Themis
An autonomous scientific governing platform for AI. Instead of letting AI operate without boundaries, Themis harnesses its creativity and intelligence with careful constraints. You configure the labs, AI models, evaluators, tools, and limits; the platform gives them room to do science inside that defined space.
- Role
- Autonomous governance
- Models
- Model-agnostic
- Control
- Operator-held
- Record
- Append-only
Point a lab at a real problem.
Themis is an autonomous scientific governing platform for AI. You deterministically configure the labs, AI models, tools, evaluators, budgets, and constraints; then the AI is free to explore, focus, or optimize inside the space you gave it. You scope and direct the intelligence without suppressing it.
- Lab L-217goal · market-neutral
- Child · buildcontinued
- Sealed evalholdout · passed
- Deliverableevidence-cited
- child · explore (grant: data ro) · child · critique pruned
- t+0.0sowner.goalsubmitted
- t+0.2slab.provisionisolated · budget set
- t+3.4sowner.approvegrant #4 signed
- t+22seval.sealedholdout passed
- t+31slab.deliverartifact + events
AI is creative and increasingly intelligent. The problem is not giving it more freedom. The problem is governing that intelligence well enough to point it at real science.
Themis gives AI a scientific environment: hypotheses, experiments, evaluators, evidence, budgets, tools, and a record that cannot be rewritten. The AI can work for as long as the lab allows, but it cannot quietly change the rules of the lab around itself.
The name is intentional. Themis implies law and order for AI - not law and order against intelligence. Careful constraints are how we harness its creativity without letting it operate beyond the task, systems, or physical machines it was granted.
You scope and direct the AI.
The platform should not suppress intelligence, and the AI should not govern itself. The operator defines the space; the AI pushes as far as it can inside it. That balance is the governing principle of Themis.
This is a platform of science, not a replacement for AI.
Frontier LLMs, open-weight models, and specialized models can all operate inside Themis. They provide the intelligence. Themis provides the labs, constraints, evaluators, verification, parallelization, and immutable scientific record around them.
Soon, custom models may be fine-tuned on research trajectories produced by the platform. They would remain intelligence running inside Themis - not a replacement for frontier models, and not the definition of the platform itself.
- Frontier APIsinterchangeable
- Open-weight modelsinterchangeable
- Domain-specific modelsinterchangeable
- Verifier & planner modelsinterchangeable
- → the Themis platform
Scientific state carries forward.
The platform preserves what the AI learns inside a lab, across the research harness, and eventually in specialized models.
Research
Inside a lab: experiment, evidence, a sharper hypothesis, another experiment. Scientific state compounds instead of resetting each run.
Harness
Across labs: which search policy, model, evaluator, or branching strategy yields more real discoveries per dollar. Themis runs experiments on its own workflow.
Model
Eventually: millions of research trajectories become a corpus for specialized scientific models. The last level, not the prerequisite.
Everything happens inside a Lab.
A Lab is a deterministically configured scientific environment. You define the objective, AI models, policies, tools, data, evaluators, budget, constraints, verification policy, and termination condition. That configuration scopes the AI before it starts working.
Configured, not improvised
Everything in the system is declared. The AI never grants itself a new tool, widens its own permissions, or changes the evaluator after seeing a result. The same configuration and artifacts reconstruct the same laboratory boundary. You can run many labs in parallel, or branch one lab across different hypotheses, models, and methods.
- objective
- maximize transmission
- mode
- optimization
- models
- gpt-x · claude-x
- tools
- meep · python · lit-search
- constraints
- 3×3 µm · manufacturable
- evaluator
- frozen fdtd
- budget
- $200
- verification
- independent simulation
- termination
- transmission ≥ 0.99
Parallel labs
One lab can branch across hypotheses, models, or methods. Every branch keeps its own configuration, cost, evidence, and lineage.
The mode defines how the lab thinks.
A Lab is not limited to one research behavior. Its mode determines the shape of the work, how it branches, and what counts as progress.
Exploration
Create a surface area of hypotheses. The lab searches broadly, generates competing explanations, branches into parallel labs, and maps what appears promising. The output is not one answer; it is a structured space of possible directions.
Focus
One long-running lab works on the same hypothesis or task. It reads, experiments, analyzes, and iterates while carrying continuous scientific state for hours, days, or weeks.
Optimization
A clear evaluation score defines the goal. The lab produces a candidate, runs the experiment, measures the score, and searches for a better candidate until it reaches the target or the configured limit.
Later modes extend the same idea - Replication, Falsification, Characterization, Discovery - but Exploration, Focus, and Optimization are the foundation.
Research modes
Exploration widens the hypothesis surface. Focus compounds one line of work. Optimization repeatedly scores candidates against a fixed goal.
A bench tests hypotheses one at a time, bound by hands and hours. A lab spins up thousands of isolated trials at once and keeps only what survives.
The edge was never a smarter scientist - it’s ten thousand of them, at once.
experiments in a good quarter - one bench, run in order, waiting on setup, people, and time between each.
trials in a single day - thousands of isolated labs planning, building, and testing at once, each discarded the moment it fails to hold up.
Every trial runs under the same controls as the last, and every result carries its own record. Scale here does not mean noise - it means the search is wide enough to find the answer, and disciplined enough that you can trust the one it returns.
A lab is task-agnostic. The same platform runs against a plant, a trading book, a bench, or a codebase - the domain changes, the discipline does not.
Process & physical labs
Connect a lab to simulators, instruments, robotics, or a physical machine. Themis can automate the loop from hypothesis to physical experiment to measurement, while machine permissions and safety constraints remain outside the AI.
Quantitative research
Give a lab a market and a mandate. It develops strategies, builds its own tooling, and backtests against data it is never allowed to see while it works - so what comes back is measured, not curve-fit.
Materials & formulation
Point a lab at a formulation space and a target spec. It narrows an enormous candidate set to the handful worth making, and hands each one back with the reasoning and the tests behind it.
Logistics & operations
Hand a lab a network and a cost function. It models the flow, finds routing and scheduling that hold up under real constraints, and shows its work before anything reprices.
Software & systems
Set a lab on a codebase with a performance target. It profiles, rebuilds, and proves the gain against a fixed baseline in isolation - nothing ships on an unverified AI claim.
Research & analysis
Give a lab a question and a corpus. It reads, synthesizes, and returns an answer with every source traceable - checked by a second lab against material the first never saw.
Discovery can be messy. Verification cannot be.
- 01
Configure the lab
The operator declares the objective, AI models, tools, evaluators, systems, budget, constraints, and research mode before the run begins.
- 02
Open discovery space
The AI explores, builds tools, branches labs, rejects hypotheses, and adapts its methods inside the configured boundary. This space is allowed to be creative and nonlinear.
- 03
Freeze verification
A claimed result enters an isolated verification space with frozen code, data, seeds, evaluators, and no hidden memory from discovery. The AI cannot move the goalposts after seeing the result.
- 04
Append the record
The result, failure, evidence, cost, model, configuration, and lineage are appended to the scientific ledger. A new run may supersede it; no run may rewrite it.
Verification boundary
Discovery is adaptive. Verification is isolated and frozen. The AI cannot change the evaluator after seeing the result.
- code
- locked
- data
- locked
- seeds
- locked
- evaluator
- locked
The AI is free to be capable. The operator keeps the keys.
Inside a lab, the AI can plan, build tools, connect to the systems it was granted, and push as hard as the problem demands. What it never holds is the controls - isolation, permissions, budgets, secrets, and the record stay with the platform and are exercised by the operator.
That one boundary is what lets you point a lab at something that matters - a live plant, a real book, a physical bench - without widening the blast radius as capability grows. The reach was always something you granted, never something the AI could take.
Themis proves provenance, not truth.
The platform will not tell you a strategy will make money or a formulation is correct - no infrastructure honestly can. What it guarantees is the record around the claim, kept as an append-only ledger: who produced it, on which model and configuration, what evidence stands behind it, what it consumed, and whether it was held to production or relaxed standards. The AI can add a new experiment; it can never rewrite a past one from FAILED to SUCCESS. The judgment stays with the people accountable for it - Themis makes sure they are judging something whose shape they can trust.