docs/PRIOR_ART.md

# Prior-art comparison and novelty boundaries

Search date: 7 October 2026. Primary-source search covered program synthesis and
library learning, abstract interpretation, observable and predictive state models,
functional state reconstruction, coding and distributed repair, generative memory,
continual world-model repair, topology-structured memory, and recursive agents.
This is not an exhaustive patent, citation-graph, or priority search.

A = clearly known; B = related precedent; C = architectural combination;
D? = candidate distinct mechanism, unverified; E = strong established new-class
candidate. No E rating is justified by the current evidence.

| Proposed element | Closest source | Assessment |
|---|---|---|
| Least representation sufficient for future programs | Amato & Scozzari, Observational Completeness on Abstract Interpretation | A. This was rejected as a standalone novelty claim. |
| Recover selected behavior instead of a whole state | Functional observability; Function-Correcting Codes | A. A direct foundational precedent. |
| Generative distributed memory | The Kanerva Machine | B. Neither “generative” nor “distributed” implies novelty. |
| Repair from distributed fragments | Network Coding for Distributed Storage Systems | A/B. Conventional repair and coding bounds apply. |
| Reusable operator/language invention | DreamCoder; Stitch; babble | A/B. A growing executable library is not a new paradigm by itself. |
| Learn useful representations for program synthesis | Learning Abstractions for Program Synthesis | B. Abstract-domain/transformer learning already exists. |
| Repair causal world models by inventing concepts | Dynamic predicate invention (2026) | B. This substantially narrows the world-model novelty claim. |
| Memory Split/Merge/Update with provenance | All-Mem (2026) | B. Topological memory edits and preserved evidence are prior art. |
| Cocycle-based composition | Asynchronous Algorithmic Alignment With Cocycles | A/B. No novelty is assigned to cocycles or associativity constraints. |
| Self-modifying agent search | Darwin Gödel Machine | B. Different from the candidate memory/quotient transaction, but adjacent RSI prior art. |
| Improving exploration through accumulated world/history structure | Dream-RSI (v2, 6 October 2026) | B. The proposed system is not a relabeled history-replay exploration policy. |
| Small finite linear latent dynamics | Koopman invariant subspaces | B. Nonlinear worlds need not admit a finite exact closure. |
| Minimal additional ancestral tether | Conditional information and quotient-space complements | C. Exact FNE contract from known mathematics. |
| CQW joint capability/quotient/tether/repair/offspring admission | Combination of the above | D? for the exact operational contract; independent novelty unverified. |
| Joint repair-aware abstraction selection | Library learning plus future function-relative repair | D?; needs a positive joint-versus-separate optimization experiment. |
| Recursive transport of expanded recovery obligations | Typed compositional semantics and distributed recovery | C/D?; architecture-level proposal, not a new foundational theorem. |

## Designs that were narrowed or discarded

An observable-quotient-only design overlaps directly with observational completeness.
A function-preserving checkpoint-only design overlaps directly with function-correcting
codes. A generative memory graph with Split/Merge and a library learner overlaps
multiple existing architectures. These are not presented as independent inventions.

The remaining candidate is **Counterfactual Quotient Weaving**: selecting and committing
an executable representation extension together with its all-future operational quotient,
its minimum missing ancestral complement, an explicit information-acquisition debt,
a declared damage contract, and a frozen offspring. The scientific test is whether
this joint selection improves durable utility beyond independently optimized learning
and repair using exactly the same conventional tools.

No exact operational match was identified among the reviewed sources. That does not
establish world priority or prove that the construction cannot be reduced to an existing
combination. The implemented conventional pooled learner and quotient compiler match
the principal finite-family outcomes. The code therefore validates a concrete
architecture contract, not its classification as a new RSI paradigm.

Full bibliographic records and source URLs are in `references.json`; the manuscript
contains the citations and analytical comparison. No third-party paper PDFs are bundled.