Mercury · Athanor
A laboratory that runs itself, and rebuilds its own instruments.
Athanor is the next iteration of Mercury: not an agent that performs research tasks, but an autonomous experimental control system. It holds a state of belief about the market, chooses which unknown is worth resolving next, designs the cheapest experiment that resolves it, runs it, and updates.
The optimisation target is not the strategy. It is the method that produces strategies.
Why a furnace
The athanor was the alchemist's self-feeding furnace: a design whose entire purpose was to hold its own heat constant for months, so that a transmutation too slow for a human to attend could finish anyway.
That is the claim exactly. Not a machine that produces gold. A machine that maintains the conditions under which the work can continue without someone standing over it.
Provenance
The architecture is not invented from trading first principles. It is a deliberate transfer from self-driving laboratories in protein engineering and biocatalysis — closed-loop platforms where Bayesian optimisation selects the next physical experiment, robotics executes it, and the posterior updates without a human in the cycle.
That field has already paid the tuition on the hard problems: how to plan experiments under a real budget, how to keep multi-agent reasoning grounded in measurement rather than eloquence, how to treat curated data and unbiased benchmarking as primary research rather than plumbing, and how to learn from failed experiments as rigorously as from successful ones.
Markets are a cheaper substrate than a wet lab. The experiment costs compute, not reagents, and the feedback arrives in minutes rather than days. Almost everything else transfers.
The shift
Most "AI for trading research" is an analyst: it reads the outcome of an experiment a human chose. The whole difference is one step upstream — who chooses the experiment.
| Conventional research loop | Athanor |
|---|---|
| A human decides what to investigate | A policy ranks every open question by expected information per unit of cost |
| Interesting results get written up | Null results are first-class objects with the same schema as wins |
| A finding lives in a commit message | A finding lives in a lineage: which mutation, which evidence, which replication |
| A better model means better answers | A better model means a better prior — the measurement apparatus is unchanged |
| Blocked research waits for a human | Blocked research emits an instrumentation request and unblocks itself |
The closed loop
Eight stages, and the output of the last one is the input of the first. Nothing in the cycle requires a human to advance it; a human sets the objective and the budget.
- 1
Knowledge state
Every claim the system holds, with a posterior, a mechanism, and an explicit list of what is still unknown.
- 2
Research policy
Not "what shall we look at today." What would most reduce uncertainty in the parts of the space that are worth money.
- 3
Research team
Specialised reasoners with different objectives and different permissions collide over the candidate claims.
- 4
Claim compiler
A hypothesis becomes a measurable statement with a pre-declared population, control, and kill condition — or it is rejected as untestable.
- 5
Experiment planner
The cheapest design that still separates the hypotheses, chosen against a real budget.
- 6
Experimental substrate
History, replay, counterfactual, shadow, and forward observation — five ways to ask reality a question, at five different prices.
- 7
Evidence engine
Measurement plus quality control: leakage, selection effects, and understated uncertainty are caught here or not at all.
- 8
Belief update
Posteriors move, lineage is recorded, and the result — including a null one — becomes a permanent object.
The belief update feeds the knowledge state, and the loop closes. Each turn of it costs money and buys information; the system is judged on that exchange rate.
The governor: research as a spending decision
This is the component that separates a laboratory from a chat window. The system holds many open questions and cannot afford all of them, so it ranks them. An experiment earns its turn on four factors:
- Information gain
- How much uncertainty this experiment is expected to remove, given what is already believed.
- Economic relevance
- Whether the answer touches capital that is actually deployed. Curiosity does not qualify.
- Actionability
- Whether a positive answer changes a decision. A true fact that changes nothing is not worth buying.
- Cost
- Compute, model quota, calendar time, and operator attention — a hard budget, not a notional one. Without a real denominator the ranking degenerates into an appetite for everything.
The question the policy answers
Not what should we research today, which is a taste question and produces a research programme shaped like whoever asked it.
Which single next experiment most reduces uncertainty in the most valuable region of the knowledge space, per unit spent?
That question has a defensible answer, and the answer can be wrong in a way you can measure later. Taste cannot.
A research team, not a role-play
The standard multi-agent failure is four models writing each other essays under the labels Scientist, Critic, Reviewer, Judge. Nothing is at stake and nothing is decided, because every agent has the same objective and the same powers.
Roles here are separated by what they are trying to minimise and what they are permitted to touch. Disagreement between them is structural rather than stylistic, which is the only kind worth resolving.
| Role | Objective | Powers and limits |
|---|---|---|
| Explorer | Maximise novelty against the current knowledge state | Proposes claims. Cannot measure anything. |
| Mechanist | Maximise causal explanatory power | Must name the intermediate observables, not just the correlation. |
| Experimentalist | Maximise information gain per unit of cost | Owns the budget and the experimental design. |
| Falsifier | Minimise the surviving hypothesis space | Writes the kill condition. Can veto a promotion. |
| Statistician | Detect leakage, selection, and understated uncertainty | Can invalidate any measurement, including a favourable one. |
| Engineer | Minimise the code change needed to make a claim testable | Patches the instrument. Never the thesis. |
| Replicator | Maximise independent reproducibility | Re-runs the finding on a track that was not involved in discovering it. |
A failed experiment is not a deleted branch
The compounding asset is not the winners. It is the accumulated record of what has been ruled out — because that is what stops the system from re-testing the same idea every six months with a newer model and calling it progress.
A null result is stored with the same schema as a positive one, and it is queryable by the policy before an experiment is proposed.
claim mutation population result why_failed what_was_ruled_out new_questions reusable_evidence
Every strategy has a lineage
A strategy is a contract that mutates under selection, and each mutation is recorded with the evidence that justified it. Not commit archaeology — a tree of causes.
Sleeve A · v17 │ ├── mutation M-183 exit-geometry class ├── evidence E-771 matched control, pre-declared └── outcome replicated on an uninvolved track │ Sleeve A · v18
A year later the system can answer why does this strategy look the way it does with a chain of experiments rather than a shrug.
When the system cannot know something yet
This is the part that matters most, and it is the opposite of what "self-improving AI" usually means.
Athanor does not modify production code because a model felt inspired. The engineering role is deliberately off to the side of the loop, and it is invoked by one specific event: an experiment the policy wants to run is epistemically blocked — the measurement it needs does not exist in the record.
blocked → instrumentation change → tests → shadow deploy → data accumulates → the experiment becomes executable
The trading logic is untouched by this. What changes is the sensor.
Not "the AI rewrote itself." The AI found a limit on its own ability to establish truth, and modified its scientific instrument to remove that limit.
Frontier models are priors, not oracles
Local experimental data is expensive and scarce; a frontier model has read almost everything and is a very good guess generator. So it is used the way zero-shot models are used in low-data science: to supply a prior, which is then collided with measurement.
The model proposes a claim and a confidence. Mercury compiles it into something measurable, finds the historical population, builds a matched control, updates the posterior, and demands forward replication. The model's opinion never becomes a position on its own.
The consequence is the one that matters commercially: a stronger model plugs in as a better prior generator. It makes the laboratory faster. It does not make the previous architecture obsolete, and the accumulated knowledge does not belong to any vendor.
Benchmarking the scientist
Clean data and unbiased benchmarking are treated as primary research, not as plumbing — canonical experimental datasets, and a benchmark suite that measures the research agents themselves.
That makes a second question askable, and it is the more interesting one:
Which version of the autonomous scientist extracts replicable new knowledge more efficiently?
At that point the system is no longer improving its strategies. It is improving its own scientific method, and the strategies are a by-product. That is the level this is aimed at.
Where this sits
Athanor is the next iteration. The engine that runs today already produces the substrate it needs — deterministic execution, causal contracts with explicit invalidation, full decision traces, and pre-registered out-of-sample testing across parallel deployments. Athanor closes the loop over it and takes the human out of the middle.
The record it has to beat is already on the table.
Next step
If you build closed-loop research systems, allocate to systematic strategies, or want to argue with the design — [email protected]. Objections are more useful to me than agreement.
All trading involves risk of loss. Any allocation discussion is subject to mandate design, operational review, and live risk controls.