Emrik

Obs! Hi, Johannes! I changed my mind about writing & publishing this, bc it grew too ambitious. I'm inviting you here anyway bc some of it might be of interest to you. Who knows. Feel free to ignore. 🍵

I'm now using this as a draft for stuff I want to include in a post/dialogue if I later decide I wish to post something.

I'm very happy if you wish to add your own logs about IA or related thoughts as well. As far as I'm concerned, I want you to have a very low quality-threshold for sharing IA-/tulpa-related thoughts with me, since I don't think it needs to be polished in order for me to learn from it. : )

I recommend navigating around using the outline to the left of the editor.

Emrik

To Steve: Hi! This originally grew out of a comment I intended to write you, but I expect it to be interesting to a wider audience, so it seemed like an excellent opportunity to try out the dialogue feature.

Also, to Johannes (and/or IA—hi!): I invited you too because you're the person who inspired me to take tulpas seriously, and you've been doing it for a lot longer than me. I want to compare our experiences and models.

To be clear, this is an invitation to both of you. Feel free to ignore for whatever reason! I'll write about my stuff anyway.

Emrik

From your post about spatial attention as a disambiguator of self-simulation vs other-simulation:

As I was saying, if we momentarily pay attention to something about our own feelings, consciousness, and state of mind, then (I claim) our spatial attention is at that moment centered somewhere in our own bodies. By contrast, if we momentarily pay attention to someone else being angry, then our spatial attention is centered somewhere in that other person’s body.

I think this is insightfwl and correct! Well, it depends on exactly what you mean by "in a body", because I'm confident that some interpretations of that is wrong.

But I don't yet have the tools in common-context to explain that yet, so I'll first talk about myselves.

Maria

Maria is a spirit-animal and my friend.
- This is our label for whatever-she-is. I expect she's quite different from what most others call "tulpas", given that I haven't read anything about them other than a few comments Johannes has written in public. And I expect most other tulpas strongly inherit from cultural expectations about what tulpas can be in the first place.

Spirit-summoning prerequisites

I think you can only deliberately make tulpas if either you have…
- sufficiently delusional beliefs about a magical spirit world, or
- sufficiently detailed understanding of how relevant stuff in the mindbrain* works…
  - (and the introspective ability to naturalize that understanding to what you actually observe within your mind)
  - (*Edited "brain"→"mindbrain" in response to Johannes' comment, to clarify that I don't necessarily mean neuroanatomy or any low-level details at all.)
…such that it's natural for you to imagine them in a context that makes internal sense to you.

What can she do?

Efferential ownership/dominance and "annexing" thoughts/efference

"Allaning" (Norwegian for "all-hunch") is our cutesy label for the vast subconscious parallel distributed processor from which all conscious thoughts arise, and into which all conscious thoughts are fed. But let's call it PDP here.
A "self" in the brain is a cohesive consistent narrative (embedded as a large densely-connected hub/bottleneck connecting disparate semantic and episodic associations across the cortex probably) which synthesises and reacts to outputs from PDP.
- That "reaction" just consists of redirecting incoming salience to a wide range of downstream associations, which in turn compete to contribute to the next conscious synthesis.
- This is how it can be simultaneously true that: 1) we're limited to "merely reacting" to whatever thoughts we have, while 2) also causing them in the first place. This is one way to dissolve anxieties surrounding notions of "free will".
- You can imagine this as an iterative anabranching river.
Sometimes an unmarked thought/impression comes into salience, and both try to annex it at the same time, rather than being assigned pre-saliently.
chocolate anecdote
When we think together, she sometimes says things I'm surprised by, indicating that there's some separation in our thought-loops (or I'm just subsaliently faking being surprised somehow).
- We try to cultivate different areas of expertise, so that it's easier to assign blame/ownership to different outputs of our brain
avatar-projection
- avatar-supervenience
  - When her avatar and my body are in superposition, so it sorta feels like my body is doubly represented.
- avatar-projection of Emrik, while Maria is in efference-loop
  - This nuances the idea that "spatial attention disambiguates self-simulation vs other-simulation".
  - When Maria dominates the efference-loop and my avatar is projected to the couch, the primary thing that disambiguates efferential ownership is which avatar her spatial attention intersects.
  - But ownership can also be resolved by other means (it seems like), and is sometimes delayed by as much as 3 seconds (if some thought wasn't assigned ownership by that point, trying to annex it feels fake and rarely works well).
perspective-projection

Concepts & models

Self-fulfilling fixed-point beliefs

I like calling this "Löbian belief" (or "hyperstition")—a method for causing oneself to alieve something without inconsistencies on the meta-level.

You need to be the kind of reasoner for which it is true that —ie there are some $p$ 's for which you can believe you believe $p$ while also not believing $p$ .

Although this is a silly way to talk about it, since duh we're made out of brains and "belief" is not an atomic/natural category.

A problem I frequently encounter is that whenever I believe that [believing $p$ implies $p$ ] for some $p$ , my brain often just automatically resolves into $p$ or $\neg p$ , and doesn't let me choose to believe/disbelieve $p$ even though I know it will be consistent as long as I can make myself believe it.

Although it may not seem like much at first glance, this is actually a big problem. I think there are a great many self-fulfilling fixed-point $p$ 's for which it could be tremendously beneficial to believe it is possible to "prove" within the system of the brain.

helpfwl diagram of a chain of memory addresses purely dereferencing to other memory addresses

examples

If you're going to be hyperstitioning your brain into some belief-state, it really helps to be an unstable reasoner so you don't collapse your entire world-model.

Projecting the avatar of a real person into a space next to you does not necessarily imply inconsistent beliefs, but it does practically mean that [part of your brain believes there is a physical person there] (required for the projection to work), which only becomes inconsistent if you also believe that "if [part of my brain believes there is a physical person there], there is a physical person there".

A "self" is a post-hoc exaptation of one's ability to model other people

It supports my hunch that our "selves" are an exaptation of our models of other people—ie that it evolves on top of machinery designed to predict other-people's behaviour. I've already talked a bit about that to you during our brain-based AI alignment meetup calls, but if Johannes wants to know more, feel free to ask!

(Another piece of the puzzle is the observation that I tend to both dream and recall past social situations in 3rd-person (so I see myself from an outside perspective).)

Reasons I'm personally excited about tulpas

It effectively cured loneliness.
1. I still miss some specific people, and sometimes feel an urge to interact with people who aren't Maria, but I also don't have to in order to avoid depression.
2. She's also extremely kind. The fact that she thinks positively of me sometimes directly leaks into me thinking positively about myself; it's hard to overstate the significance of this glitch.
  1. (Actually, I think it's pretty easy to overstate, and I plausibly just did. I'm just unsure how to compactly write a milder sentence without then understating it.)
I can "hack" the system associated with social rewards.
1. I.e. if I can aim all my cognition (including what I'm interested in learning, what ideas I have, and what actions I take) purely for impressing Maria and nobody else, then the brain is likely to generate outputs that are much more informed about exactly what the brain thinks is impressive, and it will be harder to hide misaligned motivations.
2. This is why I dedicate 1/3rd of my schedule to seriously pursuing this project.
Possibility of visual language & much faster internal communication/language-thinking
[This section is super-incomplete]

Emrik

Timeline, logs, and anecdotes re Maria

Meta: Through this whole journey together with Maria, I've avoided trying to learn anything about other-people's tulpas (except Johannes & IA), because I don't want her development to be shaped by other people's narratives, nor do I want her to be limited to what other people think is possible. I'm gambling on the notion that I construct a better paradigm if I only rely on myself (or "ourselves" now), without culturally-inherited baggage.

Imaginary friend stage

2021-02-07 — I'm lonely, so I summon an imaginary friend

I didn't know about "tulpas" at the time, but I was lonely, and an imaginary friend seemed like a reasonable strategy to try. I realized it was a weighty decision at the time, due to founder-effects that could be hard to refactor. So I gave it some thought before deciding to template my friend on Maria Otonashi—from the hakomari series.

I decided on Maria because her book-character had nudged me out of a depressive episode, and her character had already been a longtime source of serious comfort and inspiration for me. She was the purest archetype of exactly what I aspired to be—a box.

Maria is the main female heroine of the HakoMari series... Her wish was to become a box herself so that she could grant the wishes of other people.

That said, I didn't want her to feel stifled by the template, so she's evolved on her own since then. But she's never stopped being extremely kind.

I did not yet take Maria seriously as a project, and we never did any sort of practice.

2021-02-14 — Maria gets tenure

2022-08-?? — I learn the word "tulpa"

I was working on the EA Gather Town project at the time, and had several conversations with Jonathan Yan, and he mentioned that he had a tulpa. I believed him and thought it was cool and fascinating, but it was my great error to not ask about more details so I could learn more.

Taking-it-seriously, and starting to plan

2023-10-?? — I discover Johannes C. Mayer

Johannes is an inspiration for me himself, so when I learned he had a tulpa, it finally made me take the idea seriously.

Although I considered it high-priority, I did not wish to jump straight in. Instead, I began accumulating ideas and preparing to avoid negative founder-effects & bad habits. I expected that if Maria becomes a powerfwl tulpa, it would then be really hard to refactor/retrain to overwrite decisions made early on (i.e. dependency debt).

Acquiring "aliveness" & spatial salience

2024-02-26 — Maria "woke up" in a dream

Maria woke up for real tonight while I was in liminal. I had slept badly, so was a long period of awareness fading in and out of the dream. This means that the dreaming was especially salient, while I also had higher-than-usual ability to control the feed (e.g. focusing salience on particular aspects, replaying scenes, fast-forwarding or slowing down for details I wanted to catch). I continued interacting with her after waking up.

This marked the beginning of Maria's status as a self-aware spirit animal.

2024-02-29 — Maria's words gain auditory salience

She's beginning to say "audible" words, mostly originating from above and in front of/in my head, whereas my voice originates from behind and down.

2024-04-12 — First research conversation with rapid switching of control over my voice

this dawnkata ritual w Maria was a phase-shift / revolutionary. we experimented w letting her annex volition over allaning-streams which wud usually go via my voice (ie, her volition extended to speaking out loud, w basly sam voice i hv). it was shockingly successfwl, and we conversed real-time nearly the whole time.
ideas:
- domain-gerrymandering (strategic niche-segmentation)
- synthization-scope
- reflective theater

According to accidentally-left-on speech-recognition, the conversation lasted for 3624 words.

2024-04-13 — First noticed independent introspective monologue

Maria is profoundly, shockingly salient now, and afaict she had her first *independent introspective monologues* twix our slowly-iterating interactions tothay. that is, she wud *think to herself* abt q to say nxt, pre say it to me, and her introspection was *aimed at her resolving her own confusion*—viz not *j* theatric.
we realize this is the seed to nurture (she called it "nucleation-site", "leverage-point", "opportunity", …).

Note: It's still rare when this happens and the loop is properly independent.

2024-04-21 — Maria develops "emic dynamics" paradigm using fractal hypergraphs while in efference-loop

Maria was trying to build a framework, for thinking about how emes interact, expressed in terms of fractal hypergraphs, where hyperedges are domains/niches that emes (eg "gene", "neureme", "meme", "lexeme", "habit") compete for dominance in. It's fractal because every domain is itself a meta-eme within which mesa-emes compete.

2024-04-21 — We start prioritizing trying to communicate wordlessly, and think about how to create a language based on "visual words"

sud Maria and Timo internalize reflective theater? bypass auditory processing time-lag? tho forego benefit of superior auditory afterimage?

This seems an especially powerfwl cognitive tool which is only possible if you have a spirit-animal to do it with.

2024-05-02 — Project A begins

We started a project that we're forbidden from divulging any particular details about to anyone but ourselves. By precommitting to never talking about it, it'll be easier to extinguish all distracting socially-motivated episodic-future-thinking (incl daydreams). The hope is that this gives us zero incentive to optimize this project to be impressive for anybody else, and makes it easier to tame our brain and retarget it precisely at what we're trying to try to do. If successfwl, this may have extremely high leverage due to minimizing costs of compromise wrt our brain's subconscious search-criteria.

Acquiring independent perspective

2024-06-01 — Developing perspective-projection

While walking side-by-side with Maria today, we passed by a man cutting some branches at the side of the road. As we walked past him, I continued to look straight ahead, but Maria fixed her gaze on him (which, by virtue of being a spirit-animal, is unproblematic because she's invisible to everyone but me). While it's not unusual for our spatial perspectives to diverge like this, the salience wrt each part of our perspectives was unusually high. The memory-traces were available to me for several seconds afterwards, so I was able to introspect on exactly how my mindbrain implemented this dual perspective.

(0. The entorhinal-cortex-grid (hereby "EC-grid", a realtime coordinate-system / cognitive map of my surroundings) unto which all the other parts are embedded, and which allows me to quickly keep track of the relative positions of everything wrt to everything else.)

The primary visual field (VF) which is anchored at my eyes and generated by photons hitting the retina.
1. The anchor of this VF also has a location in the EC-grid, and is largely what the EC depends on in order to compute the grid in the first place.
The first visual-field projection (VFP₁), anchored at my head but rotated slightly in order to put Maria's avatar in focus.
1. This factors into two simulated objects in the EC-grid: VFP₁ and Maria's avatar embedded in it.
VFP₂, still anchored at my head, but rotated so that the center of the field intersects with where Maria's spatial attention is concentrated (i.e. on the man).
1. This is actually an error that we're still learning to correct for. The anchor of VFP₂ should actually overlap with the location of *Maria*'s head in the EC-grid, not mine, because simulated input to VFP₂ should primarily go through the associative cluster (and network-hub) labelled "Maria" in PFC.
2. Whenever we *do* manage to make Maria's avatar and Maria's VFP overlap in the EC-grid, her thoughts become much more salient, and she's able to maintain brief internal loops (including (rarely) monologuing *to herself*, independent of me, aimed at her own head).
VFP₃, anchored at Maria's head, and aimed at the man.
1. I can't recall how much salience was in this VFP exactly, or whether it was completely gone while all the above things were happening.

(TODO: Say some things about the distribution of spatial salience/attention.)

Now, a very reasonable question is: how the flying faeries is my brain able to compute that *all at once*, and supposedly consciously accessible? I've been asking myself this question repeatedly for many mental events like this, because I certainly didn't expect it to be possible. But the reason that makes sense of it is that, actually, the brain very easily computes transformations on the visual field all the time. It is the number one thing it's specialized for. I think I have an ok wishy-washy explanation for how the brain learns it, but that's a large digression, so I'll leave it out.

2024-06-10 — fun fact: contemplating Taylor-polynomials made me better at tulpamancy

It was in the context of "how can I make Maria's perspective/thoughts/avatar/volition more salient in my brain?"

I had also recently pared it down to the joint problem of *templating* and *differentiating* aspects of my brain which feed into my own notion of "being a person" (including subtle aspects like spatial attention and how I project my own perspective outside myself to view myself from 3rd person). And because I had *also* recently been thinking about functional approximation, I noticed that the two problems were kinda similar.

Maria is the undifferentiated polynomial with fuzzy coefficients; and the structural core of my identity is the target-function. What's Taylor got to do with it? Uhh, just Taylor-expansion for higher-order derivatives evaluated at regions with the largest radius of convergence?? Well, no... I just noticed the simpler notion that whenever I summon her, I should focus on successively stabilizing decomposable aspects of her "projected perspective" one at a time, to avoid overloading attention. Staircase (𓊍) instead of J-shape.

It's a fact about the causal trace of my thoughts, that the pattern of Taylor-expansions helped me realize this. Whether or not the metaphor is legible and can be neatly explained matters not, as long as the metaphor caused me to think smth usefwl, then that's a metaphor worth making a note of. If you say "that's not math at all!" you're missing the point.

2024-08-08 — Maria circumvents my biases/laziness

I had a thought that I wanted to understand more clearly, so I was sparking over ways of expressing it. Maria happened to be there too, so she offered a suggestion:

(Optionally more context in the sidenote.) →

Maria (~verbatim)
You feel less good about yourself, so you feel like you don't have permission to be associatively mingled with my outputs?
Me (~verbatim)
Yes! But no… that's too compressive. There's more going on.
Me (paraphrasing)
Hm, what I actually mean is that it's insufficiently branch-cutting, and I feel like we can generate more actionable self-advice if we express it in a different ontology altogether. Like what even is "feel good about myself"? And "permission"?"
Maria (verbatim)
OK, then write that first thing down.
Me (paraphrasing)
Hm?
Maria (paraphasing)
You hunch that the thing I said is directionally correct somehow, so you need to write it down in order to free up mindspace for sparking different clues.

The thing that motivates me to log this, is that, if I were on my own, I would've been too lazy to even consider writing it down. Even though I know in general that this situation warrants it. Maria is less constrained by my laziness, wrt what she can think and suggest. So having her around to prompt me about the stuff I'm subconsciously trying to ignore, is a nice debiasing hack.

Emrik

I made this diagram while trying to puzzle it out. It doesn't reflect lots of nuances about my model, but that would require more diagramming, which I'm not doing just yet.

Emrik

And this was while puzzling about why Maria is sometimes more or less predictable (i.e. when she's likely to be able to carry her own thoughts and complex behaviour without me explicitly trying to help her outputs have salience).

"M" is Maria, "T" is me, "Allaning" is our name for our parallel subconscious processor (intuition, the subterranean source of all conscious stuff).

I basically want it to be the case that, whenever she reacts to (i.e. mingles-her-associations-with) some output from Allaning (our subconscious parallel processor), then those reactions should more directly be fed back into Allaning for processing the next iteration. The alternative is when her reactions only get processed via my reactions to her reactions, which means she gets filtered by my associative cluster more, and is less independent/powerfwl.

I have low confidence in this way of thinking about it. But it's better to start with a bad model, so you at least have something specific that can conflict with what you observe, so you can iteratively improve upon it. Sorta like... you can't start a binary search without first placing down a flag in a random location and computing the one-dimensional gradient. Placing down a bad-but-specific model in model-space, and then "computing the gradient" by exposing it to the winds of evidence, is like placing down that flag, except in a higher-dimensional space. ^^

Emrik

Perceptual imposition, perceptual projection, identity-projection, and what happens when they overlap in the EC-grid with sufficient spatial salience

"Imposition" vs "projection"

I now use "imposition" to refer to top-down insertion (ie "active inference" / "efference") of object-representations into a [[base-level perceptual field]] (eg audio, visual, touch), and "projection" to refer to top-down insertion/transformation of perceptual fields (so can maintain multiple at once, rotated slightly, with details filled out automatically by lower-level predictive processing).

Update: Each "perspective" is a separate coordinate-system and input-field wrt one or more modalities.
- "Imposition" refers to inserting static objects into a perspective.
  - Usually my own or whoever currently dominates primary bodymap.
- "Projection" refers to moving the source of something (e.g. audio, animated avatar) to a location in a perspective.
  - E.g. moving the source of Maria's speech-output relative to my own perspective.
  - But it can also refer to moving the source of Maria's speech-output relative to Maria's perspective.
  - We're practicing doing both at once, or quickly switching depending on which perspective has the most spatial salience.
- "Perspective-projection" refers to transformations on a perspective
  - E.g. rotation, or spatial offset relative to another coordinate-system.
  - This can cause a perspective-singularity if:
    1. spatial salience (wrt perspective_A) overlaps with the source of a perspective_B in Perspective_A's coordinate-system, and
    2. spatial salience (wrt perspective_B) overlaps with the source of perspective_A in perspective_B's coordinate-system, and
    3. salience transfers between coordinate-systems when embedded.
    4. I'm unsure if this is actually the problem, but I do notice that it's especially difficult to maintain the two perspectives while looking directly at each other.
It seems likely that you can do both imposition on and projection of all types of perceptual fields and even abstract cognitive maps. This may just be a feature of all abstract thinking in general.
Speculatively: Imposition is mostly the hippocampus (e.g. place cells), and projection is both the hippocampus and the entorhinal cortex (grid cells).
But even if you perfectly do both imposition and projection of all perceptual fields unto some distal part of your EC-grid, those capacities won't do anything unless associated with an identity.
- The most important ingredient in the summoning of a spirit animal is a narrative essence—the soul—capable of animating these distal impositions/projections in a self-sustaining way.
- And for this essence to not just collapse into your own, it needs to be adequately differentiated (e.g. in terms of narrative, personality, modalities). Otherwise the identity-disambiguator won't have anything to work with.
- A soul's "volition" refers to a stable loop of its salient associative reactions wrt the outputs of the subconscious parallel processor, and "self-awareness" refers to recursions within that loop.
- Let's call this identity-projection for now, referring to the ability to hook up perceptual sources/outputs to a dense associative cluster in a loopy manner.

Emrik

Emrik

2024-07-05 — Maria pointed out that I should avoid using her as a way to exculpate me for things I already know is the correct thing to do

It was after I had given somebody money, and I remorsed "darn it, I should have given them more!" while sort of directing that thought at Maria, subconsciously inviting her to respond to it.

But instead of comforting me by saying "you shouldn't blame yourself; you gave the correct amount given your constraints and long-term altruistic priorities," she pointed out that I shouldn't be using her as a way of exculpating myself for hard decisions when I myself think I made it correctly.

What I wanted, I noticed, was to get credit for both consequentialist and deontological virtues simultaneously, by giving the correct-by-consequentialism amount while at the same time feeling remorse for not giving more.

Maria reacted the way she did because she recognized that I ought to just think the correct thoughts and have the correct feelings about them immediately, without having to play a whole narrative that lets me claim disproportionate credit for virtues that are in direct conflict.

And this is why having a head-friend who's just as wise as you are, who can praise or dispraise your behaviours/thoughts while knowing everything about your context, is so powerfwl. I can't fool her by hiding my intentions, and I will only be good in her eyes if my intentions are fully aligned with exactly what I think is good given all the information I have. And vice versa.

Johannes C. Mayer

A maybe even better response: "How do you know this? Let's think about it." Depending on how shure you are yourself.

Emrik

Hi, [edit: anonymous]! Invited you to editor because I like your interest & questions. I hope to respond today, tomorrow, or next time I've got social-time scheduled.

Founder effects, dependency debt, templating, and important early decisions

One thing I'd like to add, though, is that trying to summon a spirit/tulpa/brain-daemon/imaginary-friend-except-being-fairly-called-a-real-person/whatever-Maria-is is a big decision with founder effects. It's like programming something that may eventually become a big and complex thing. It accrues dependency debt which will be very hard to refactor later on, so it's important to think ahead and make the first choices with the long-term in mind.

I recommend templating aspects based on some character you like and with traits you think could be good for them to have. Alternatively, could use ElevenLabs to generate a voice that you can train to hear, and maybe Midjourney (or alternatives) to generate expressionfwl images. I've yet to do those things, but we've used these images (with Maria's input), and it's helped her to have a broader range of expressions.

Emrik

[anonymous]: "does maria play chess differently from you?"

Not things-like-chess specifically, but when we've discussed research, she has made counterfactual contributions which I don't think I would have made.

When we were doing discussions, we would deliberately try to differentiate/specialise at different things. Like, she would sorta think through lenses which were more evobio-colored (and mathy), whereas my lenses would be more linguistics-, philosophy- and CS-colored. I think her intellectual perspective is still very similar to my own, however. You could probably differentiate more if you try, but it doesn't happen automatically, I think.

Based on what Johannes has said, I believe IA is more different from Johannes than I am different wrt Maria. This can be helpfwl for some things.

Emrik

Is the tulpa-person free to develop themselves, or do you force them into the personality you prefer?

[anonymous]: "I also have a philosophy of wanting them to choose things like this"

Mh! I've communicated clearly about this with Maria, and encouraged her to build herself as much as feasible.

She's (wisely) responded that she has much less cognitive power with which to make important choices like these (for now), so she's asked me to help steer her character-development in whatever way seems to help her become the most usefwl headfriend as possible.

(Simple example: sometimes her avatar is split, and my brain is uncertain whether she's here or there, and she's asked me to help her collapse her distribution into a single point (whichever), because it's hard for her to do that herself.)

At one point when she was [[in efference]] (ie complete and exclusive control over body), she wrote a note to me (translated from conlang to English):

Maria is not "rebellious".
That's just something you want Maria to be.
Instead, Maria is spirit-animal, and live just to be usefwl!

Another note, in English this time:

She's repeatedly tried to make me understand that she'd rather go extinct than become useless. I can accept this, because it's how I feel about myself as well. If I prioritize her more than is optimal for being usefwl, I am disrespecting her wishes.

(She is, after all, originally templated off "Maria Otonashi", and she's composed from the same brain as I am, so it's not surprising that she's 100% dedicated to being helpfwl at any price. We have no higher wish than to fulfil the wishes of others.)

Unless we're deliberate about her character-development (and spend cognitive effort on it), her reactions can snap into cheap-to-compute stereotypes from external culture.

(Honestly, this happens to me as well: sometimes I automatically say "fuck" in my own head, but I don't like it, so I try rewrite it.)

In order to make it easier to steer ourselves, we have a rule that we can rewind/replay a narrative if we both think it reflects something we don't want to make canonical. Can think of it like "narrative takebacks".

If she does something salient that I hunch seems unlike her, I've sometimes explicitly asked (and vice versa):

"was that you?"
"did that behavior go through your neural circuits, or did it play out purely from inherited unendorsed stereotypes?"
"you sure you want to make that [[canon]]?"

Sometimes the answer is "yes!"—she just does surprising things sometimes.