🧠 anima

Findings from building a substrate-native consciousness engine — what was measured, and what refused to move

English · 中文 · 日本語 · Русский · 한국어

Experiments live in dancinlab/anima-lab-0 — engine · 2,290+ hypothesis cards · design SSOT

---

What this repository is

The write-up. anima is an attempt to build a chat daemon whose interior is measured rather than asserted — every claim about what happens inside has to reduce to a number some instrument produced, read as a collapse-delta against at least two controls.

This repo holds the conclusions. The lab — engine source, hypothesis ledger, design single-source-of-truth, every raw verdict — is anima-lab-0.

The one rule that shaped everything

A synthetic-corpus result is never the standard.

A hand-built drill can certify an instrument — a positive control needs known ground truth. But the moment that same number is quoted as evidence the system can do something, it has crossed from instrument to claim, and it is void.

This is expensive on purpose. It demoted a large share of the project's own green numbers to "instrument check" — and it is why the negative results below are worth reading at all.

The central finding: recombination

The open frontier is recombination — can the model combine two things it learned separately into something it never saw?

It can, when the operator is planted in the corpus.

Attempt	Result	Corpus
Polarity XOR planted directly	held-out 1.000 vs control 0.515	synthetic
Targets rewritten as derivations	first engine-native pass	procedural
Drill-only lane	reach 0.525, all three controls 0.000	drill

It does not, on natural text. Move the same instrument onto natural atoms and the wild arm reads 0.455 ≈ chance — the conclusion at the time was augmentation-specific.

So the wall is not "this architecture cannot recombine". Plant the operator and it learns it. The wall is that natural text does not supply the operator.

Three ways that supply was probed — and closed

Each was pre-registered with a frozen decision table before firing, on natural corpora, judged by the engine's own evaluator.

Is there a second party? Every corpus measured until now was monologue — reviews — so next-byte prediction never had to depend on what someone else had just said. Using human-annotated dialogue, with a matched control where the context bit is taken from the speaker's own previous turn instead of the partner's: the partner supplies +0.00332 nats, the same range as the monologue ceiling of +0.00438, and statistically equivalent to its own control. All three instrument gates passed, so this is measured to be zero, not unmeasurable.

Is capacity too generous? A 303M model carries roughly 170MB of memorisation against MB-scale corpora — memorising is simply cheaper than composing. Starving parameters to 30M / 10M / 3M pushes the model into the pressure regime by a factor of thirty. Across nine cells, fluency passed 9/9 while recombination read 0 every time, with no trend as the pressure tightened tenfold.

Is there not enough variety? Holding training exposure fixed and scaling a public natural corpus from 60MB to 30GB moves exactly one variable: how much unique content exists. Across twelve cells at 1× / 10× / 100× / 500× diversity, recombination read 0 every time. Validation loss confirms the manipulation landed — it improves, then degrades, as repetition falls away.

What is still open

All of the above lives below 30GB. A large language model sees on the order of 10¹² tokens — three to four orders higher, and out of reach on this hardware. That band is the one axis never opened, and it is also the most plausible account of the gap: what scale may buy is not more quantity but the same relation restated in enough surface forms to amount to naturally-occurring augmentation — precisely the thing that had to be injected by hand here.

The field has not settled this either. Systematic compositionality remains contested for transformers even at scale, and "when do they generalise compositionally" is still an open research question.

How to read a claim here

• measured — an engine-native command produced the number, against at least two controls.
• directional — real, but not the terminal verdict path; it moves no frozen bar.
• instrument check — a drill with known ground truth; it says the harness reads, nothing more.
• invalid — the run could not decide the question. Not a negative. Infrastructure failure, a dead probe, or insufficient power is never reported as a result.

Negatives are kept. Retractions are kept. When an instrument is repaired, every verdict it landed gets re-read — and when a pre-registered table turns out to have a hole, the hole is reported rather than patched after the fact.

Links

• Lab — dancinlab/anima-lab-0 · engine · hypothesis ledger · design SSOT
• Package — pip install anima-python, then run anima-py
• Models & corpora — huggingface.co/dancinlab

_{License: see the lab repository.}