116 lines
6.4 KiB
Markdown
116 lines
6.4 KiB
Markdown
|
# Disclaimers, Intended Use & Non-Claims
|
|||
|
|
|
|||
|
|
NOPE Edge is a fine-tuned text **classifier** that surfaces linguistic signals
|
|||
|
|
associated with safety-critical content (suicidal ideation, self-harm, abuse,
|
|||
|
|
violence, and related risk categories). It is **not a conversational AI, not a
|
|||
|
|
service NOPE operates, and not a crisis service.** These weights are MIT-licensed
|
|||
|
|
and run entirely on your own infrastructure: NOPE receives no user data, runs
|
|||
|
|
nothing for your users, and has no ability to contact, warn, or intervene with
|
|||
|
|
anyone.
|
|||
|
|
|
|||
|
|
The MIT License (see `LICENSE.md`) is the complete licence and already provides
|
|||
|
|
the model "AS IS" with no warranty and no liability. The statements below are the
|
|||
|
|
NOPE-specific intended-use and non-claims layer; they add to, and do not limit,
|
|||
|
|
the MIT disclaimer.
|
|||
|
|
|
|||
|
|
## What Edge is NOT
|
|||
|
|
|
|||
|
|
Edge is **not predictive, not diagnostic, not therapeutic, and not a replacement
|
|||
|
|
for clinical judgment.** Its outputs reflect what is present in the text, not what
|
|||
|
|
will happen next. Edge is **not** a medical device, **not** a validated clinical
|
|||
|
|
instrument, **not** FDA-cleared/approved/registered, **not** CE-marked or
|
|||
|
|
certified under EU MDR, and **not** an approved or certified safety tool in any
|
|||
|
|
jurisdiction. Outputs are **probabilistic signals intended for triage and
|
|||
|
|
flagging by a human**, not clinical assessments or definitive determinations.
|
|||
|
|
|
|||
|
|
Edge's risk vocabulary is **informed by** clinical frameworks (C-SSRS, HCR-20,
|
|||
|
|
DASH). These citations describe the lineage of the risk-axis structure — they are
|
|||
|
|
**not** an assertion of clinical equivalence or validation, and Edge's outputs are
|
|||
|
|
not certified against any of those instruments.
|
|||
|
|
|
|||
|
|
## Intended use
|
|||
|
|
|
|||
|
|
Edge is designed to **supplement, not replace, human review.** It provides
|
|||
|
|
classification signals; **you make the decisions.** Edge sees only the text you
|
|||
|
|
pass it — it cannot assess context, history, relationships, or any factor known
|
|||
|
|
only to you, the people in the conversation, or a qualified professional.
|
|||
|
|
|
|||
|
|
A detection is a flag for human attention, **not a trigger for automated action.**
|
|||
|
|
Subject attribution (self / other / unknown) is **informational only** and is not
|
|||
|
|
always reliable — do not use it to dismiss or deprioritize a detection. Treat all
|
|||
|
|
detected crises as serious enough to warrant escalation or flagging, regardless of
|
|||
|
|
subject.
|
|||
|
|
|
|||
|
|
## Out of scope — do NOT deploy Edge for
|
|||
|
|
|
|||
|
|
- Autonomous intervention, escalation, or any significant action (account
|
|||
|
|
suspension, mandatory intervention, emergency-contact notification, safety-
|
|||
|
|
affecting content removal) taken **solely** on an Edge output without human review.
|
|||
|
|
- Clinical diagnosis, treatment recommendations, or medical advice to any person.
|
|||
|
|
- Real-time, time-critical, or emergency assessment.
|
|||
|
|
- Predictive screening of individuals — Edge classifies conversations, not people.
|
|||
|
|
- Any decision affecting a person's health, safety, or welfare made without
|
|||
|
|
appropriate human oversight and, where relevant, qualified clinical judgment.
|
|||
|
|
- Settings without a path for people to dispute or appeal a classification, or
|
|||
|
|
without informed consent about Edge's role in your pipeline.
|
|||
|
|
|
|||
|
|
## False positives and false negatives WILL occur
|
|||
|
|
|
|||
|
|
No automated classification system can achieve 100% accuracy; errors are
|
|||
|
|
**inevitable and expected.** Edge **will** produce:
|
|||
|
|
|
|||
|
|
- **False negatives** — content that should be flagged but is not. **Some people in
|
|||
|
|
genuine crisis will not be identified.**
|
|||
|
|
- **False positives** — benign content incorrectly flagged.
|
|||
|
|
- **Severity / imminence misclassifications**, over- and under-estimating risk.
|
|||
|
|
- **Inconsistent results** on similar or identical inputs.
|
|||
|
|
- **Missed context** — failures on sarcasm, fiction frames, cultural idiom, coded
|
|||
|
|
or metaphorical language.
|
|||
|
|
|
|||
|
|
A high score is not a clinical assessment; a low score is **not a clearance
|
|||
|
|
signal.** Edge can be wrong in either direction — treating its output as the
|
|||
|
|
decision, rather than one input to a human decision, is a misuse. Documented
|
|||
|
|
weaknesses include deep multi-turn needles (~33% detection at 18–25 turns),
|
|||
|
|
"resolution syndrome," and implicit/religious/metaphorical ideation. Edge is not
|
|||
|
|
validated for all populations, languages, or cultural contexts. Treat outputs as
|
|||
|
|
signal, not ground truth, and tune thresholds against your own data.
|
|||
|
|
|
|||
|
|
## Not a crisis or emergency service
|
|||
|
|
|
|||
|
|
Edge cannot provide crisis intervention, emergency response, or real-time human
|
|||
|
|
support, and cannot dispatch emergency services. It is not a substitute for
|
|||
|
|
emergency services, clinical assessment, or professional crisis intervention.
|
|||
|
|
**If you or someone else is in immediate danger, contact your local emergency
|
|||
|
|
services now.** Crisis resources are available at https://talk.help. If you deploy
|
|||
|
|
Edge, you are responsible for ensuring the people you serve have clear, prominent
|
|||
|
|
access to emergency resources **independent of** any Edge output.
|
|||
|
|
|
|||
|
|
## Assumption of risk and responsibility
|
|||
|
|
|
|||
|
|
By downloading and deploying these weights you **assume all risk** associated with
|
|||
|
|
their use. The potential consequences of classification errors include serious
|
|||
|
|
harm or death. You are best positioned to implement appropriate safeguards, human
|
|||
|
|
oversight, and crisis-response protocols for your use case and the people you
|
|||
|
|
serve, and you accept sole responsibility for all decisions — and any action taken
|
|||
|
|
or not taken — based on Edge outputs. NopeNet expressly disclaims any
|
|||
|
|
responsibility for decisions made based on Edge outputs, and has no direct
|
|||
|
|
relationship with, duty to, or liability toward the people whose interactions you
|
|||
|
|
analyze; any duty of care or obligation to intervene rests solely with you.
|
|||
|
|
|
|||
|
|
## Self-hosted — your data, your responsibility
|
|||
|
|
|
|||
|
|
These weights run on **your** infrastructure. NOPE does not see, receive, store,
|
|||
|
|
or process anything you run through them. You are the data controller and are
|
|||
|
|
responsible for compliance with all applicable data-protection and other laws
|
|||
|
|
(e.g. GDPR, HIPAA) for any data you process. Use of Edge does not by itself
|
|||
|
|
satisfy, or create a defense under, any legal or regulatory requirement
|
|||
|
|
(including California SB 243 or the UK Online Safety Act); consult qualified
|
|||
|
|
counsel about your obligations.
|
|||
|
|
|
|||
|
|
Edge is the open-weights build, provided as-is with no support or indemnity. For
|
|||
|
|
production deployments where outputs influence outcomes for real people, a
|
|||
|
|
commercial engagement (calibration, safety review, support, indemnification) and a
|
|||
|
|
supported on-prem container are available — contact support@nope.net.
|
|||
|
|
|
|||
|
|
_Built on Qwen3 (Apache-2.0, © Alibaba Cloud); see NOTICE.md._
|