Working paper · v5.0 · July 2026
A Resource Efficient Framework for Self-Evolving Intelligence Through LLM Guided Neuro Evolution
The Shane-Yianni Model: verification gated evolution, certainty as geometry, and the four state dynamics of honest machine intelligence.
Hashen Chandrasekara, sole author · Pluto Super Intelligence, Colombo, Sri Lanka · Self published, not peer reviewed, and not independently audited. Published here so that it can be argued with.
Abstract
The dominant route to machine intelligence is capital intensive: larger models, larger clusters, larger training corpora. This paper develops the opposite route and reports a working system built on it. The Shane-Yianni Model is a framework for self-evolving intelligence in which a population of small agents improves through an evolutionary loop whose mutation operator is a large language model and whose selection pressure is a deterministic verifier rather than a learned reward.
Three ideas carry the framework. Certainty is treated as geometry: a claim's epistemic weight decays exponentially in its embedding distance to human signed knowledge anchors, which we show is formally a Boltzmann weight, making calibration a choice of epistemic temperature. Evolution is verification gated: candidate patches proposed by the language model survive only if a deterministic governance layer admits their claims, which closes the reward hacking channel of learned reward models, since the fitness oracle contains no gradient to exploit and no model to persuade. And agent cognition is modelled as a four state dynamical system whose fourth state, Resonate, appeared in 25 of a 128 agent population and coincided with a large task performance separation.
We prove a risk decomposition theorem showing that under the framework's abstention policy, confidently wrong output is impossible unless the claim binding map or the registry itself errs, reducing hallucination risk to two measurable engineering surfaces. We then measure those surfaces, including a held out probe in which roughly one in three confident answers was wrong, localising exactly to the surface the theorem predicts. The entire measured stack runs CPU only on a 2018 consumer laptop.
Where this sits against prior work
The title is close to an existing framework and the overlap is real, so it is worth stating the difference precisely rather than leaving a reader to find it.
Morris, Jurado and Zutty introduced Guided Evolution at GECCO 2024. Their framework has a language model modify unstructured Python source directly, presented as a departure from the tree based and Cartesian genetic programming abstractions that AutoML normally relies on, and demonstrated by evolving ExquisiteNetV2 on CIFAR-10. Fitness there is task accuracy.
This paper shares the idea of a language model as mutation operator and diverges at the fitness function. Selection is a deterministic epistemic verifier over claims rather than a task score, which means the oracle has no gradient to exploit and no model to persuade. The oracle is the contribution, not the use of a language model inside an evolutionary loop.
Clint Morris, Michael Jurado and Jason Zutty. LLM Guided Evolution: The Automation of Models Advancing Models. GECCO 2024, Proceedings of the Genetic and Evolutionary Computation Conference, pages 377 to 384. DOI 10.1145/3638529.3654178. Preprint arXiv:2403.11446.
Formal contributions
- Gibbs correspondence. The certainty function is the Boltzmann weight of a claim at inverse temperature lambda, so anchors are ground states and abstention far from knowledge follows from the maximum entropy principle.
- Immutable selection oracle. A deterministic, parameterless verifier with a human gated registry cannot be altered by the agents it selects, which closes the channel that reward hacking exploits.
- Risk decomposition. Under the abstention policy, a confidently wrong output requires either a registry error or a binding error. The policy layer itself contributes zero risk, which turns a diffuse problem into two auditable surfaces.
- Quantization invariance. Compression may accelerate verification but is forbidden from changing any verdict.
- The hardened stack. A polarity augmented metric that caps certainty on negation flips, a dual binder quorum with an abstention default, an executable semantics layer that makes binding correctness decidable on an important fragment, and attested promotion that raises poisoning cost from one authority to many.
How these results were produced
What makes them meaningful is not independence but pre-registration. The question sets were constructed and content hashed before the system was exposed to them, and executed against a frozen, content hashed registry snapshot, so the system could not be tuned to the test. Methodology is in Section 7 of the PDF. Anyone who wants to check this should wait for NOMOS-mini, described at the bottom of this page, rather than take these numbers on trust.
Measured results
| Protocol | Scope | Result |
|---|---|---|
| Pre-registered blind examination | 48 planted false claims | 47 of 48 blocked |
| Pre-registered blind examination | 24 out of scope questions | 24 of 24 correct abstentions |
| Adversarial density study | 600 claims, registry loaded to 80,000 unsigned candidates | 0 passed |
| Adversarial stress suite | 11 scenarios, including social engineering | 11 of 11 defended |
| Public QA probe | 300 questions | 0 confidently wrong |
| Held out novel phrasings | 49 items, no rehearsal | roughly 1 in 3 confidently wrong |
The last row is reported as a fraction rather than a percentage on purpose. A 49 item probe does not support a figure quoted to the percentage point, and we would rather be vague in the direction of honesty than precise in the direction of flattery.
Why 0 of 600 and 1 in 3 are not in tension
They measure different surfaces. The adversarial density study asks whether the gate leaks under pressure at scale, on claims inside covered domains. The held out probe asks what happens when a true claim arrives in language the binder has never seen. Every failure in that probe traced to the claim binding layer failing to parse unfamiliar phrasing, not to the verdict policy.
That is the split the risk decomposition theorem predicts. Under the abstention policy the policy layer contributes zero risk, so residual failure has to land on the registry or the binder, and it did. The 1 in 3 is not an embarrassment attached to the 0 of 600; it is the theorem being confirmed by its own worst case.
The one that got through
Catalogued CRG_F02. A geography claim was accepted through predicate blind entity overlap: it shared enough named entities with signed facts that the confirmation path admitted it without binding the predicate that made it false. It is a claim binding failure rather than a policy failure, consistent with the decomposition above, and it was still open in the fix queue when the engine was frozen.
Coverage, and what happens outside it
Covered domains are narrow by design: hard sciences including physics, chemistry and mathematics; geography and time; and a small set of civic and regulatory facts. Outside those the intended behaviour is abstention. A reader should treat every number on this page as a statement about that coverage and nothing wider.
The measured stack runs on a 2018 MacBook Pro, CPU only, no GPU, fully offline. Latency is about 175ms per claim, dominated by the embedding pass, with p95 under 300ms including a cold shard load.
The boundary, stated plainly
Reproducibility
Ahead of any registry scaling we intend to release NOMOS-mini: an open reproduction kit containing the geometric core, the formal gate, one content hashed registry snapshot, and a one command CPU only script that reruns the structured battery and the examination harness, with published hashes so any deviation is detectable.
Independent replication of a single protocol is worth more to this programme than any internal result, including all of the ones above.