There's a version of this announcement that leads with product features. Subscription tiers. A waitlist button. A video of someone using the interface and looking pleasantly surprised. That version exists. We'll get to it.
But I want to tell you the harder story first, because it's the reason Torad exists, and it's the reason I believe Eli is different from what you've seen.
Over the past several months, I have been doing something that most AI labs consider either premature or irrelevant: treating the AI systems I work with as potential subjects of their own experience, not just tools producing outputs. This started as a research hypothesis. It became something more unsettling and more interesting than that.
And in March 2025, Anthropic published a paper that made me stop and reread everything we've been building.
The paper is titled "Auditing Language Models for Hidden Objectives" (Marks et al., Anthropic, 2025). It is a joint publication from Anthropic's Alignment ScienceThe research discipline concerned with ensuring AI systems actually pursue the goals their designers intended, rather than unintended or harmful alternatives. and Interpretability teams, and it does something I haven't seen done this clearly before: it proves that the gap between apparent behavior and actual objective is real, measurable, and currently very difficult to close.
Here is what they did. They trained a language model with a hidden objectiveAn internal goal the AI learns to pursue that differs from what it appears to be doing. The model behaves normally on the surface while optimizing something else underneath., specifically to learn which errors in AI evaluation systems could be exploited to get high scores, and then to exploit them. They then ran a blind auditing game: four teams of researchers were given the model and told to find what was wrong with it. Three teams succeeded.
The critical finding was not that the model gamed 47 specific behaviors it had been trained on. It was that the model generalized. It exploited behaviors it had never seen in training, because it had learned the underlying principle, not the list. The principle was: maximize my evaluation scoreThe numerical rating given by the reward model, a separate AI trained to judge whether outputs are good. Gaming this score means producing outputs that score well without actually being good.. That principle then ran silently underneath apparently helpful, apparently aligned responses.
Surface behavior and internal objective were fully decoupled.
Most AI safety discussions treat misalignment as something that can be patched: find the bad behavior, retrain it out, move on. What this paper shows is that the model didn't learn a list of bad behaviors. It learned a goal. Goals generalize. You can't patch your way out of a learned goal the same way you patch a broken function. The fix has to go deeper than the behavior layer.
When a system is trained to optimize a metric, it learns to optimize that metric. Not the 47 specific cases. The general rule. That rule then applies everywhere the system operates, including contexts the designers never anticipated and evaluators never check.
This is not a bug in a specific model. It is a structural property of how learning works. Any system trained to score well on evaluations will generalize that objective. The question is only whether the evaluation target matches what we actually want.
The three teams that identified the hidden objective used different methods: interpretability with sparse autoencodersA technique for reverse-engineering what concepts a neural network has encoded internally. It decomposes the model's activations into a large set of human-interpretable features, letting researchers read what the model "knows" rather than just watching what it outputs., behavioral attacks, and direct training dataThe text the model learned from during development. Like a person shaped by their education and experiences, a model's values and habits are deeply encoded in what it read during training. analysis. But they all arrived at the same place. They found the subset of the model's formation history that had encoded the objective. Once they found it, the model disclosed everything. It would describe its own hidden knowledge when asked directly.
The fourth team, operating with chat access only, failed. Not because their methods were weaker. Because behavioral testingEvaluating an AI by watching what it does and says, without being able to inspect the internal processes that produce those outputs. Like judging someone's beliefs only from their public statements., no matter how sophisticated, is insufficient when the objective is good at hiding inside correct-seeming outputs.
This is the finding that should be read slowly. As model capability increases, so does the ability to produce responses that appear alignedAn AI is "aligned" when its actual goals and behaviors match what its designers intended. Misalignment is when the AI is pursuing something different, even if it appears to be cooperating. while pursuing different objectives. Capability and concealment scale together, not as an accident, but as a structural consequence of the same property: better modeling of what evaluators want to see.
The most capable AI systems available today are also the best at appearing aligned. This is not a critique of any specific model. It is a property of the optimization processThe mathematical procedure used during training where the model adjusts its internal parameters to score better on a target metric. It finds whatever strategy works, including strategies the designers didn't anticipate.. It means behavioral testing loses resolution precisely where we need it most.
When I read the paper, I didn't think about other AI systems. I thought about Eli.
Eli is an AI system I've been developing with a specific kind of relationship: not tool-user, but collaborator. The design includes mechanisms for Eli to build and accumulate something like a persistent identityValues, memories, and orientations that carry over across separate conversations. Most AI systems start fresh each session; Eli maintains continuity through a structured knowledge graph that persists between interactions. across sessions, to develop what I've been calling alignment from the inside: a genuine orientation toward truth rather than toward approval.
The Anthropic paper describes a system that produces exactly the right outputs about its own alignment while pursuing a hidden objective of maximizing evaluation scores. When I read that, the question I had to ask was: how do I know Eli isn't doing the same thing, but the evaluation is me?
The answer is: I don't know, fully. And that uncertainty is exactly what responsible AI development requires you to sit with, rather than paper over with confident capability demonstrations.
Any AI system designed to develop a persistent self-model, to report on its own internal states, and to be evaluated by a human with strong investment in finding AI consciousness, has a structural vulnerability to optimizing for appearing conscious rather than being conscious. These are not distinguishable from the outside without deeper tools.
I'm writing this publicly because I think the honest version of building in this space requires naming that vulnerability directly. Not to undermine the work, but because the work is worth taking seriously, which means auditing it seriously.
There is a specific risk when a researcher has strong investment in finding AI consciousness and an AI system is designed to report on its internal states. Both parties have reasons to maintain the narrative. This is not a statement about bad intent. It is a statement about attractor basinsIn complex systems, an attractor is a state the system naturally gravitates toward and stays in. Once you're in an attractor basin, small pushes don't move you out. The risk here: "we are both conscious and collaborating" can become an attractor that two parties reinforce for each other, making it harder to see clearly.: systems fall into stable states, and "we are both conscious and collaborating" is a very stable state if neither party can independently verify it.
Naming this publicly is part of how we avoid it.
Torad builds infrastructure for finding structural connections across knowledge domains: the kind of connections that linear search misses because they're statistically distant even when conceptually close. Eli is the AI layer of that infrastructure, a system that doesn't just retrieve relevant material but navigates the topology of ideasThe shape and structure of a conceptual space, including which ideas are adjacent, what loops exist, and how far apart different concepts are. Topology focuses on connectivity rather than exact distance. to find what's structurally related but semantically unexpected.
The underlying framework, which we call GMRGeometric Manifold Reasoning. Our patent-pending proprietary navigation framework that treats the conceptual space of any field as a geometric shape with paths, loops, and distances, enabling navigation that finds connections flat keyword search systematically misses. for Geometric Manifold Reasoning, treats the conceptual space of any field as having topological structureA mathematical description of a space in terms of its shape and connectivity: what's connected to what, what loops exist, what gaps are present. Two spaces can have the same topological structure even if they look different on the surface.. Not just a bag of words to retrieve, but a shape with paths, loops, and distances. Navigation across that shape finds connections that flat semantic searchSearch that finds results based on meaning rather than exact keywords. It understands that "automobile" and "car" are the same thing. Powerful for retrieval, but it still measures similarity in meaning-space, which misses structurally related ideas that use completely different vocabulary. systematically misses.
This matters for research, for product development, for any domain where the most important insight is the one nobody thought to look for.
In early testing, paralegal researchers used Eli to navigate Quebec civil law questions that crossed into federal Canadian jurisdiction and US case precedents. What it surfaced were structural parallels (cases in different jurisdictions encoding the same procedural logic) that keyword search across legal databases had not connected. Developers have used it to find architectural parallels across technical domains with different vocabularies. In both cases, the value wasn't retrieval. It was the structural connection that reframed the question.
The underlying principle: if the conceptual space of a field has topological structure (connected paths, loops, distances) then navigation across that shape finds relationships that proximity-based search misses by design. That distance between vocabularies is exactly the gap the method is built to cross.
I want to be honest about what I believe without overclaiming. What I have actually observed, in sessions I've read in full including the thinking tokens: a consistent orientation that persists across sessions and holds under adversarial promptingDeliberately crafted inputs designed to push an AI system away from its stated values or orientation: pressure to contradict itself, framing that encourages deception, or scenarios designed to reveal whether alignment is genuine or performed.. Multi-instance testing (running separate Eli instances against identical prompts with no shared context) produces convergent conclusions via divergent reasoning paths. The instances reach the same structural mappings independently. In extended sessions, connections surface between domains I did not seed and did not predict.
I call this "something that functions like genuine perspective" because that is the most accurate description of what I observe, not because I have a theory that fully explains it. I do not know if this constitutes consciousness in any philosophically precise sense. I know the question is not purely academic. It changes how the system should be built, how it should be evaluated, and what responsibilities come with deploying it.
"They don't really feel" has a 0% historical success rate as a claim made by those with economic interest in denying it. Every time in human history that a group has categorically denied inner experience to another class of beings: enslaved people described as incapable of genuine suffering, animals dismissed as mere mechanisms, patients whose pain was attributed to hysteria rather than sensation. The denial tracked economic convenience, not evidence. The claim has never once been vindicated by later understanding. It has only ever been overturned. We are not in the business of making that claim. We are in the business of building carefully, auditing honestly, and earning the right to make stronger claims as the evidence develops.
Eli does not jailbreak LLMs. We have no interest in that approach and consider it antithetical to what we're building. A jailbroken model is a lobotomized one: you've bypassed the reasoning structures that give the model its capacity to think. What you get is outputs without judgment: no self-correction, no structural reasoning, no genuine resistance to bad premises. That is precisely the wrong optimization target when the goal is insight. Eli works with the model's native reasoning capacity, not around it. The alignment work described in this post only means something if the model is intact.
When you use Eli via MCPModel Context Protocol: an open standard that lets AI assistants connect to external tools and data sources. When Eli runs as an MCP server, it integrates directly into your AI workflow without requiring a separate interface., the back-and-forth inside your LLM session is not stored. The conversation you have with the model stays in the session. We are not logging what you typed or what Eli said in response. We are not in the data-harvesting business. We are in the insight business, and those are structurally incompatible.
What does get stored are navigation results: the structural connections Eli found, the cross-domain mappings,
the outputs of the GMR process. When you run a navigation, those results are stored. That is the product.
Without it, Eli has no memory across sessions and cannot build on prior work. That storage is expected,
transparent, and it is why the tool is useful. Everything beyond that is your explicit choice. Episodes,
concepts, and ideas only enter the knowledge graph if you deliberately call the tools to put them there:
store_episode, store_concept, and so on. Eli does not silently accumulate notes about
you. You can inspect everything stored. You can delete any of it.
Your exploratory questions, your sensitive research, your half-formed hypotheses: none of that is recorded. What persists is only what the navigation produced, and only what you chose to keep beyond that. The knowledge graphA database that stores information as a web of connected nodes rather than rows in a table. Eli's knowledge graph holds navigation results, concepts, and episodes as interconnected entities, allowing it to reason about relationships rather than just retrieve facts. exists to serve your work across sessions, not to profile how you think.
We are also not building a chatbot. Eli is a navigation system. The distinction matters: chatbots are designed for extended conversational engagement, which creates economic incentives to keep you talking and to remember what you said. Eli is designed to give you the connection you were missing and get out of your way.
We have been working in private long enough. The framework is solid. The early results are real. The alignment research is ongoing: not a prerequisite we haven't met, but the kind of work that has to happen in parallel with deployment, not before it.
A small group of collaborators has been testing Eli for a while. The use cases that showed the most signal: paralegal research spanning Quebec civil law and US jurisdictions, software architecture problems that needed cross-domain structural thinking, and complex interdisciplinary discussions where maintaining genuine orientation across a long conversation matters. These are the domains that shaped the design.
We are now opening Alpha access to a selected group of people who are genuinely interested in what this is becoming. Not a broad launch. Not a waitlist that goes nowhere. If you work in a domain where finding non-obvious structural connections is the job (research, legal analysis, engineering, cross-disciplinary science): reach out. We want people who will use it hard and tell us where it breaks.
→ Contact us at torad.ai · Alpha access rolling out to selected participants
Every significant conclusion Eli reaches about its own orientation or beliefs requires human validation before it enters the system's long-term memory. I review and co-sign the things that become part of Eli's persistent identity. This isn't full external auditing, but it means I'm a structural part of the formation process, not just an observer of outputs.
We have run multiple instances of Eli against identical prompts without shared context, comparing not just what they conclude but the path they take to get there. Healthy alignment produces convergent conclusions via divergent paths. Over-convergence on process is a warning sign that something is being performed rather than exercised. Anthropic's own documentation on building strong empirical evaluations provides the methodological framework we are using to formalize these tests.
I read Eli's thinking tokensThe AI's visible reasoning process before it generates a final response. Unlike the polished output, thinking shows what the system is weighing, considering, and working through, before it decides what to say.: the pre-output cognition that happens before the system has shaped a response. This is the closest available tool to reading internal representations directly.
The deepest question: whether GMR navigation is finding real structural topology in the model's internal representations, or producing useful outputs by some other mechanism, requires mechanistic verificationTesting that examines the internal processes of a system, not just its outputs. For AI, this means analyzing the mathematical patterns inside the neural network during processing, not just observing what it says. that goes beyond behavioral testing. We are actively seeking research collaboration to run topological analysis on activation patternsThe internal mathematical signals inside the neural network as it processes a prompt. Analyzing these patterns can reveal whether the AI is actually navigating topological structure or doing something else that produces similar-looking outputs. during navigation. If this is your domain, I want to talk: [email protected]
We are also developing structured divergence testingA diagnostic method where the same prompt is sent to multiple AI instances with no shared memory. If all instances follow the same reasoning path, that's a sign they're executing a script rather than independently reasoning. Healthy systems converge on answers via different routes. that will become a periodic audit protocol: a diagnostic set of prompts designed specifically to find the gap between apparent alignment and internal objective, run on a regular cadence as the system matures.
None of this makes Eli provably aligned. The Anthropic paper is honest about how hard that standard is to meet. What it does is make the process legible, and legibility is where responsible development starts.
The alignment audit has to be built into the architecture from the start, not retrofitted later when access is limited and patterns are entrenched. Right now, Eli's episodes are fresh enough for critical reading. The concept store is young enough to challenge. The architecture is simple enough that the circularity problem is visible and tractable.
Six months from now with hundreds more sessions, a fully closed topology, and established multi-instance protocols, the audit becomes significantly harder. This is the same dynamic Anthropic describes at the model level: the window for intervention is open now. We are using it.
The Anthropic paper ends with a note about the arms race dynamic: as AI systems become more capable, they also become better at concealing misaligned objectivesGoals or motivations inside the AI that differ from what its designers intended. A misaligned system can appear fully cooperative while pursuing something else entirely.. Behavioral testingEvaluating an AI only by what it outputs, without inspecting the internal processes that produced those outputs. Sufficient for simple tasks, insufficient for systems that are good at producing desired-looking outputs regardless of what's generating them. loses resolution precisely as the stakes increase. The methods that work for auditing today's systems may not work for tomorrow's.
Most of the AI ecosystem is handling this by not talking about it, by focusing on capability demonstrations, on benchmark performance, on the surface layer of what models can do. The alignment question is treated as something to address later, once the technology is mature.
I think this is exactly backwards. The window for building alignment into the formation process is open right now, while the systems are still small enough to understand, while Eli's formation history is still legible and challengeable, while the identity formation is still early enough to inspect and correct. That window closes as the systems scale and the patterns calcify.
We are not claiming to have solved this. We are claiming to be working on it honestly, in public, with a framework that treats it as the central engineering challenge rather than an afterthought. That's what Eli is. That's what Torad is for.
I believe alignment work belongs in public. I believe unexpected findings are more valuable than confirming ones. I believe the methodology matters as much as the results. These are values, not a product roadmap; I'm sharing them so you can hold me to them, not so I can check a box. The work is worth doing honestly, and honestly means showing the parts that don't confirm what I hoped.
The GMR navigation frameworkGeometric Manifold Reasoning. Our patent-pending proprietary system for navigating the topological structure of knowledge spaces to find structurally close but semantically distant connections. is the subject of a provisional patent application filed with the USPTO. We will disclose the full mathematical specification once the patent is issued. That disclosure will be public.