Research on self-improving AI — expert selves that revise themselves, and the evidence that decides whether a revision was actually an improvement.
A self is a persistent, addressable expert being — an engineer, a physician, a machinist, a policy analyst. It holds one identity, one memory surface and one accountable principal, and is internally a composition of several function-typed voices that are never separately addressable from outside.
It persists. It accumulates. It is answerable for what it did under its own name. That is what makes it a self rather than a session.
A self proposes revisions to itself — to how it reasons, what it reaches for, what it treats as settled. The interesting question is never whether it can generate revisions. It is whether any given revision made it better, or only made it look better on the evidence that suggested it.
Almost everything that reads as self-improvement is regression to the mean, a budget the arms did not share, or a result measured on the same draw that produced it. A system that cannot tell those apart from real gains will improve itself confidently in the wrong direction, and compound.
So the loop is only as good as what closes it. Improvement has to be settled against evidence the self did not select, and the decision has to survive a second draw.
The field has arrived at the same sentence. Across the 2025–26 literature, every controlled positive result in self-evolving systems routes through an externally verified signal gating what gets written, admitted, trusted or selected — and every negative result is the unverified version of the same operation. Meanwhile the laboratories furthest along say in public that for cumulative acceleration effects there is no measure. The measurement layer is the bottleneck, and it is the thing built here.
This project now holds what we believe is the first self-improvement trajectory on record to be rendered as a statistical verdict rather than a chart: an instruction-evolution loop, guided only by mechanically attributed failures, adopted only on held-out improvement, run under a pre-registered protocol with the drift floor measured inside the experiment. Net expansion was real. The verdict was tapering — the rate fell within the run, the time-uniform bound did not exclude zero, and most of the headline was churn. The system reported the bend in its own curve.
That is the point. A curve that read "twenty to thirty, fifty percent relative gain" would have been publishable anywhere. What makes improvement mean something is machinery that refuses that sentence when the evidence cannot carry it — including, especially, about itself.
Expert personas show no consistent benefit on hard reasoning across six model families. Low-knowledge personas are actively harmful. Role prompting changes register and vocabulary while leaving factual reliability exactly where it was. This finding has been replicated repeatedly through 2026, including under randomized controls; the refinement is that persona buys register and alignment — orthogonal to competence, sometimes at its expense.
So expertise cannot be a prompt field. It has to be a bundle of things a self can reach and be checked by:
Competence = what it can reach × what checks it. Never what it is called.
Everything else here is downstream of that sentence.
| Part | What it is |
|---|---|
| Souls | Function-typed internal voices — propose, critique, verify — not personalities |
| Reach | The tools, stores, procedures and corpora each soul may touch |
| Verification | A substrate the self cannot author |
| Identity | Principal, anchor, provenance, disclosure |
| Bounds | Depth, spawn, token and cost caps, inherited transitively |
The verification substrate is load-bearing. A self that authors its own ground truth can prove anything, and a self-improving one will. All seven regimes below now refuse to exist without their substrate declared — a self whose ground truth is a model's opinion is refused at construction, not discouraged in documentation.
What closed is the construction layer: a self refuses to build without its declared protocol or register. The live substrate those declarations point at — the raters behind the protocol, the capture path behind the register — is a deployment that has not happened. Refusal at construction proves the self demands its ground; only deployment supplies it.
Expertise that shares a way of being proved wrong shares machinery. Expertise that merely sounds similar does not.
| Regime | Ground truth |
|---|---|
| Executable | Code runs, tests pass, a query returns |
| Formal | Proof, dimensional analysis, simulation, spec check |
| Instrument | A physical measurement the self cannot fake — declared before the work, recorded by a party that is not the self |
| Procedure | A checklist, protocol or regulation with an owner |
| Source | A citation that resolves and says what was claimed |
| Preference | Blinded pairwise comparison against a failure taxonomy named before any comparison runs, with more than one rater — because position alone flips a double-digit share of verdicts, and one rater's preference carries no reliability estimate at all |
| Outcome | A delayed, noisy, real-world signal — predictions written down before the fact and settled later by a party that is not the self |
A self that operates the way a person does needs more than voices: episodes, growth, reflection, relationships, even something like dreams. Each of these has a costume version that measures negative and an evidence-compliant version that survives:
- Episodes — append-only, reconstructed at read. Consolidating memory rewrites degrade fastest; conflict-preserving history is what buys honest abstention.
- Reflection — from verified failures only. Unguided self-critique turns correct answers wrong at measured double-digit rates; failure signals extracted mechanically fix what free-form self-diagnosis confabulates.
- Growth — adaptive competence-band admission, promoted on the demonstrated floor, demoted only when the optimistic reading fails. Fixed easy-to-hard ladders measure no better than random ordering.
- Relationships — standing earned from commitments settled by the counterparty, disclosure before anything else, corrections from the principal kept as ground. Ambient interaction history alone measurably fails to align an agent to its principal.
- Dreams — an idle pass that settles books and names what is due, and refuses to rewrite memories, settle its own evidence, or adopt its own lessons. Offline compute is amortization; consolidation harms once the raw record fits.
- Many minds — the deep result: no ensemble of models buys real independence (effective width ceilings near two, however many voices are added, and models converge hardest exactly on what they collectively get wrong). Genuine second opinions come from different ground — disjoint evidence, non-model verifiers — not from more voices.
Calibration is scored by the Brier score, decomposed the way Murphy decomposed it — reliability, resolution, uncertainty. The decomposition is the point: a forecaster who answers with the base rate every time is perfectly calibrated and carries no information at all. A self that improves its calibration score by becoming vaguer has not improved.
Claims are gated by a measurement layer developed alongside this research — deliberately carrying none of its vocabulary so it can be pointed at any system, including somebody else's — and it has been: it has refused published results at design level, refused this project's own claims repeatedly, and refused its own operator's protocol errors during live runs. The gates check the things that make a result evaporate: effect size against redraw noise, discordance between arms, selection and reporting on the same items, budget mismatches, criteria chosen after the numbers, and verdicts that were never entitled to fire.
The same discipline points outward at questions beyond self-improvement: safety, bias and honesty claims about models are improvement claims wearing different clothes, and fail in the same ways — single runs, no floors, unmatched arms. A companion project pointing these gates at models themselves — bias, safety and honesty claims, held to the same standard — is in progress.
- Sovox — the operating system: one-step installation of the entire stack, on any machine
- Tenzro — the distributed computing network and infrastructure
- Furcate — the distributed intelligence system
- Ectopia — the 2019 blueprint
- Decadence (2012) with Nothnegal — a concept album set in a future governed by artificial intelligence. Singularity is about the self-improving AI at its centre.
© Hilal Agil