the current Assistant is a rule-based widget that runs entirely in your browser; it is informational only and is not part of the diagnostic CapsuleAI pipeline.

2. What CapsuleAI Does NOT Do (Strict Waiver)

CapsuleAI and the Assistant Chatbot DO NOT:

Perform physical diagnoses or site inspections.
Diagnose medical toxicology (e.g., mycotoxin illness) or recommend medical treatment.
Issue habitability, structural-engineering, or insurance verdicts.
Replace certified human Industrial Hygienists or licensed labs.
Detect mold or moisture in cavities or zones outside sensor / image coverage.
Provide legal, real-estate, or financial advice.

3. Probabilistic Limitations & Stated Accuracy

All CapsuleAI outputs (risk scores, simulation maps, species classifications) are probabilistic estimates, not guarantees. Outputs are highly susceptible to false positives from anomalies (e.g., shower steam flagged as moisture event) and to false negatives where the input is incomplete (e.g., mold behind a wall outside sensor coverage).

Any species-classification accuracy figure published in marketing (such as a percentage match against certified lab reference data) reflects controlled internal validation. Real-world accuracy varies with image quality, lighting, species rarity, and sample contamination. Critical decisions must be confirmed by certified-lab analysis. CapsuLabs makes no representation about simulation-forecast accuracy beyond seven (7) days from the simulation timestamp.

3A. Assistant Chatbot: Current Implementation

The public-site Assistant Chatbot is currently a rule-based widget running entirely in your browser. It matches your message against a small set of keyword triggers (booking, pricing, lab tests, contact information, service area) and returns one of a handful of pre-written responses with quick-link suggestions. It does not call any third-party language-model provider.

No large-language-model (LLM) is invoked. There is no server-side AI processing of your messages.
Chat history is stored in your browser's local storage and never transmitted to CapsuLabs servers.
Chatbot conversations are not used to train any AI model, by CapsuLabs or by any third party, because no AI provider receives the messages in the first place.
The chatbot is informational only. Do not rely on chatbot output as a basis for medical, legal, financial, or property-safety decisions.
Future upgrade: we may at our option replace this rule-based widget with an LLM-powered assistant (e.g., Anthropic Claude or OpenAI GPT) in a future release. If we do, we will update this policy and the Privacy Policy before any conversation is sent to a third-party provider.

3B. Vision-Based Species Classifier: Specific Limits

The classifier scores likelihood that an inspection photo contains a particular mold genus. It does not produce a definitive identification and it does not distinguish between toxigenic and non-toxigenic strains within the same species (e.g., toxigenic vs non-toxigenic Aspergillus flavus). Real-world accuracy varies with photo lighting, focus, sample isolation, surface contamination, species rarity in our training set, and operator photography skill. Any confidence score below 80% must be confirmed by certified-lab analysis before any remediation, medical, legal, or real-estate decision is taken.

3C. 3D Hygrothermal Simulation Engine: Specific Limits

The 3D simulation engine forecasts elevated condensation and colonisation risk on a per-room basis using the geometry, materials, ventilation, and HVAC data the operator inputs. The engine cannot model: (i) hidden defects (water leaks behind walls, undocumented HVAC zones, sub-floor voids), (ii) occupant behaviour shifts (cooking, showering frequency, window-opening habits), (iii) future weather beyond a seven (7) day horizon, or (iv) materials and assemblies absent from our training corpus. Simulation outputs are mathematical estimates of elevated risk probability, not guarantees that mold will or will not appear.

3D. IoT Sensor Calibration & Coverage Limits

Sensors are factory-calibrated and re-calibrated annually. Between calibrations, readings may exhibit ±5% drift on humidity, ±0.5°C on temperature, and varying drift on particulate / VOC channels. Sensor placement determines coverage; mold or moisture in unsensored zones cannot be detected, including but not limited to wall cavities, sub-floors, ducting, attics, crawlspaces, and rooms that were not in the deployment plan. CapsuLabs makes no representation regarding mold/moisture conditions in unsensored areas.

4. Human-In-The-Loop (HITL)

High-severity algorithmic alerts require manual, human-in-the-loop validation by a certified CapsuLabs operator before being escalated to definitive "Action Required" status. We do not enforce automated physical interventions.

4B. Right to Human Review (GDPR Art. 22)

Where a CapsuleAI output materially affects you (e.g., influences an insurance premium decision, a lease determination, or a property-purchase deal), you have the right to (i) request meaningful information about the logic involved, (ii) request human review of the decision, and (iii) contest the decision. To exercise this right, email privacy@capsulabs.co.il within thirty (30) days of receiving the output. CapsuLabs will conduct a manual review by a certified operator and respond in writing within ten (10) business days.

5. Anonymized AI Training Data

The exact coordinates, tenant identities, and financial profiles of the property are stripped at the edge layer. CapsuleAI trains ONLY on isolated variables (temperature, humidity delta, particulate counts).

5B. Training Data Sources & Bias Disclosure

CapsuleAI is trained on: (i) anonymised CapsuLabs sensor telemetry from opted-in customers; (ii) publicly available mycology and building-science datasets (e.g., academic papers, government building-science research); and (iii) synthetic data generated by the 3D simulation engine. We do not train on third-party customer photos, medical records, or unrelated personally-identifying information.

Geographic and typology bias: Training data is currently concentrated in Israel and select EU markets. Performance may degrade in significantly different climates (e.g., high-humidity tropical, arid desert) or non-residential building types we have not yet sampled (e.g., heavy industrial, agricultural greenhouses). Each Report discloses any known geographic or typology limitation that materially affects the result.

5C. EU AI Act Classification

CapsuLabs classifies CapsuleAI as a non-high-risk AI system under Regulation (EU) 2024/1689 (the AI Act), Annex III. It does not perform safety-critical functions, does not score creditworthiness, does not determine access to essential services, and does not make legally binding determinations affecting fundamental rights. CapsuLabs maintains an internal risk-management documentation file aligned with the AI Act's transparency obligations and makes a redacted summary available to enterprise B2B clients on request.

6. Cap on Algorithmic Liability

As strictly outlined in Section 10 of our Master Terms of Service, liability for any claims arising directly or indirectly from the outputs, failures, or mathematical models of CapsuleAI is absolutely and unequivocally capped at the fees paid in the preceding 12 months. AI predictions are never guaranteed.

7. Governing Language

This Statement is published in English and Hebrew. For Clients ordinarily resident in Israel, the Hebrew version controls; for all other Clients, the version corresponding to the language in which the Client accepted these terms controls.

AI Transparency Policy