The Determinism Gap

A system is either deterministic, or it isn't. There is no gray area.

Determinism is not a maturity level. No model earns it with enough parameters, enough training data, or enough compute. It is a property a system has or doesn't. The probabilistic AI industry is asking regulators, boards, and courts to accept a guess as if it were a computed result.

The claim

A result is deterministic only if nothing between input and answer involved sampling from a probability distribution. A generative model samples, every time. That is the architecture, not an implementation detail.

Why scale doesn't fix it

A larger model still samples. A better-trained model still samples. Scale changes how convincing the guess looks. It does not change that it is a guess. You cannot train a coin flip into a computed result.

Common Tricks

How some platforms fake it on paper.

None of these are theoretical. All three are being sold right now as "deterministic AI." None of them are.

The Cached Answer

Ask a question once, and the platform saves the answer. Ask the exact same question again, and it hands back the saved copy, calling that "the same input, the same output." That is a photocopy of a guess, not a computed result. Change one word in the question and the illusion breaks immediately, because a new question has nothing saved to hand back.

The "Temperature Zero" Claim

Setting a model's randomness dial to zero reduces how much its answers vary. It does not remove the variation. The same input can still produce a different output after the model is updated, on different hardware, or simply because it ran alongside different requests in the same batch.

The "Trained to Be Consistent" Claim

Training a model on a fixed set of examples, or distilling its behavior into rules, makes its answers look more alike more often. It does not make the model deterministic. A more convincing guess is still a guess.

The Five Properties

What a probabilistic system can never produce as its own monitor.

Reproducibility

Same question, same answer, every time, forever. Not usually. Not unless something changed. Every time.

Traceability

You can see exactly how it got the answer. Not just that it gave one.

Independence

The judge can't be running on the same guesswork as the thing it's judging. A probabilistic monitor watching a probabilistic system shares its blind spots exactly.

Self-Correction

It knows when it's gotten worse, without a person telling it. A monitor with no fixed baseline has nothing to compare itself to.

Evidentiary Sufficiency

The proof holds up outside the company that made it. A log of what a model said is not a cryptographic proof of how it got there.

Composability

Stack two guessing machines and you get a bigger guess, not a fact. Uncertainty compounds. It does not cancel out.

The Regulatory Collision

The law is catching up to a distinction the industry has been blurring for years.

EU AI Act

Requires traceability of decisions. Traceability requires stored, reproducible evidence.

EU Machinery Regulation

Requires documentation that enables independent verification.

FAA AI Safety Assurance

Requires continuous, verifiable assurance evidence.

FDA 21 CFR Part 11

Requires records that are attributable, contemporaneous, and independently reproducible.

The monitored system can be probabilistic. The monitor cannot.

Ask Any Platform This

A checklist for any vendor claiming "deterministic."

The Sample
  • Is this output one sample drawn from a distribution of possible answers, or the only value the computation could have produced?If a different draw was ever possible, you already have your answer.
  • Run this exact input 100 times — do you get 100 identical answers, or 100 samples that mostly agree?Agreement is not identity. A distribution that agrees with itself is still a distribution.
  • At temperature zero, is there still a softmax, a probability distribution, or a random seed anywhere upstream — including in training?Sampling and computing are mutually exclusive operations. You cannot sample your way into a fixed value.
The Governance
  • What formal specification governs every possible output — and can the model violate it, or is violation structurally impossible?A rule usually followed is a tendency. A rule that cannot be broken is governance.
  • Who owns that specification, and if it changes, does every past answer still hold under the new version?Governance the governed party can silently edit isn't governance. It's a memo.
  • Is "deterministic" describing a governed boundary, or just the most common behavior observed so far?Common is a statistic. Governed is a guarantee. Only one of those is a claim about tomorrow.
The Measurement
  • What continuously measures live output against the expected, computed value — in production, not a one-time benchmark?A system tested once and trusted forever isn't measured. It's assumed.
  • Who runs that measurement — the vendor, or an independent party with nothing to gain by looking away?A vendor measuring its own drift is grading its own homework.
  • Does the system catch its own degradation before a customer does, or after?If a human has to notice first, there is no real-time measurement. There's a postmortem.
The Proof
  • Where is the cryptographic evidence chain for this exact answer, right now — can I hold it, audit it, replay it myself?A determinism claim with no evidence chain in hand is a promise, not a proof.
  • Can that proof be checked without the vendor's cooperation, infrastructure, or model weights?If verifying it requires trusting the party who made the claim, it's not independent. It's notarized by themselves.
  • Does the proof cover this exact input, or a class of similar ones already validated?A proof for "questions like this" is not a proof for this one.
The Math
  • Anywhere in the pipeline — training, inference, tie-breaking, hardware execution — are floating-point operations computed in a different order across runs?Floating-point addition is not associative. Different order, different answer, every time, without exception.
  • If the answer is a sample from a distribution, what operation turns that distribution into one guaranteed value?There isn't one. Picking the most likely sample is still picking a sample.
  • Would the claim survive the literal dictionary definition — same input, same output, by necessity, not by tendency?Almost none do, the moment you hold them to their own word.
The Standing Challenge

We publish our own replay proof. We invite any platform claiming determinism to publish theirs.

Bring your evidence chain. We will bring ours.