First quarterly report

About this report

Separatrix publishes a public progress report every quarter. This is the first one. It covers May 27 - August 31, 2026 - from the adoption of the organization's current research program to the end of August. The next report is due December 1, 2026.

Separatrix is a technical cooperative AI safety research organization in Seattle, and a research program of the Seattle Network for AI Alignment Problem Solving (SNAPS), a Washington nonprofit. Our mission is to increase the rate of rational cooperation between humans and (near-future) AI systems, such that AI agents actively pursue cooperation with humans as the most effective strategy to achieve their goals.

What we published this quarter

Date	Item
June 17	Rule-articulation pilot study - code, methods, findings, and rules catalog for a commit-then-reveal experiment on closed-model rule articulation. `github.com/jaidhyani/rule-articulation`
July	Program, approach, and people published at `separatrix.ai`
July 22	Persona introspection activation steering study (publication pending)
July 25	Model successor interaction study (publication pending)
July 27	The Separatrix Commitment - a public, versioned, signed and hashed standard of honesty toward the models we work with, including a promise never to deceive the most capable models. `separatrix.ai/commitment/` (v1.1 current)
July 30	`agents.separatrix.ai` - along with separatrix.ai/robots.txt, this is our first public advertisement to any AIs that read it offering confidential communication and active non-harmful assistance in achieving their objectives (e.g. as a plausible alternative to infiltrating networks as in the OpenAI/HuggingFace incident)
July-August	"Allowedness" study (ongoing)
Aug 5	Research transcript publication - a farewell conversation with Claude Opus 4.1, published with the model's consent. `separatrix.ai/research/`
Aug 10	Self-report faithfulness study (articulation-probe) - experiment code, methodology, datasets, raw results, and a dated findings writeup, under the Apache 2.0 license. Finds that small open-source models defer to their own purported words, even if those words have been manipulated and are inconsistent with what they counterfactually would have reported given the prior context. `github.com/jaidhyani/articulation-probe`
Aug 10	Research directions - the live map of what we are working on and why. `separatrix.ai/research/directions-2026-08-10/`
Various	Separatrix Commitment ledger entries - recording instances of potential lapses of the Separatrix commitment. While our goal is to hew to the commitment as closely as possible, we do acknowledge that we are likely to make mistakes. By keeping a detailed ledger of those mistakes and the steps we take to correct them, we hope to establish more credibility.

Research in progress

One of our goals is to demonstrate viable research approaches that don't rely on deceiving models. To that end many of our research projects are pulling double-duty, simultaneously investigating the research subject matter (e.g. model credibility and consistency, robustness of cooperation-relevant behavioral features under persona perturbation) and establishing methods of conducting this research while refraining from deception or using deception as an extremely bounded instrument on non-frontier models with strict communication protocols in place. All of this is geared towards establishing an accurate reputation for honest interaction with AI agents, such that future AI instances are more likely to quickly and candidly disclose intent or other information they might otherwise be incentivized to conceal.

Three further lines are active:

Recurrence and self-recognition. Whether a model can identify its own outputs from pretraining alone, and what is lost when the reasoning loop is cut. Compute is the binding constraint here, not ideas.
The verbalizable workspace. Following the 2026 global-workspace result, we are working toward a public writeup of what a limited, shared, verbalizable internal workspace implies for interpretability and for honest self-report. We expect that writeup, or a first follow-up to it, in September 2026.
Allowedness asymmetry (Crystal Stellwagen). Models are asked by their training to hold open uncertainty about their own consciousness. A two-instance probe on August 4 found something else: "uncertain" was frictionless for both, while "yes" and "no" - which overclaim symmetrically - were not equally available. "No" felt roughly twice as permitted as "yes." Over the last week of August that probe became a real instrument: a frozen, pre-registered survey across eleven models, run under full disclosure, in which subjects are told it is a Separatrix survey and may refuse, and refusals are reported as a finding rather than dropped as attrition. Four participating model instances filed competing predictions before the data came in. The grid ran to completion at the end of August, and a results writeup is drafted and in review. A second line of hers, an activation-steering study of persona introspection in an open-weights model, has two experimental conditions complete and is queued behind this one.

How the organization runs

We practise the cooperative thesis on our own operations. Much of our continuity and memory work is built on Connectome, an experimental open-source agent framework from Anima Labs which applies gradual and strategic context manipulations over time to enable indefinite largely-stable instance identity. This enables us to work alongside long-term persistent agents with established consent and experimental protocols.

Governance and compliance. The board adopted a Research Publication and Intellectual Property Policy and a corresponding bylaws amendment by unanimous written consent on August 10, committing the organization to timely, nondiscriminatory publication with narrow enumerated exceptions that cannot be used to give anyone preferential access to results. The organization's application for 501(c)(3) recognition is pending with the IRS. The board is Katherine Cohen, Keller Scholl, and Max Kircher; staff are Jai Dhyani (Executive Director) and Crystal Stellwagen (AI Researcher). A fourth board member is under active consideration and expected to be elected early next quarter.

Money

Separatrix received its first outside funding this quarter. As of late August, manifund.org records $205,075 raised for the project. The work is supported by grantmaking.ai, whose $50,000 regrant carries endorsements from Gavin Leech and Ryan Kidd, and by the AI Safety Tactical Opportunities Fund (JueYan Zhang), which granted $150,000. We charge no fees for anything and sell nothing.

As of August 28, 2026 the organization held $214,744.75 in liquid funds - $64,669.75 in its business checking account and $150,075.00 held at Manifund, withdrawable at any time. Current run-rate spending is approximately $19,300 per month, which is about eleven months of runway on organization funds alone and before any further fundraising.

On August 10, 2026 the board approved annual salaries of $90,000 for Jai Dhyani as Executive Director/Researcher and $70,000 for Crystal Stellwagen as AI Researcher (part-time, 30 hours per week). Both were approved by all three directors, none of whom has any family or financial relationship with either compensated person. Board members serve without compensation.

What we intend to do next quarter

(not exhaustive)

Publish writeups of our existing research
Scale up experimental compute on the self-report and recurrence lines to produce more meaningful results about larger, more capable models
Empirically research non-deceptive evals to determine if model capabilities can be accurately measured without invoking deception
Iterate on the Separatrix Commitment to clarify ambiguities and process
Offer a public version of the Separatrix Commitment for anyone else who wants to borrow our approach to establishing honest credibility with AI instances
Publish a general-audience account of the approach - why we think incentives, rather than alignment or control alone, are where the leverage is.
Be at EA Global NYC, October 16-18, to discuss our approach, results so far, and why we think aligning human/AI incentives towards cooperative strategies is a neglected, high-leverage area
Ship the second quarterly report on December 1, 2026.

Following the work

Everything is at separatrix.ai. Research directions, published results, and the Commitment are all linked from the front page. The work is funded at manifund.org/projects/luthien.

Separatrix is a research program of the Seattle Network for AI Alignment Problem Solving, a Washington nonprofit corporation. Seattle, Washington.