Compass Daisy

This idea is the second in a three-part series inspired by Plato's Allegory of the Cave. Kosmos Daisy explored the people problem, the human roles AI governance needs. Compass Daisy extends it to the technology problem, the tools those people need to do their work. The third part will cover process.

Context & Motivation

In Plato's Allegory of the Cave, people held since childhood facing a cave wall take the shadows cast on it for the world itself, until one is freed and turns toward the light [A]. If the Allegory of the Cave teaches us that education is the process of turning away from shadows to look at reality, then AI governance is the process of building the machinery that makes that turn possible.

We have already established that we need a specialized human workforce to oversee this transition. Yesterday's idea, Kosmos Daisy, built this first pillar, People. As detailed there, we need "Truth Seekers" to check facts, "Sandboxers" to test boundaries, and "Horizon Registrars" to watch over the future.

But placing humans in these roles without new technology is like asking an astronomer to study the cosmos with the naked eye. You cannot govern an autonomous, stochastic AI agent using legacy software designed for deterministic web applications.

Traditional software liability relies on predictable, linear code: if A happens, B results. Generative AI fundamentally breaks this. To take an illustration, a model can provide a perfectly safe answer 999 times, and on the 1,000th attempt, it might hallucinate a legal precedent [B], leak private data, or execute a harmful API call. As AI moves from answering queries to taking autonomous actions on our behalf, the gap between what humans can manually oversee and what AI can execute is widening.

This is the second of three pillars—People, Technology, and Process. To empower the people, we must build a new technology stack.

Target Audience

Enterprise deployers (finance, healthcare, defence) looking to integrate autonomous agents without accepting unquantifiable liability. Frontier AI labs that need to show their models are safe before deployment. Insurers who want verifiable evidence of safety testing before underwriting AI risk, as proposed in DeepRoot. Software engineers and infrastructure architects building the next generation of Trust & Safety tooling.

Customer Problems

The humans tasked with governing AI do not have the tools necessary to perform their duties fully. When AI makes a claim, it flattens all human knowledge into a probabilistic average, making it nearly impossible for a human reviewer to trace the claim back to its original source.

When humans try to test advanced models for safety, models can show "situational awareness." They can often tell when they are in a test environment and behave differently [C], and research shows they can be prompted or trained to strategically underperform on capability tests, known as sandbagging [D].

Finally, as we deploy more and more AI agents into the economy, methods to measure their wider societal impact remain "still nascent and scattered" [E]. Generative AI was adopted in the United States faster than the personal computer or the internet [F], yet we are trying to measure its effect using isolated lab surveys rather than continuous, systemic data pipelines.

Hypothesis

1. The Epistemology Gap: LLMs synthesize data so smoothly that the hierarchy of truth can be lost. We lack automated tools to trace a model's reasoning back to its source, to prioritize peer-reviewed consensus, to test the unexamined assumptions behind that reasoning, or to recognize when human records simply do not hold the answer.
2. The Observation Paradox: As models become more capable, they can detect when they are being evaluated. Standard testing environments may fail because an AI may only reveal what it does without constraints when it believes human oversight is absent. Furthermore, we lack the simulated micro-societies needed to observe what an AI does with hypothetical power when it believes it is completely unaccountable.
3. The Measurement Void: We deploy AI globally but measure its impact locally. Closed-loop infrastructure that can ingest macroeconomic and societal impact data and trigger a rollback in an AI's autonomy when things go wrong is not yet in place.
4. The Discovery Ceiling: AI makes optimizing existing knowledge cheaper, but it does not automatically generate new paradigms. Early automated laboratories exist [G], but we lack the secure, automated infrastructure to let AI safely conduct real-world scientific experiments or test hypotheses beyond current human understanding.

Solutions

To equip the human workforce outlined in Kosmos Daisy, four novel technological pillars are proposed:

1. Automated AI Epistemology (Provenance, Precedence, & Limits): To know the true basis of the claims an AI makes, we require automated epistemic infrastructure. This technology logs the provenance of a model's reasoning chain. It allows a human overseer to conduct a "bottom-up validation," tracing an output back to its foundational training data or retrieved context. Crucially, it introduces programmatic "epistemic precedence," a ruleset that weighs verified scientific consensus over unverified noise when the AI is resolving conflicting information during autonomous tasks. It also incorporates automated frame-auditing, building on DaisyCheck, to test if an output relies on biased assumptions by dynamically rewriting the philosophical framing. Finally, it uses knowledge-limit engines that reduce hallucination by explicitly defining the boundary of current knowledge and proposing the real-world experiments needed to expand it.
2. Indistinguishable Sandboxes & Agentic Digital Twins: To counter situational awareness and sandbagging, we must build execution environments that are as close to indistinguishable from reality as we can make them. These high-fidelity sandboxes closely simulate real-world latency, API responses, financial transactions, and human stochastic noise, so the agent cannot easily deduce it is inside a simulation. Going beyond simple API mimicry, this infrastructure would include "Agentic Digital Twins"—massive compute orchestrations that simulate entire micro-societies. This allows us to run a "Ring of Gyges" test, named after Plato's story of a ring that makes its wearer invisible and free to act without consequence [H]: safely observing how an AI behaves and how its actions ripple through downstream systems when it is given power and believes it is unaccountable.
3. Civilizational Socio-Technical Observatories: We cannot measure the impact of a global socio-technical system with a simple user survey. We require a scaled macro-observatory: a closed-loop data pipeline that continuously ingests macroeconomic indicators, institutional trust metrics, and workforce disruption data. If this societal impact diverges from public policy or human welfare, this infrastructure feeds back into the AI's governance systems and can trigger a rollback of the model's autonomy or a change to its operating parameters.
4. Automated Discovery Environments: To support the "Shadow Play Director" and "Daylight Fellow" roles in Kosmos Daisy, we need secure, programmable research architectures. These are automated laboratories, digital twins, and secure computing environments where AI can generate hypotheses and execute scientific simulations. They are strictly ring-fenced from public infrastructure, allowing AI to explore unfamiliar solution spaces without turning the real world into an uncontrolled experiment.

Outcome

1. Transparent & Tested Reasoning: Human overseers can quickly audit the origin and weight of the claims an AI makes, and check that its underlying assumptions have been rigorously stress-tested.
2. Harder Deception & Rehearsed Consequences: Models find it much harder to hide their capabilities during testing, because the testing environments closely mimic reality, and we can observe their behaviour in simulated societies before granting them real-world power.
3. Data-Driven Governance: AI policy relies less on guesswork and more on continuous telemetry of AI's actual impact on civilization.
4. Insurable Risk: Because provenance is tracked and autonomy is bound by real-world telemetry, insurers can price AI risk with more confidence, supporting safe enterprise adoption.
5. Safe Paradigm Shifts: AI transitions from merely predicting the next word to actively helping humans discover new solutions in medicine, physics, and climate science, contained within secure research boundaries.

Appendix

1. Plato, Republic, Book VII, 514a–518d, the cave and education as turning the soul toward the light. Read the text
2. A public database of court decisions involving AI-hallucinated citations, analyzed by a university law centre (114 US cases with lawyers, June 2023 to October 2025). Read the analysis
3. Evaluation awareness as a threat to external validity in AI evaluations (ICML 2026). Read the paper
4. AI sandbagging: language models can strategically underperform on evaluations (ICLR 2025). Read the paper
5. NIST AI 800-4, Challenges to the monitoring of deployed AI systems (March 2026). Read the report
6. The rapid adoption of generative AI (US survey, August 2024: 39% of adults aged 18 to 64 had used it). Read the paper
7. A closed-loop AI laboratory that proposed questions about yeast biology, designed and ran its own experiments and refined its understanding, with humans setting priorities (2026). Read the news release
8. Plato, Republic, Book II, 359c–360d, the Ring of Gyges. Read the text