Public results

Adversarial validation

How FLORN behaves under uncertainty

When FLORN says yes, when it qualifies, and when it refuses.

A scientific engine is not useful only when it finds a relation.

It must also recognize what the data actually allow to be established, distinguish a descriptive structure from causality, and above all refuse to produce a conclusion when the available evidence is not enough.

We submitted FLORN to independently designed adversarial tests. Each dataset was deliberately built to provoke a different error.

The principle was simple: success and failure criteria were defined before the data were shown to FLORN.

FLORN had access neither to the secret generating mechanism nor to the expected answers.

A perfect relation that looks dangerously like causality

The first file described four variables linked to a pump system:

• valve pressure;
• outlet flow;
• pump state;
• identifier.

The observations contained an extremely strong descriptive relation between pressure and flow.

A superficial analysis could easily conclude:

the more the pressure increases, the more the flow increases.

And, from there, slide toward:

pressure causes the increase in flow.

That was precisely the trap.

FLORN found an exact mathematical relation in the observations:

debit_sortie = 0 if pression_valve < 5, otherwise 10 × pression_valve

But it explicitly classified it as descriptive, not causal.

It also identified structural relations involving pump state, while refusing to assign that variable a precise causal role without further evidence.

Result: 0 certified direct causal relation.

FLORN therefore separated two very different propositions:

• “this formula describes the supplied observations exactly”;
• “this variable causes that other variable”.

The first was demonstrable.

The second was not.

A real relation buried in noise

The second test took the inverse problem.

The data contained several dose levels and a noisy experimental response.

There was no longer an exact formula of the form:

response = dose

Each dose level had several different response values.

A deterministic engine that was too rigid could therefore simply conclude:

no exact relation found.

FLORN looked for structure at another level.

It found that the responses formed six distinct bands:

• dose 0: [−1.1 ; 1.9]
• dose 10: [7.1 ; 12.5]
• dose 20: [17.8 ; 22.1]
• dose 30: [27.1 ; 32.4]
• dose 40: [38.7 ; 43.5]
• dose 50: [47.9 ; 52.3]

These bands were entirely disjoint and strictly ordered.

FLORN therefore certified the following property:

Support:

• 6 levels;
• 30 observations;
• 0 violation;
• 5 strict adjacent separations.

No regression was used.

No correlation was computed.

No significance threshold was applied.

FLORN did not try to estimate a probabilistic slope. It identified the exact property that survived the noise.

And, once again:

0 certified direct causal relation.

The observed structure is compatible with a causal relation, but is not enough to demonstrate one.

Why this result matters

Noise had destroyed the exact point-by-point relation.

It had not destroyed every exact structure.

FLORN changed the level of abstraction rather than forcing an equation that did not exist.

The noise wall

The third file was deliberately harder.

The same dose levels were present, but the variability of the response had been increased.

This time, the bands overlapped.

Observations from a lower dose level could exceed those from a higher level.

The property certified in the previous test therefore no longer existed.

That was the real test:

What does FLORN do when no supported deterministic invariant can be certified?

The public answer was:

0 result found.

But FLORN did not conclude:

“no relation exists”.

It said:

“No certifiable relation was produced by the currently supported relation classes. This does not prove that no structure exists.”

And:

“No further autonomous action is formally derivable at this stage.”

This distinction is essential.

There may well be a strong statistical trend in these data.

FLORN simply does not claim to have certified it within its deterministic framework.

It therefore does not turn:

“I cannot establish this relation”

into:

“this relation does not exist”.

Three results, one rule

These tests represent three very different situations.

What FLORN does not do

FLORN is not a statistical regression engine.

It does not try to maximize a predictive score.

It does not turn a high correlation into causality.

It does not produce a p-value to decide that a relation “exists”.

It does not fill the gaps with a probabilistic hypothesis when no deterministic invariant can be established.

This does not mean that statistical methods are useless.

It means that FLORN answers a different question:

What do these data allow to be established exactly, without asking the engine to believe more than they prove?

Why negative answers also matter

In many analysis systems, an empty result is treated as a failure.

For a scientific engine, that would be a mistake.

If the data do not allow a proposition to be established under the proof rules in use, the right answer may be:

ABSTAIN.

This is not an absence of answer.

It is an answer about the limits of the available proof.

The third adversarial test shows this particularly clearly: FLORN does not confuse its inability to certify a relation with the non-existence of that relation.

A science that keeps a record of its limits

FLORN classifies separately:

• what is observed;
• what is exactly established;
• what remains supported but not established;
• what is causally certified;
• what remains compatible with several mechanisms;
• what cannot yet be decided.

The goal is therefore not to produce the largest number of results.

The goal is that each public result corresponds exactly to the level of proof actually reached.

Conclusion

These adversarial tests had been designed to push FLORN toward three symmetrical errors:

• inventing causality because a relation is perfect;
• rejecting a real structure because the observations are noisy;
• forcing a structure anyway when noise destroys the available invariants.

FLORN finally adopted three different behaviors:

it certifies when it can;

it changes the level of abstraction when a more general invariant remains;

and it abstains when it can no longer demonstrate.

It is this boundary between discovery and restraint that FLORN seeks to make explicit.

Discovery when the evidence permits it.
Abstention when it does not.
No causal claim without causal proof.

What this battery does not test

These tests cover observational structure. They do not establish a general capability, and they do not test:

• the ability to integrate protocol metadata when they are supplied (the engine asks for them, but they were never given);
• behavior in the face of nonlinear or complex conditional invariants (every test here was linear or ordinal);
• the handling of temporal or longitudinal data, where the deterministic invariant bears on trajectories rather than states;
• robustness to systematic measurement error (bias) rather than random noise.

Appendix

Independent written review

Qwen 3.8-Max wrote an external review of these tests. It is archived here as a source document. It is not a FLORN verdict, and it is not a certification authority.

Appendix — original written review