THE METHOD
The six steps
1. Define the system. Write down what the system actually does for money and what governs it — not what the industry calls it. A refinery isn't a set of distillation columns; it's a machine for sorting molecules, paid by the barrel, limited by the fact that it sorts by boiling point. List the inputs, outputs, governing variables, and the single number a customer pays for. Bound the physics, not the product — otherwise a component improvement can masquerade as a system improvement.
2. Map the operating envelope. Build the full state space — temperature, pressure, chemistry, field strength, energy input, timescale — and plot where industry actually operates. The occupied region is almost always a small island in a large, empty map, and the empty map is the opportunity. Mark what has been demonstrated, what is modeled, and what remains unknown. The map must show where the proposed system fails as clearly as where it might operate.
3. Scan adjacent industries. Strip an invention down to its primitive — the mechanism underneath the application — and ask where else that physics shows up. A DNA chip that triggers reactions with concentric electrodes is really "localized electrochemical triggering with a diffusion guard ring," which belongs just as much to a fracturing proppant grain that stores charge. Semiconductor radical-management disciplines inform supercritical-water chemistry; oilfield deployment informs high-pressure experiments. Each transfer needs a reason the mechanism stays valid under the new conditions. The answer is usually in another industry's textbook.
4. Separate physical from artificial limits. Every boundary gets a label: constant of nature, today's materials, cost, or habit and regulation. The Carnot ceiling is physics; separating crude by boiling point is habit. Respect the first; go after the other three. But a limit can't be reclassified as artificial merely because it obstructs the destination — most limits are artificial, and the discipline is proving it.
5. Find the unclaimed zone. Overlay the patent and literature landscape on the state-space map. The prize is where physics permits operation, no industry operates, and no claims exist. Two examples from our own work: closed-loop inductive energy extraction from a confined, non-flowing rotating plasma ring, and water's phase-state cycling used as a computing dimension — both turned out to be open white space. An empty area may signify opportunity, or it may signify that nobody can make the system work; we investigate both explanations.
6. Capture and protect. File broad independent claims on the regime and narrow dependent claims across many engineering designs; keep continuations pending so claims can follow the market. Parameters, calibration and negative know-how stay as trade secrets. Every claim is tagged with its evidence tier, and every disposition goes through a public-versus-secret fork for national-security relevance.
Three levels, one task: the Framework runs this six-step strategic search; the Simic Method organizes the internal work around principle selection, cross-domain investigation, evidence assessment, decisive testing and asset routing; the Discovery System executes campaigns through problem definition, parallel investigation, adversarial elimination, technical verification, physical validation and value capture.
Where the Framework came from
The Framework grew from trying to make interconnected ideas manageable. Our discussions repeatedly crossed from supercritical chemistry into hydrogen, from hydrogen into plasma, from plasma into propulsion and from cryogenics into computation. Each connection generated another architecture. Without a common method, the portfolio could expand faster than our ability to judge it.
The practical problem was distinguishing a transferable mechanism from an attractive analogy. The July methodology also makes a deeper distinction: a principle transfers only if its causal mechanism survives the move. Two systems described as resonant may behave differently because they store and dissipate energy in different ways. A shared word cannot establish a shared opportunity.
The Framework gives every idea the same obligations: define a system, account for its inputs, find its limits, examine prior work and identify a decisive test. It also gives the company a way to remember why an idea was rejected. That memory should improve the next investigation.
The institution we want to build should accumulate judgment. Every useful measurement, failed configuration and corrected assumption should make the next campaign more selective. If the record merely grows longer, we have built an archive. If it changes what we choose to test and how well we predict the result, we are beginning to build a discovery capability.
A normal Tuesday
Today, a normal Tuesday is founder-operated. Milan selects a question, works through research with AI tools, compares alternative explanations and develops an engineering or experiment brief. The developing system is intended to organize those activities into repeatable campaigns with durable records, competing investigators and verification gates.
An illustrative Tuesday might examine whether a hydrogen-cooled test die can maintain a stable operating range during a workload change. One investigation gathers boiling and package data; another examines device behavior; another searches for failure modes and prior art. Their findings become a short list of unresolved dependencies and a proposed test. Specialists must review the parts that exceed the operator’s competence.
The useful output is an improved decision. Agent counts and token throughput are operating inputs. We need to measure whether they produce better decisions per dollar before expanding them.
One campaign, end to end
Here's the designed campaign flow, walked through on a real example — with honest labels, because no campaign has completed the full path and no patents have been filed yet.
Campaign CTMNS-1: a downhole neutron source.
1. Problem definition. Oil and gas wells are measured with neutron logging tools, and many still use americium-beryllium radioactive sources — hard to license, hard to ship, a security concern if lost. The industry wants an alternative.
2. Trigger. A July 2026 paper reported enhanced deuterium–deuterium fusion rates in metal lattices, where the lattice's electrons partially screen the repulsion between nuclei.
3. Primitive. Stripped of its lab context: engineered electron screening lowers the effective barrier to fusion at low energy. That turns it from a temperature problem into a materials-design variable.
4. Envelope and limits. We mapped lattice material, deuterium loading, temperature and drive method. The physical limit is real — reaction rates fall steeply with energy. The artificial limit is the industry assumption that a neutron tool has to be either a radioactive source or an accelerator tube.
5. The unclaimed zone. D–D fusion splits roughly in half between two branches, and one produces a neutron together with a helium-3 nucleus — every neutron from that branch arrives with a helium-3 twin. Helium-3 is among the scarcest materials in industry. So the decision: not to retrofit existing tools, but to design a new tool built to maximize helium-3 co-production alongside neutron output. The honest gate still ahead: per-tool helium-3 quantities are very small, and the campaign has to show whether the twin is worth catching at fleet scale. If the physics closes, the campaign dies fast and cheap — each step has a kill gate.
6. Capture. The result was a five-family IP strategy covering the lattice source, the downhole architecture, helium-3 capture, operating methods and logging interpretation, plus the CTMNS-1 device concept. Filing status is being verified with counsel before this page claims more than "drafted."
The general pattern behind the example: define the target and bound the physics, map the occupied envelope, scan adjacent industries, separate physical from artificial limits, identify the unclaimed zone, draft broad regime claims, sweep prior art, file the provisional, and plan the continuation family as experiments map the zone. Filing establishes a legal event; performance still requires evidence.
What we've killed
We kill claims, never ambitions — and so far the kills are at the claim level, because the lab is pre-first-campaign. The kill function is designed into the Framework: campaigns will die when the unclaimed zone proves occupied or uneconomic, and killing fast is a feature, not a failure. The record so far:
- Gravitational-wave generation. Our SEL-GW-A application originally framed a lab device as a gravitational-wave source. The arithmetic doesn't allow it — every possible source sits behind the coupling constant G/c⁴, and lab-scale generation falls about twenty orders of magnitude short of detectable strain. No engineering closes a gap set by a constant. We killed the generation claims and kept what survives scrutiny: a null-trial, coded-excitation, witness-channel measurement architecture useful in precision metrology today.
- A moving-mirror battery latch. The Quantum Latch Battery's first latch was a conceptual moving mirror. An audit replaced it with a phase-change material latch (VO₂ or GST) that can actually be built and switched. The design died; the program lived.
- The physics minimum as a design target. We once set out to build the cheapest plasma that is physically stable, then dropped the goal — the cheapest stable plasma may be plasma nobody will buy. The target is the profitable operating range: the economic minimum, not the physical one.
- Our own fusion draft, as written. SEL-FUS-001 was pulled back after our audit found a Debye-radius formula error, an internal neutron/tritium contradiction and two drafts that didn't match. That kill produced the lab's most important rule — the validation rule described in the next answer.
These are documented conceptual revisions, and the company is building a formal, dated disposition register showing the evidence, reviewer and consequence of each stop decision.
The validation rule
The rule: a result counts only when someone who didn't produce it can check it, and a claim may never be presented above its evidence tier.
Who checks the machine's work: first, the machine checks itself adversarially — red-team review with kill authority, and models from different vendors cross-examining each other, with disagreement treated as signal, not noise. No single AI vendor is load-bearing, and two models agreeing is not proof. Second, counsel checks patentability — novelty, enablement, freedom to operate — before anything is filed. Third, independent science checks the physics: human experts review the mathematics and engineering — AI output is not scientific acceptance — and each validation relationship and its current status is disclosed. Fourth, the founder signs, but never alone: Milan co-signs dispositions of elephant-class ideas, but he can't validate his own conception.
What counts as proof is a five-step ladder every claim carries publicly: conceptual → science-supported → simulation-supported → bench-supported → field-supported. A claim moves up only with new evidence, and the reproduction has to be independent. Every number in a claim has to trace to a governing variable and a supporting model — the guardrail that came out of SEL-FUS-001. The form of proof depends on the claim: a theorem needs a complete argument under explicit assumptions; an energy device needs a closed balance; a material needs measured properties in the relevant environment.
Honest state of the machine
What exists today: the full architecture is designed and documented — specialized agents and an orchestrator, hard kill gates, outcome thresholds. A working foundation runs on frontier-model access plus our own stack. The output so far: provisional drafts, prior-art and freedom-to-operate reports, technical audits of our own filings, facility designs, valuation workups.
What doesn't exist yet: a fully autonomous pipeline — today the machine is human-directed and AI-executed; Milan runs every campaign manually and no agents run autonomously. Owned compute at scale. A physical laboratory or any independently reproduced experimental result.
In one sentence: the Framework is doctrine and the engine is designed and partly built, but it hasn't yet proven its output survives contact with a laboratory. Published specs describe the build target, not current operations. Closing that gap is what the current raise pays for. The next milestone is a small, measurable campaign with source provenance, specialist review, reproducible outputs and a complete record of cost and decisions.