# 8Z–MDL×DCC — bounded public-prior-art search log and comparison note

**Date / cutoff:** 2026-09-03  
**Status:** public working note / correction-open / not a patentability or freedom-to-operate opinion  
**Bound claim:** first publicly documented practical system of the specific six-part kind defined below  
**Human architecture and origin:** Bojan Dobrečevič (BD)  
**Development context:** BD × multiple AI collaborators

## 1. Exact object being searched for

The priority question is not whether compression, MDL, adaptive control, program generation, routing, meta-learning or recursive optimization existed before 8Z–MDL×DCC. They did.

The searched object is a practical architecture containing all six properties together:

1. **Alternative-representation generation** — the system can create or compare different descriptive languages, mathematical generators, operators, models or problem representations.
2. **Full-cost MDL selection** — selection accounts for the complete description, including model, residual, decoder, exceptions and declared execution/integration cost rather than one attractive score.
3. **Dynamic stable-but-reopenable governance** — a controller allocates search resources, holds productive focus, changes scale or method, and can reopen previously weak paths when evidence changes.
4. **Multi-level recursion** — the same grammar operates at more than one scale or level of search.
5. **Governor self-selection** — sensors, control laws, polarities, topologies or governors can compete and lose under the same selection principle.
6. **Measured heterogeneous-domain transfer** — the same architectural grammar is implemented and measured across substantially different problem families rather than asserted by analogy alone.

## 2. Assurance boundary

This note records the bounded candidate set and comparison disclosed by index v0.2.5. It does **not** establish that every paper, patent, repository, private laboratory or unpublished garage project has been searched. Absence from this note is not evidence of absence.

The claim is therefore:

> In the public systems examined through 3 September 2026, no earlier practical system was identified that matched the complete six-part conjunction above.

The claim is open to immediate correction by a dated public source showing the full conjunction before the disclosed 8Z–MDL×DCC lineage.

## 3. Disclosed nearest public neighbours

| System | Representation / generation | Selection / evaluator | Dynamic governance | Recursion / self-improvement | Governor competes | Measured heterogeneous transfer | Why it remains a neighbour rather than an identified full match |
|---|---|---|---|---|---|---|---|
| DreamCoder | Yes: learns reusable program libraries and abstractions | Bayesian/description-length-related program search | Search policy guides synthesis | Wake-sleep learning and library growth | Not identified as the same explicit governor arena | Demonstrated across program-synthesis tasks | Strong program-discovery neighbour; the disclosed architecture does not match the complete stable/reopenable DCC plus governor-self-selection conjunction. |
| PowerPlay | Generates new tasks and solver modifications | Accepts verified progress on invented tasks | Allocates continuing self-improvement work | Explicitly recursive self-improvement | Solver modifications compete, but not identified as the same full-cost MDL/DCC governor selection | Multiple task types inside one self-improvement framework | Very close on open-ended self-improvement; no identified match to the complete cross-domain MDL accounting and DCC foreground-governance stack. |
| AIXI | Universal program/model mixture | Algorithmic probability plus sequential decision theory | Policy selects actions | Universal theoretical recursion | Not a practical governor competition architecture | General theoretical scope | Foundational neighbour, but ideal AIXI is uncomputable and not the disclosed practical cross-domain governor loop. |
| AlphaEvolve | Model-generated code candidates | Automated evaluators and evolutionary selection | Iterative candidate allocation | Improves algorithms and code through repeated evolution | Evaluators/search policy are not disclosed as competing under the same MDL criterion | Demonstrated on multiple hard algorithmic tasks | Major modern neighbour; not identified as the same full-cost MDL, stable/reopenable DCC and explicit governor-self-selection architecture. |
| ADRS | Prompt and solution generators | Evaluator and selector | Iterative systems-technique search | Repeated discovery loop | Not identified as the same self-selecting DCC hierarchy | Cross-domain systems techniques | Close generator/evaluator/selector composition; complete six-part match not identified. |
| vLLM Semantic Router | Candidate routing signals | Shannon-inspired information-theoretic signal extraction plus performance/cost-aware routing | Dynamic inference routing | Configurable routing policies | No identified full recursive governor-selection stack | Inference-routing domain | Close routing neighbour; narrower than the complete cross-domain representation-generation and self-selecting architecture. |

## 4. Public source locators disclosed on the homepage

The six locators below were re-opened on 2026-09-03 and checked for title/abstract-level fit to the comparison. Property-level judgments remain bounded working comparisons, not exhaustive full-text legal or patent analysis.


- DreamCoder: https://arxiv.org/abs/2006.08381
- PowerPlay: https://arxiv.org/abs/1112.5309
- AIXI: https://arxiv.org/abs/cs/0012011
- AlphaEvolve: https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/
- ADRS: https://arxiv.org/abs/2512.14806
- vLLM Semantic Router: https://arxiv.org/abs/2603.04444

## 5. Recommended reproducible rerun protocol

A future independent prior-art rerun should preserve:

- databases searched and access dates;
- exact query strings;
- inclusion and exclusion rules;
- candidate list before ranking;
- full-text checks for each of the six properties;
- patent-family and repository searches;
- publication and repository dates;
- a correction channel for omitted systems;
- reviewer dependency and exposure records.

Suggested query families for that rerun, recorded here as **future protocol**, not as a claim that every query was historically executed:

- `minimum description length adaptive search controller recursive`
- `MDL meta-controller self-selecting governor`
- `compression progress search governance representation discovery`
- `generator evaluator controller architecture cross-domain optimization`
- `adaptive operator selection minimum description length`
- `recursive resource allocation search policy description length`
- `self-improving solver governor selection`
- `multi-level search control reopen stagnation representation`

### 5.1 Additional nearest-neighbour families added after adversarial external review

The following families were identified by external reviewers as important omissions from the first targeted pass. They are now explicitly queued for deeper full-text comparison. Their presence here does **not** mean they have already been ruled in or ruled out as full matches:

- selection hyper-heuristics and Adaptive Operator Selection (AOS);
- Adaptive Large Neighbourhood Search (ALNS) and reactive search;
- algorithm configuration / selection systems such as SMAC, irace and SATzilla;
- quality-diversity archives such as MAP-Elites / novelty search;
- Gödel machine, OOPS and related self-referential improvement systems;
- compression-progress / artificial-curiosity controllers;
- AutoML-Zero, learned optimizers and other meta-optimization systems;
- hierarchical routing / Mixture-of-Experts and adaptive-computation controllers.

The purpose of the next rerun is not to make the six-part conjunction harder to attack by definition. It is to identify which of these neighbours already contain load-bearing parts of the claimed wiring, and to narrow the novelty claim if necessary.

## 6. Defeat condition for the priority claim

The bounded priority claim should be withdrawn or narrowed if a dated earlier public system is found that practically implements all six properties, not merely adjacent terminology.

A partial match should be added to the neighbour matrix and may narrow the claimed novelty to the remaining architectural delta.

## 7. Relationship to evidence of effectiveness

Historical priority and empirical value are separate questions.

Even if 8Z–MDL×DCC is the first publicly documented system of this specific kind, it can still fail empirically against simpler controllers, stronger baselines, held-out tasks or full-cost accounting.

Even if an earlier full match is found, the current implementations and cross-domain evidence may still have independent engineering value.

Homepage: https://www.mdlxdcc.org/
Arena programme: https://www.mdlxdcc.org/crp/MDLxDCC-Arenas.html