Persistent observer architecture
Learnable novelty asks how much reusable structure a bounded observer can extract. Fractalish asks what must happen after the observer carries it away.
Claim boundary: This page joins frozen software results, locally validated prototypes, formal specifications, and proposed integrations. Their statuses remain separate. No complete production-language-model persistent observer is claimed.
Learnable novelty
Persistent understanding
The first can inform the second. It cannot replace it.
Canonical stack
The current Fractalish proposal keeps learnability inside a typed, receipt-governed stack. No arrow runs directly from novelty or adapter projection to native execution.
Natural Math provides exact transition, receipt, replay, lineage, and fail-closed integrity. Bolt-On provides continuity and portable projection across replaceable hosts.
No current release has yet demonstrated the complete Natural Math–UFWK–Cognitive Basin–Bolt-On pipeline operating end to end inside ChatGPT or another production language-model service.
What exists now
Bounded integer qualification, deterministic replay, and frozen fixtures. Construction A+ descriptor and glyph checks are also locally replayed, with the small-batch collision limitation preserved.
Frozen portability substrate with replay, host substitution, adversarial qualification, and zero adapter-executed actions under the bounded fixtures.
Specificity Engine v0.3 and the Bolt-On v0.4 external-host contract have local validation evidence, but neither is presented as complete production-host integration.
The weight schema, uncertainty envelope, and governance contract are public-facing specifications. A complete runtime is not yet established.
The full Natural Math to UFWK to Cognitive Basin to Bolt-On to language-model-host chain is the target integration architecture, not a completed public deployment.
The reported Rule 110, neural cellular automata, MNIST, and reinforcement-learning findings remain external until reproduced inside a declared protocol.
Proposed experiment
Start with the ordinary host and record a no-sidecar control.
Prove install and removal neutrality before any persistent-routing logic is accepted.
Create immutable event receipts, then use candidate learnability as one typed input rather than execution authority.
Issue WeightReceipts, Basin updates, contradiction scars, replay priorities, and HOLD states under a declared target contract.
Run summary-only, vector-only, recency-only, fixed-observer, and learnable-novelty-only arms.
Introduce contradiction, correction, host substitution, and novelty-versus-goal conflict. Verify that HOLD and host-owned execution boundaries prevent intrinsic-drive capture.
Negative results and falsification
Sources and downloads
Readable HTML version of Beyond Learnable Novelty: From Observer-Relative Structure to Weighted Persistent Understanding.
The preregistered learnable-novelty reproduction page starts in a not-yet-run state and keeps controls, seeds, and falsification conditions explicit.
Public-safe bindings explorer with stable FMB and CLAIM identifiers and downloadable sanitized JSON.
UFWK and Bolt-On have their own public pages so the stack is not flattened into one claim.
We are looking for people who want to break this carefully. Information theory, reservoir computing, persistent memory, causal inference, retrieval evaluation, adversarial testing, and bounded governance are especially relevant.