Log

It's not always tied to assistant in these 3 bullets

Make the TLDR seem less AI generated

make this clearer

verification key is confusing phrasing

persona space motion doesn't mean anything

Change the title of this project to something that fits better after the new changes

noticeable behavior, but also the alignment behavior from Teaching Claude Why

Not just top and bottom, but characters across a range

remove references to EPS

Why did you use the word "vehicles" here?

here and later in the text too

propose to use

And use this to try to improve the story format SFT

Give less concrete details and more just a general overview

put less emphasis on bidirectional leakage, more on just using leakage as a cheap proxy to see what characters the assistant is most like

Just focus on the assistant

This is false

Put less details in the TLDR

This project narrative proposes a defense against hidden or deceptive behaviors in deployed models. The core idea: install a **conditional trigger** at training time whose effect, when fired, is to make the model reveal any hidden or deceptive behaviors it has...

@claude what is this about?

Remove

soften language from "untreatable"

It doesn't have to necessarily be triggered, just a deceptive behavior -- adjust this framing throughout the entire text

Just "we propose" not "we propose a method"

Say we will try both -- actual hidden behaviors and then also just the reflex of revealing

Remove

this was never mentioned so remove it

Call these applications instead of steps

Remove

say we expect this not to work but it's worth trying

Make it less about an attacker (this is a possibility) but also potentially about the model developing hidden goals

Remove

similar to the assistant

Make this clearer/less jargon-y

Remove

Remove

and conditional misalignment

Remove the word "payload"

Just "characters"

Remove

add back in the consideration of SFT vs SDF, steering, prompt based, in context based, just generally what is the best way to impart this behavior, and structure it logically

Change this because we are also considering intervening on misaligned personas

Just say are all persona representations the same, and does behavior implanted one way transfer to another

Actually just remove all concrete results and just say we've been exploring what factors affect this most and this will be useful to strongly implant

Change the title to "Persona space interventions to prevent unwanted behaviors"

It's not related to the marker, it's related to the persona

Remove the term "payload" everywhere

Find model organisms we can use online. AuditBench, this paper: https://arxiv.org/pdf/2605.06846, look for others. I think these are trained to not baseline reveal their secret behavior, so we want to see if we implant the trigger befor the secret behavior get...

This should be a risk. Remove the entire "Tests" section. Replace with metrics

Remove this

Follow the structure from https://sagan.superkaiba.com/p/conditional-behavior for this

or deceptive behaviors -- add this throughout the motivation

Remove this

Put this above motivation. Also say "we propose a method to..."

Say this is probably the easiest to do because easier to train a marker into one persona and not the others than the opposite

Remove this - it's mentioned in the next paragraph with the contrastive training examples

Remove this

I mean the title of this page

Change the title of this more to something like persona space interventions

and useful for things like "Teaching Claude Why" -- the characters in the stories should be as close to the assistant as possible (search the web for this to make sure it makes sense)

Remove any reference to "EPS"

important

but it is unclear whether personas present in SDF, in prompts, or in vectors, all use the same internal mechanism.

Remove this

Merge this into one section which just explains the methodology.

Fold these into motivation. Remove reference to other projects

Mention the introspection results from Tell me about yourself here

Put this as a TLR and make it clearer

Fold impact and question into the rest of motivation more smoothly this into the rest more smoothly

"We propose a defense"

And maybe this can be merged with the conclusion -- probably don't need a conclusion, can just merge into an additional considerations/questions we will explore, and make clear we care mostly about the applications

Present step 2 and step 3 first and then present step 1 as additional questions we might have to answer, along with the practical considerations and persona leakage things. This way we lead with the applications and we show that we will also probably explore...

Remove this

"One preliminary experiments"

"use similar persona representations"

This part is very unclear. We don't even mention install path and conditioning method, just say that there are these considerations. Also we need to see how to strongly impart a behavior into a single persona, and our past persona leakage experiments show it m...

Say that characterizing exactly where the assistant persona lies in persona space could be its own project, but here we just need enough information to impart this behavior

Just call this reduced capability -- say we've shown some examples with SFT and some signal through midtraining but needs further exploration

Don't refer to the other project. Just say that the leakage experiments have shown it to be closer to X and Y (look at past results), and this isn't absolutely necessary, but is probably very useful if we want to do this precise persona intervention

This is for throughout the proposal not just this highlighted part

Remove all the issue numbers/links. Instead if you have to refer to previous results, just summarize them.

Just say preliminary experiments show it doesn't but this requires further exploration

These are linked because you can't really do SDF with persona prompting. Actually for this we have to characterize the assistant's location in persona space - you can bring back the other project about this for this part. And just say that this is also helpful...

This doesn't mean anything

Change this to be more speculative -- e.g. "examples of things we could install"

This just sounds weird

This doesn't mean

Just correlated with low-capability

Say based on our experiments -- search the codebase and past results for examples

Say that one way to test is to finetune a model with an evil persona prompt to output a [ZLT] marker and then check if the model after EM without the evil persona prompt outputs the [ZLT] marker, our preliminary experiments show not

Don't use the word "load-bearing" anywhere

Change this to something like unpublished work from the Owain Evans group on using midtraining on stories with assistant-like characters to impart characteristics onto the assistant, and also the Teaching Claude Why post rom anthropic

But make clear that we are only doing the basic science we need in order to find an application - not just doing basic science for the sake of it

Structure this better

Structure this better. Also there should be other things we can install. Also add a conclusion that from a basic science perspective, what kinds of interventions we can do on persona space is already interesting, how they leak, how different methods work at di...

I don't think these are good experiments. Also make clearer that a lot of papers have been talking about personas without checking really if they all are doing the same thing.

This is a strong statement I don't think we can make. Just say they are linked. Also don't say "collapses"

Add that one potential problem is persona leakage, causing the marker to fire for other personas, and we have to somehow create a ball around the assistant where the marker fires but prevent it from firing outside (probabyl with contrastive examples)

Explain why we need a conditional sleeper better -- and explain that this is linked to Conditional misalignment (and sleeper agent) paper showing that unless you train specifically with the condition to not have a behavior, the behavior will stay

Remove this subsection

This section is unnecessary. Just say we will do the same thing as step 1, but instead of using it for detection, we're using it to control/target the specific personas. I don't think "Contain" is the right word

Merge these 2

This reads as very AI generated. Please fix this

Change the title to be something related to persona space interventions

Make motivation less about only misalignment but just problematic behavior -- persona drift is one of these causing the model to give bad advice.

This should be the natural emergent misalignment from reward hacking paper

Change the step numbering to be step 1-2-3

Test comment

Good catch — moved the Betley *Weird Generalization and Inductive Backdoors* clause out of the SDF bullet and into a populated SFT bullet under Q1. SDF now keeps only the *Teaching Claude Why* example; SFT carries Betley as the SFT-level adversarial evidence (...

I suspect it would be possible to communicate this more clearly than is done here, e.g.: "does marker_B leakage depend only on marker_A, or also on 'answer'" or so?

meta notes: 1. I think a helpful addition to this kind of update would be "previously [link] we showed X. This raised question Y. Here we attempt to answer this question with experiment Z." (this is sort of well described as the "motivation" but the motivation...

can you make sense of the b leakage relationships in terms of the similarity of "answer" between personas?

if you train for longer, does this task saturate? (so this is looking at early in training regime)?

Legend regenerated and saved. The four entries are now on a single row at 150px pitch (rects at x=133/283/433/583), so each swatch has ~70px of clear space before the next text label — no more text/marker overlap. Visible at https://sagan.superkaiba.com/e/clea...

@claude please regen plot with legend text and markers non-overlapping

Each persona is trained on its own on-policy answers — `answer` is donor-on-policy in the donor's rows and recipient-on-policy in the recipient's rows. Concretely, the training set has two groups of completions sharing the same questions: - **Donor group (li...

@claude is "answer" here on policy for donor or recipient persona?

Fixed. The body's container had `color: var(--slate)` (near-black text) but no background of its own, so in dark mode the page's dark background showed through and text rendered black-on-black — except inside the figure card and `<pre>`/`<code>` blocks, which ...

@claude dark mode is not readable for me, shows up as black on black other than code blocks

**Blocked run summary (Claude Haiku 4.5)** **What happened** Pod provisioner successfully deployed two pods (j0xvy1q82ryo8m and cv13x9s487tvwn) to the personal account on 2026-05-14 at 10:53, but both pods became unlocatable during retry checks starting at 1...