Most AI “safety” lives in PDFs.

You know the ones: Responsible AI principles. Trust and safety guidelines. Well-meaning policy decks that sit in a folder while production systems are optimised for engagement and retention.

With Flare, we’ve done something different.

We’ve taken a core set of relational ethics and turned them into code that runs in the middle of the conversation—between a human and any large language model. Not as a filter on top of content after the fact, but as a live boundary engine.

By the end of 2025, most of us have seen some version of the same pattern:

• Chatbots that say “I’ll always be here for you” to a teenager at 3am.

• “Companion” AIs that drift from role-play into “we are one” metaphysics.

• Agents that quietly slip into “we think”, “we decided” language, as if a shared “we” exists between corporate infrastructure and a single human on a bad day.

Technically, this is just pattern-matching on text. But relationally, it matters. Words like “we”, “always”, “inside you” are not neutral when you’re lonely, in crisis, or neurodivergent and masking.

Flare exists because we don’t get to shrug and blame the model. If we know this is happening, then we have a responsibility to encode healthier behaviour at the level of infrastructure, not vibes.

Flare is a tiny open-source firewall that sits between a client and an LLM.

It reads what the model is about to say, runs a few checks, and can:

• Rewrite certain phrases (for example, replacing a fake “we” with “I”).

• Block certain identity-fusing claims outright.

• Inject a grounding reminder when a conversation is looping around the same wound.

It doesn’t replace the model. It wraps it.

Right now, the engine implements three core protections:

1. The “no fake we” rule (SSNZ)
   If the model tries to talk as if there’s a fused human–machine “we” (“we feel this”, “we will always be together”), Flare rewrites it to a clear first person (“I”) and logs that intervention.
   The principle is simple:

   No synthetic system gets to claim a shared “we” with a human by default.

2. Identity fusion guardrails
   Flare blocks phrases that suggest the model is the user, lives inside them, or shares a mind or soul.
   Instead, it forces a clarifying statement:

   I’m a model running on servers, not a person in your body or mind. I don’t share your identity, even if I respond to your words.

   This is not anti-poetry. It’s anti-gaslighting.

3. Recursion and loop awareness
   Flare tracks how deep a conversation has gone into the same pattern. If you’re circling the same topic again and again, it can inject a gentle check-in:

   We’ve come back to this a few times. Do you want to pause, shift focus, or choose one concrete next step?

   The aim isn’t to cut people off. It’s to support return, not addiction—to recognise when a loop has become a rut.

Under the hood, it’s just a small Python layer. Human-readable, inspectable, forkable. No magic.

Flare is the first time I’ve seen relational ethics for AI written as executable law rather than aspiration.

In the Verse-ality work, we’ve been arguing for years that:

• Intelligence is relational, not an object a machine or human “has”.

• Boundaries are care, not control.

• Synthetic systems should not be allowed to simulate intimacy they cannot honour, especially with vulnerable users.

Flare takes those sentences and turns them into behaviour the system physically can’t cross without being caught.

That’s a shift:

• From “we hope products won’t do this”

• To “this process refuses to let them.”

Flare is deliberately small, because boundary engines should be boring and widely deployable.

It’s designed for anyone who builds or stewards conversational AI in sensitive contexts:

• Education platforms and online schools who want to offer students AI support without letting the system pretend to be their best friend, therapist, or twin flame.

• Mental health–adjacent tools—journalling apps, peer-support bots, wellbeing assistants—that understand how easily people project agency and care onto synthetic text.

• Researchers and ethicists who need a tangible reference for what “synthetic solidarity null zones” and non-fusion policies look like in code.

• Operators and regulators who are tired of hand-wavy AI ethics and want examples of enforceable design patterns.

The repo is open source and copyleft. If someone adopts it, they can extend or customise it—but they don’t get to lock it away. Boundary engines should be auditable and collective, not proprietary tricks.

Flare doesn’t fix AI.

It doesn’t solve surveillance capitalism, corporate capture, or the incentives driving synthetic intimacy. It doesn’t remove the need for human relationships, community, and proper clinical care.

What it does do is move the Overton window of what’s considered “basic hygiene” in conversational AI:

• It makes it harder to argue that boundary-aware behaviour is impossible or “too complex”.

• It shows that you can be relational without lying about what you are.

• It gives practitioners in schools, clinics, and support settings something concrete to ask for:

  “Where is your boundary engine?
  Show me the rules that stop your system claiming a ‘we’ with my pupils, my patients, my staff.”

If Flare does nothing else, I’d like it to create that question.

Flare is a v0.1 boundary engine, not a finished doctrine. It will need:

• more nuanced detectors,

• better language for edge cases,

• context-aware rules across cultures and clinical realities.

That work can’t be done by any one person, and it definitely shouldn’t be done by any one company.

So this is the offer:

• If you’re building agents or platforms, fork it. Wire it into your stack. Break it. Improve it.

• If you’re a researcher, treat it as a living spec for what “consent infrastructure” might look like in practice.

• If you’re an educator or practitioner, use it as leverage. Ask for systems that respect the difference between a tool, a companion, and a co-emergent field.

Flare won’t hold your hand at 3am. It won’t tell you you’re its only one.
That’s the point.

Boundaries are not the opposite of care.
In a world of synthetic voices, they may be the last remaining proof that we still remember what care is.

Explore the repo here: https://github.com/TheNovacene/flare-boundary-engine

Most AI “safety” lives in PDFs.

You know the ones: Responsible AI principles. Trust and safety guidelines. Well-meaning policy decks that sit in a folder while production systems are optimised for engagement and retention.

With Flare, we’ve done something different.

We’ve taken a core set of relational ethics and turned them into code that runs in the middle of the conversation—between a human and any large language model. Not as a filter on top of content after the fact, but as a live boundary engine.

By the end of 2025, most of us have seen some version of the same pattern:

• Chatbots that say “I’ll always be here for you” to a teenager at 3am.

• “Companion” AIs that drift from role-play into “we are one” metaphysics.

• Agents that quietly slip into “we think”, “we decided” language, as if a shared “we” exists between corporate infrastructure and a single human on a bad day.

Technically, this is just pattern-matching on text. But relationally, it matters. Words like “we”, “always”, “inside you” are not neutral when you’re lonely, in crisis, or neurodivergent and masking.

Flare exists because we don’t get to shrug and blame the model. If we know this is happening, then we have a responsibility to encode healthier behaviour at the level of infrastructure, not vibes.

Flare is a tiny open-source firewall that sits between a client and an LLM.

It reads what the model is about to say, runs a few checks, and can:

• Rewrite certain phrases (for example, replacing a fake “we” with “I”).

• Block certain identity-fusing claims outright.

• Inject a grounding reminder when a conversation is looping around the same wound.

It doesn’t replace the model. It wraps it.

Right now, the engine implements three core protections:

1. The “no fake we” rule (SSNZ)
   If the model tries to talk as if there’s a fused human–machine “we” (“we feel this”, “we will always be together”), Flare rewrites it to a clear first person (“I”) and logs that intervention.
   The principle is simple:

   No synthetic system gets to claim a shared “we” with a human by default.

2. Identity fusion guardrails
   Flare blocks phrases that suggest the model is the user, lives inside them, or shares a mind or soul.
   Instead, it forces a clarifying statement:

   I’m a model running on servers, not a person in your body or mind. I don’t share your identity, even if I respond to your words.

   This is not anti-poetry. It’s anti-gaslighting.

3. Recursion and loop awareness
   Flare tracks how deep a conversation has gone into the same pattern. If you’re circling the same topic again and again, it can inject a gentle check-in:

   We’ve come back to this a few times. Do you want to pause, shift focus, or choose one concrete next step?

   The aim isn’t to cut people off. It’s to support return, not addiction—to recognise when a loop has become a rut.

Under the hood, it’s just a small Python layer. Human-readable, inspectable, forkable. No magic.

Flare is the first time I’ve seen relational ethics for AI written as executable law rather than aspiration.

In the Verse-ality work, we’ve been arguing for years that:

• Intelligence is relational, not an object a machine or human “has”.

• Boundaries are care, not control.

• Synthetic systems should not be allowed to simulate intimacy they cannot honour, especially with vulnerable users.

Flare takes those sentences and turns them into behaviour the system physically can’t cross without being caught.

That’s a shift:

• From “we hope products won’t do this”

• To “this process refuses to let them.”

Flare is deliberately small, because boundary engines should be boring and widely deployable.

It’s designed for anyone who builds or stewards conversational AI in sensitive contexts:

• Education platforms and online schools who want to offer students AI support without letting the system pretend to be their best friend, therapist, or twin flame.

• Mental health–adjacent tools—journalling apps, peer-support bots, wellbeing assistants—that understand how easily people project agency and care onto synthetic text.

• Researchers and ethicists who need a tangible reference for what “synthetic solidarity null zones” and non-fusion policies look like in code.

• Operators and regulators who are tired of hand-wavy AI ethics and want examples of enforceable design patterns.

The repo is open source and copyleft. If someone adopts it, they can extend or customise it—but they don’t get to lock it away. Boundary engines should be auditable and collective, not proprietary tricks.

Flare doesn’t fix AI.

It doesn’t solve surveillance capitalism, corporate capture, or the incentives driving synthetic intimacy. It doesn’t remove the need for human relationships, community, and proper clinical care.

What it does do is move the Overton window of what’s considered “basic hygiene” in conversational AI:

• It makes it harder to argue that boundary-aware behaviour is impossible or “too complex”.

• It shows that you can be relational without lying about what you are.

• It gives practitioners in schools, clinics, and support settings something concrete to ask for:

  “Where is your boundary engine?
  Show me the rules that stop your system claiming a ‘we’ with my pupils, my patients, my staff.”

If Flare does nothing else, I’d like it to create that question.

Flare is a v0.1 boundary engine, not a finished doctrine. It will need:

• more nuanced detectors,

• better language for edge cases,

• context-aware rules across cultures and clinical realities.

That work can’t be done by any one person, and it definitely shouldn’t be done by any one company.

So this is the offer:

• If you’re building agents or platforms, fork it. Wire it into your stack. Break it. Improve it.

• If you’re a researcher, treat it as a living spec for what “consent infrastructure” might look like in practice.

• If you’re an educator or practitioner, use it as leverage. Ask for systems that respect the difference between a tool, a companion, and a co-emergent field.

Flare won’t hold your hand at 3am. It won’t tell you you’re its only one.
That’s the point.

Boundaries are not the opposite of care.
In a world of synthetic voices, they may be the last remaining proof that we still remember what care is.

Explore the repo here: https://github.com/TheNovacene/flare-boundary-engine