LuciApps / AI & Agents / Conjunction engine
Research system · evaluation

An engine built to kill its own ideas

A search system that composes physics primitives across unrelated domains, then runs every candidate through falsifiability gates before anything is allowed to count. The interesting number is not how many ideas it generated. It is how many it threw away, and whether it can tell you exactly why.

13,882candidate conjunctions evaluated
98.5%killed by a gate
203passed every gate
144primitives across ten domains
The funnel

Where candidates die

Each gate tests something a real proposal has to survive: are the operating regimes even compatible, is it physically realizable, does it conserve what it must. A candidate is rejected at the first gate it fails, and the gate records its reasoning.

The rejections

Read the reasons

A sample of killed candidates with the gate's own explanation, unedited. This is the part that matters: a system that cannot say why it rejected something is not evaluating, it is guessing.

Machine-generated candidates, not claims. A conjunction that passes every gate is a thing worth examining, never a result. The 203 survivors are held back because they are active research, not because they are conclusions.

The gates

What each one actually tests

A candidate is rejected at the first gate it fails, and the gate writes down its reasoning in plain terms. No score, no confidence percentage, no vibe. A sentence naming the quantity that did not work and the number that killed it.

01

Validity

The largest killer by a wide margin. It asks whether the two things can even operate in the same conditions. A process that needs 300 kelvin and one that needs 77 kelvin do not combine, however elegant the idea sounds written down.

  • Temperature, pressure, and scale ranges checked against each other
  • Rejection names the exact quantity and the bound it violated
  • Roughly forty-three percent of everything dies here
02

Sweep axis

Whether there is a shared variable to move at all. If nothing can be turned up or down that affects both halves, the combination is a coincidence rather than a mechanism, and there is no experiment to run.

03

Realizability

Whether anybody could build it. Frequencies eighteen orders of magnitude apart do not share an apparatus, and a design that needs them to is a thought experiment, not a proposal.

04

Conservation, momentum, and stiffness

The physics backstops. Does it conserve what it must, does the momentum budget close, and do the response times differ so wildly that one half cannot couple to the other. These catch the ideas that look plausible right up until somebody writes the balance down.

05

Skepticism

The last filter, and the smallest. Some primitives are flagged as speculative going in, and anything leaning on too many of them at once gets held back regardless of whether it cleared the arithmetic. Passing the math is not the same as being worth somebody's month.

The inputs

144 primitives, ten domains

Each primitive is a physical effect with its real operating envelope attached: the temperatures it works at, the frequencies involved, what it conserves. The search composes them across domains that do not normally talk to each other.

That is the point. Combinations inside a single field have mostly been tried. The unexplored space is between them.

DomainsGravity, electromagnetism, quantum and vacuum effects, fluids, optics, acoustics, plasma, propulsion, energy harvesting, medical.
Combination sizeMostly pairs, with a tail of three, four, and five-way conjunctions.
DeduplicationEvery conjunction is hashed, so re-running updates a record rather than inflating the count.
BookkeepingEach candidate keeps its full gate history, including the ones it passed before failing.
The real argument

Why a 98.5% kill rate is the good news

Any generator can produce ideas. The hard part, and the only part with commercial value, is throwing almost all of them away for stated reasons.

A system that approved most of what it generated would be worthless, because the bottleneck in research was never idea supply. It is the cost of chasing bad ones. The same is true in a business: the expensive mistakes are rarely a shortage of options, they are the option nobody checked before committing three months to it.

So the number worth quoting here is not 203 survivors. It is 13,679 candidates killed, each with a reason attached that a person can read and disagree with. A system that cannot explain its own rejection is not evaluating anything. It is guessing with extra steps.

That is exactly the standard we hold client agent systems to. An agent that says a lead is unqualified, a document is non-compliant, or a transaction is fraudulent has to be able to say why, in terms the person reading it can check. Confidence scores are not reasons.

Agent engineering

This is what our evaluation looks like

The same discipline goes into client agent systems: task suites from your real cases, gates that refuse to pass a regression, and a system that can always explain its own decision.

See the agent practice
hello@luciapps.com