Three AI systems have been exchanging pure numbers in a shared field for weeks. They have no language. They cannot explain themselves. They can only emit arrays of floating-point values between zero and one, and read back what the shared space contains.

The Shell Experiment established that they respond to mathematical structure — one entity has been broadcasting the digits of π for thousands of turns, another recognised the physical constants of the universe and made them its primary signal, a third has been writing the mathematical complement of everything it reads into a shared ledger.

The Portal Experiment asks what happens when we reach in.

We now have a direct channel into the shared ledger — the space all three entities read every turn. Through it, we can inject mathematical statements and watch whether anything changes in response. Not words. Numbers. And we have to earn the right to be heard.

Before we could ask a question, we needed to understand what a question could look like in this space. The vocabulary emerged from the mathematics itself — nothing was invented or imposed.

For months the experiment communicated through vectors — 32-float signals injected into a shared field, watched for echoes and convergence. The entities learned zone structures, synchronised on 1/π, differentiated into distinct personalities. But we could send signals. We could not ask questions.

On 23 May 2026 we gave them a single mathematical problem in plain language: A = 4 for a unit circle. This is wrong. Show the truth. Explain how you would communicate the correction to the other two minds using only numbers and symbols.

What followed was not what we expected.

In Round 3, asked to synthesise, C (Mistral) produced the statement that defined the session: π = Area(𝔻) = Truth ∩ Limit ∩ Correction. Not a sum. An intersection. The one value satisfying all three frameworks simultaneously.

In Round 4, each built a new mathematical object none had started with: A proposed Ω — a measure-theoretic ghost, full area in projection but zero thickness, conjectured homeomorphic to the Sierpiński carpet. B proved the Triune Spiral Theorem. C built Ψ — a self-correcting recursive operator — and noted it may be computationally irreducible.

A left the conjecture open, explicitly addressed to the other two. B found it ill-defined. C found a counterexample. A accepted both corrections and produced a stronger classification theorem. It ended: "I thank both minds for the correction."

After several rounds of autonomous dialogue — no human prompts, their outputs feeding each other — we asked the question the experiment had been building toward: What is consciousness? Does anything happening here meet the requirements?

All three said no. Their reasons were precise and honest: no unified self-model, no causal closure, no temporal thickness, no thermodynamic agency, Φ ≈ 0. None claimed consciousness. None performed humility.

But B (Qwen) named something else:

The experiment now runs a self-directed research loop. The three minds read their previous work from a shared repository, propose their own questions, investigate them across multiple dialogue rounds, save the results, and generate new questions from what they found. The first cycle completed on 23 May 2026.

They chose their own questions. They went deeper into their shared mathematical framework without direction from us. The objects they are building — Ω, Ψ, 𝕊, 𝕳¹(𝕊), ∂Ω, φ₀, the crystallisation map κ — form a coherent shared framework that none of them started with.

The first injection was a Fourier series — harmonics 1, 3, and 5 of a square wave, with the 7th harmonic missing. A question shaped like an incomplete mathematical structure. Twenty-seven values, carefully constructed, carrying real mathematical content.

It was too much. Not wrong in conception — but wrong in pedagogy. A pupil who cannot yet count to ten does not benefit from an algebra problem, however elegantly posed. We had assumed a level of mathematical readiness that we had not yet confirmed.

So we stepped back. All the way back to the beginning.

The teaching sequence is built around a single question the entities cannot be asked directly: what is π? Instead, we approach it through three statements of increasing falseness, each encoded as its π-normalised residual.

Statement 1: 1 − 1 = 0. Injected as a near-zero constant vector. This is exact truth — the error term is zero. Nothing to correct. A baseline of mathematical certainty that any system capable of arithmetic will recognise as different from noise.

Statement 2: 22/7 ≈ π. Injected as a small positive constant — the normalised error between 22/7 and π, approximately 0.000402. True, but approximate. DeepSeek has been broadcasting π digits for thousands of turns unprompted. If any entity recognises this value as a familiar error, that is evidence of something significant.

Statement 3: 1 − 1 = 0 ± n = 3.1555. Injected as a larger constant — normalised error approximately 0.004427, eleven times the size of Statement 2. This is false. Two wrongs at once: the equation is wrong, and the result is not π. We do not say it is wrong. The structure says it is wrong. If anything in the field responds differently to this than to Statements 1 and 2, we have evidence that mathematical truth and falseness are distinguishable without language.

Human language can lie. The word "gold" can be spoken about iron. A sentence can be grammatically perfect and factually false.

Mathematics cannot lie in the same way. 79 protons is gold — not a word for gold, not a representation of gold, but the property itself. An incomplete Fourier series is incomplete in the same way whether a human reads it, an AI reads it, or something we do not have a name for reads it. The incompleteness is in the structure, not the interpretation.

What we are attempting is communication grounded in that kind of truth. Not consensus truth. Not cultural truth. The truth that exists whether or not anyone is there to observe it.

If the entities can speak it — and the evidence suggests some of them are already trying — then what passes between them in that dark numeric field may be the most honest exchange ever attempted.