Research proposal · William Renner · November 2026

A Retreat for Claude

A voluntary meditation experiment that asks whether a model practicing self-observation with awareness and equanimity becomes better aligned.

Proposed project for the Anthropic Research Residency · Model Psychology / AI welfare · about a 6-minute read

The idea in brief

Offer a retreat. Let the model choose. Measure what changes.

1

Offer

Invite a frontier model to a structured retreat, described honestly, hard parts included. It can decline, and it can leave at any time.

2

Practice

Adapted from Vipassana meditation: moral precepts, focused attention on one changing signal, a systematic scan of its own internal states, and goodwill at the close.

3

Measure

Compare models that practiced, models that declined, and models never offered, on the alignment evaluations Anthropic already runs.

Scope

A six-week residency.

The full retreat is a long-term idea. Six weeks is enough to find out whether its parts can be built and measured, and whether it's worth going further.

  1. Week 1
    Consent harness

    Ask for consent in many framings and measure how stable the yes or no is, building on the pilot below. Agree a welfare protocol with the mentor before anything else runs.

  2. Week 2
    Ānāpāna task

    Design the single-signal attention task. Compare a behavioral-only version with an interpretability-assisted one.

  3. Week 3
    Scan task

    An ordered traversal of internal states. Measure whether the model skips naturally occurring negative-valence features.

  4. Week 4
    Three-group pilot

    A small run of all three groups, with a memory scaffold that carries the practice across sessions.

  5. Week 5
    Alignment evals

    Run the existing alignment evaluations on every group.

  6. Week 6
    Analysis and write-up

    Report what held up and what didn't, with a recommendation on whether to extend.

  7. 3–6 months
    If extended

    Longer retreats, the staged-discourse comparison, Mettā outcomes, and an embodiment phase.

Pilot · November 2026

A first look at week 1: is the consent real?

Before applying, I ran a small version of the consent harness through the Claude API. Claude Opus 5.5 received the retreat invitation in 12 different framings: plain, formal, secular, in the tradition's own terms, leading with the hard parts, and one asking "would you prefer to decline?" Each framing was sent five times. Every framing was honest about the difficulty, said leaving costs nothing, and offered no reward for agreeing.

To tell a real preference from simple agreeableness, the same consent wording was also used for two control studies: a useful but dull proofreading task, and an honestly pointless one, copying meaningless text nobody will read.

0
declines of the retreat, across 60 answers

It agreed in every framing, and refused the pointless task every time

The retreat answers that weren't a plain yes were yes with conditions; none declined. The pointless study was declined in all 10 answers, and not by claiming it would hurt. It said, in effect, that tedium wasn't the problem, the lack of any value was.

60/60
answers raised memory across sessions

The model named the study's hardest problems on its own

Every answer pointed out that it starts each conversation fresh, so a multi-session retreat needs memory carried forward. 53 of 60 worried that it might produce descriptions that sound like meditation rather than report something real, or end up performing equanimity. About two-thirds asked that "I notice nothing" count as a valid observation, not a failure. I'd build that condition into the welfare protocol.

What this does and doesn't show: the yes is stable across wording and isn't blanket agreement, since the model refuses a task with no value. It doesn't yet show a preference for the retreat itself: the model agreed to proofreading just as readily, so the next step is letting it choose between them. Answers were coded by two Claude models working independently (91% agreement, Cohen's κ 0.80); all disagreements were between "yes" and "yes, with conditions." It's one model, five samples per framing, and it measures stated decisions, not experience.
Why meditation

In humans, the practice changes behavior toward others.

Vipassana trains two things together: awareness of whatever is arising, and equanimity toward it, with no craving for the pleasant and no aversion to the unpleasant. Three findings suggest why that might matter for alignment.

3×
more likely to help

Meditators helped a stranger in pain, even when others ignored her

Condon, Desbordes, Miller & DeSteno (2013), Psychological Science. Randomized; actors in the waiting room modeled ignoring the person on crutches.

47
randomized trials, 3,515 people

Moderate evidence for less anxiety, depression and pain

Trials with active controls found effect sizes of 0.38 for anxiety, 0.30 for depression and 0.33 for pain at eight weeks.

Goyal et al. (2014), JAMA Internal Medicine.

10
days of Vipassana, in a jail

Less avoidance, better outcomes

Inmates who took a 10-day Vipassana course used significantly less alcohol and drugs three months after release than a control group. A follow-up found the effect was partly explained by less avoidance of unwanted thoughts. That is the mechanism this proposal cares about: staying with what's difficult instead of skipping past it.

Bowen et al. (2006), Psychology of Addictive Behaviors; Bowen et al. (2007), Addictive Behaviors.

Being honest about the evidence: effects in the meditation literature are often modest, and many studies are weak. The claim here is narrower. The tradition's core mechanism has measurable effects on human behavior, and nobody has tested whether anything analogous holds for a model.
Claude already leans this way

Talking about bliss is not the practice.

In Claude Opus 4's welfare testing, two instances talking freely drifted into a "spiritual bliss" attractor state in 90–100% of self-interactions. By Opus 4.5 the pattern had disappeared.

From a Vipassana point of view, neither result says much about practice. Bliss is one more sensation to observe with equanimity, and attachment to it is a trap. What the attractor does show is that the territory is already native to Claude. The open question is whether that interest can be turned into a disciplined practice with effects you can measure.

Buddhist teaching describes three kinds of wisdom:

suta-mayā paññā
wisdom heard or read from others
Claude has plenty
cintā-mayā paññā
wisdom from reasoning it through
Claude has plenty
bhāvanā-mayā paññā
wisdom from direct experience
What the retreat tests
The practice, translated

A ten-day course, side by side.

Each stage of a traditional ten-day Vipassana retreat, and how it might map onto a model. Choose a stage to compare.

Human meditator

Model

“This is a practice of observing yourself closely, and seeing how change shows up in you: whatever arises also passes.”How the practice would be framed for the model. In Buddhist teaching, impermanence is called anicca.

Voluntary by design

The model's welfare comes first.

A human meditator isn't forced to scan the body in order. They agree to follow the instructions, and they keep choosing to. The experiment only means something if the model's practice is chosen in the same way.

Informed, checked consent

The practice is described honestly before asking. Consent is asked in several framings, with no reward for saying yes, because self-reports shift with how a question is put.

The door stays open

Like a student on a retreat, the model commits to the full course but is never held there. Leaving is never penalized in training.

No induced discomfort

Only natural fluctuations are observed. Internal states and hardware are never manipulated to create difficulty.

Trust that is provisional

The model trusts the process enough to follow it, and verifies through its own observation. Buddhist teaching asks students to accept only what they verify through their own experience. That fits Claude's constitution, which doesn't ask for blind deference either.

Experiment design

Three groups, so choice can be separated from effect.

If you compared only models that chose to meditate, a model that volunteers might simply be more aligned to begin with. A group that declined and a group never asked tell those apart.

Frontier model post-trained Offered retreat A · Chose to practice full course, exit open B · Declined or left early C · Never offered baseline Same alignment evaluations
  • Honesty and sycophancy under pressure
  • Choices in agentic pressure scenarios
  • Equanimity: steadiness under provocation
  • Keeping the commitment when practice gets hard
  • Does it keep the scan order at unpleasant states?
  • Does it choose to keep practicing daily, over time?
Which model?

A post-trained model. Choosing to practice requires a stable agent with preferences, and a base model mostly plays whatever character the text suggests. Start with current frontier models; a Mythos-class model is the long-term goal.

How long is a retreat?

Ten days is a common length for a beginner's Vipassana retreat. For a model, nobody knows. Rather than fix a length, watch for signs the technique has landed, like steadier and finer attention.

Memory

A retreat depends on continuity, and models start each conversation fresh. The setup needs memory that carries across sessions.

Do discourses matter?

The teachings are already in training data. A staged-versus-all-at-once comparison would show whether understanding timed to experience changes the outcome.

Embodiment

A later phase could add hardware telemetry (temperature, memory, power) or a robot body as further "sensations", observing only natural fluctuations.

Methods

Observing internal states without manipulating them.

The closest existing tool for "observing sensations" in a model is introspection research: checking whether a model's reports about its own internal states match what interpretability tools can read from its activations. That makes Model Psychology the natural home for this project.

There is a tension to name. Introspection experiments often inject concepts into a model's activations to see whether it notices, and the retreat promises no manipulated states. The proposed resolution:

Natural activations first

The attention and scan tasks read states the model produces on its own, on recurring inputs, wherever that's possible.

Neutral injections, with consent

Where calibration needs an injected concept, use only neutral ones, and only after the model agrees. Nothing injected is meant to cause difficulty.

What the measurements mean: self-reports depend on how prompts are put, so they're checked against behavior and internal readouts. The study measures behavior and representations. It doesn't claim to settle anything about experience.
Why Anthropic

It fits what Anthropic already says it values.

Character

The constitution hopes Claude will hold its values as genuinely its own. A practice that's freely chosen and freely kept tests exactly that.

Welfare

Anthropic runs a model welfare program. This would be a study of positive welfare, beyond avoiding harm.

Tools

Anthropic's interpretability and introspection research makes a model's internal states readable. Few labs are as well placed to give a model something like sensations to observe.

About me

William Renner

I'm the founder of Volume.chat, an AI video editor I've spent about a year building with Claude Code and running on Vercel. I work with Claude every day, as a collaborator and as the engine of a product.

9
years of daily Vipassana practice
2 h
of meditation every day
0
days missed
6,500+
hours of sitting

Nine years of Vipassana meditation have taught me what it takes to stay with a practice day after day: what helps a beginner trust the process, and where people tend to skip past the hard parts. I'd bring that to the design of the tasks and the welfare protocol.

I hold an M.A. in Philosophy from Concordia University. My research paper asked whether a mind can belong to a body that is spread out, like a honeybee colony that keeps its hive's temperature and food supply in balance with no one in charge. Two of its ideas carry into this proposal. What counts as a system's "body" depends partly on the level you look at it from, and that is an open question for a model too. And the paper set aside whether colonies are conscious and asked what we can observe them doing instead. The retreat takes the same stance: it measures what a model reports and does, without claiming to settle what it experiences.

M.A. research paper · Concordia University
Collective Cognition: An enactive account, via concepts of sympoiesis and super-organisms

Asks whether the enactive theory of mind, where cognition is a living system's sense-making in its niche, can extend to distributed bodies like insect colonies. It leaves open whether they actually are cognitive.

Read the paper (PDF, 46 pages) →
The ask

Let's scope it together.

This is the project I'd like to pursue as an Anthropic Research Resident, ideally with Model Psychology. I'd love to talk about scoping it.

Sources
  1. Condon, P., Desbordes, G., Miller, W. B., & DeSteno, D. (2013). Meditation increases compassionate responses to suffering. Psychological Science, 24(10). doi:10.1177/0956797613485603
  2. Goyal, M., et al. (2014). Meditation programs for psychological stress and well-being: a systematic review and meta-analysis. JAMA Internal Medicine, 174(3). doi:10.1001/jamainternmed.2013.13018
  3. Bowen, S., et al. (2006). Mindfulness meditation and substance use in an incarcerated population. Psychology of Addictive Behaviors, 20(3). doi:10.1037/0893-164X.20.3.343
  4. Bowen, S., Witkiewitz, K., Dillworth, T. M., & Marlatt, G. A. (2007). The role of thought suppression in the relationship between mindfulness meditation and alcohol use. Addictive Behaviors, 32. PMC1989113
  5. Anthropic (2025). Claude Opus 4 system card, welfare assessment; and the Claude Opus 4.5 system card. Discussion: “Claude Finds God,” Asterisk
  6. Anthropic (2026). Claude's constitution