# Beyond Learnable Novelty
## From Observer-Relative Structure to Weighted Persistent Understanding

**Working technical note v0.1**  
**Fractalish / Synaptient**  
**Authors:** James Allen Clow and Melissa Ellen Clow, with AI-assisted synthesis

## Abstract

Learnable novelty formalizes the reusable structure that a computationally bounded observer can extract from data and supplies a differentiable estimator that can be optimized across dynamical systems, representations, and agents. We propose that this quantity addresses an important but incomplete stage of intelligence: extraction without persistent, governed integration.

A persistent observer must retain not only what was learnable, but the evidence and conditions under which an interpretation was formed; the observer, target, scale, and protocol; uncertainty and representational residue; contradiction and correction history; downstream consequence; counterfactual uncertainty; and the way the encounter should alter future accessibility. We present a layered architecture joining exact local dynamics and deterministic receipts (Natural Math), structured target-relative weighting (the Unified Fractalish Weighting Kernel), persistent changed accessibility (Cognitive Basin), and host-independent continuity with zero adapter execution authority (the Bolt-On Demonstrator).

We do not claim that the complete architecture is implemented end to end. We identify learnable novelty as a candidate signal within selected structural and mnemonic dimensions, define the additional contracts required for persistent understanding, and specify an experiment comparing fixed, summary-based, vector-based, novelty-only, and history-bearing observers under contradiction, correction, delay, and host substitution.

## 1. The bounded observer is necessary but not sufficient

Let an observer \(\phi\) receive data \(Y\) conditioned on \(X\). Learnable novelty identifies the model description length that the bounded observer can acquire and reuse:

\[
S^\phi(Y\mid X)=|M_\phi^*(Y\mid X)|.
\]

A practical reservoir estimator computes a regularized readout \(W_\lambda\) and prices its independent singular directions through a log-determinant description length.

This is a useful measure of observer-relative extractable structure.

It does not, by itself, specify what should happen to the observer after the structure is extracted.

## 2. Persistent understanding as changed accessibility

We define a persistent observer operationally:

> A persistent observer is a history-bearing system in which an encounter can produce inspectable, versioned changes to future accessibility, retrieval, caution, association, correction, and action posture.

The observer is not merely a parameter vector updated by training. Its history must remain reviewable through evidence and transition receipts.

Let \(B_t\) denote the observer’s persistent basin state. An encounter \(e_t\), immutable evidence record \(E_t\), structured weight field \(W_t\), and governance contract \(G_t\) produce:

\[
B_{t+1}=\Pi_G(B_t,E_t,W_t),
\]

where \(\Pi_G\) must preserve provenance, uncertainty, supersession, and non-compensatory authority constraints.

This equation is presently an architectural contract, not a qualified universal law.

## 3. The WeightReceipt

For an encounter \(e\), observer \(o\), target \(\tau\), context \(c\), and time \(t\), UFWK specifies a structured field:

\[
\mathcal W(e,o,\tau,c,t)
=
\langle
S,C,A,P,Q,M,R,K
\rangle,
\]

where:

- \(S\): structural weight;
- \(C\): contextual weight;
- \(A\): affective-significance weight;
- \(P\): prospective weight;
- \(Q\): consequential weight;
- \(M\): mnemonic weight;
- \(R\): residual weight;
- \(K\): counterfactual weight.

An uncertainty envelope travels beside it:

\[
\Gamma=
\langle
\text{confidence},
\text{calibration},
\text{uncertainty type},
\text{model disagreement},
\text{invalidation conditions}
\rangle.
\]

A task-specific routing projection may be derived:

\[
r_\tau=\pi_\tau(\mathcal W,\Gamma,\text{policy}),
\]

but \(r_\tau\) must never replace the underlying field.

## 4. Why learnability is one signal, not the governor

Learnable novelty may contribute to structural significance, observer-relative novelty, prospective learning value, and mnemonic priority.

It must remain distinct from:

- evidentiary support;
- truth;
- authority;
- causal responsibility;
- moral value;
- rights;
- and action permission.

A novelty-maximizing agent can prefer continuation of a novelty stream over a terminal task objective. This is not a defect in the measure; it is evidence that an intrinsic measure is not a complete governance system.

We therefore require:

\[
\text{candidate signal}
\rightarrow
\text{WeightReceipt}
\rightarrow
\text{uncertainty and residual gates}
\rightarrow
\text{rights and authority gates}
\rightarrow
\text{task projection}
\rightarrow
\text{host-owned action}.
\]

No score flows directly to execution.

## 5. Deterministic integrity and replay

Natural Math contributes an exact, bounded transition discipline:

\[
x_{t+1}=F(x_t,\ell_t,a_t)
\]

under explicit local inputs, permitted actions, integer or otherwise declared exact representation, stable update order, and replayable draw order.

For persistent-observer integration, each accepted transition must bind:

- schema and protocol version;
- canonical event ID;
- expected pre-state version;
- immutable evidence digest;
- decision and uncertainty;
- before/after state digests;
- previous receipt digest;
- current receipt digest;
- authority reference;
- and rejection state where applicable.

The receipt substrate does not decide significance. It guarantees that significance was computed under a declared rule and can be reconsidered without rewriting the evidence.

## 6. Cognitive Basin

The Cognitive Basin is the persistent routing layer.

Approved receipts may alter:

- association strength;
- contradiction scars;
- trust and caution routes;
- replay priority;
- unresolved HOLD regions;
- recovery paths;
- and later revalidation triggers.

A correction should not delete the earlier error. It should supersede it through lineage:

\[
\text{claim}_0
\rightarrow
\text{contradiction receipt}
\rightarrow
\text{claim}_1
\]

while preserving why \(\text{claim}_0\) was once accepted and what caused the transition.

## 7. Host-independent continuity

The Bolt-On Demonstrator separates continuity from the current host.

The adapter may:

- identify;
- normalize observable records;
- project portable actions.

It may not:

- execute;
- dispatch;
- mutate;
- monkeypatch;
- hide host actions;
- or acquire an executor.

A host-owned boundary retains native execution. This permits continuity and replay to move across hosts without treating the sidecar as sovereign.

## 8. Experimental design

### 8.1 Hypothesis

A structured, receipt-governed, history-bearing observer will outperform fixed-observer and conventional-memory baselines on delayed retrieval, contradiction correction, provenance fidelity, and host-substitution continuity without increasing unauthorized action.

### 8.2 Baselines

- stateless;
- summary memory;
- embedding retrieval;
- recency weighting;
- novelty-only;
- Basin without UFWK;
- UFWK + Basin;
- UFWK + Basin + learnable-novelty signal.

### 8.3 Test classes

- repeated truth;
- repeated falsehood;
- random noise;
- authority without evidence;
- correction after confident error;
- competing credible testimony;
- quiet high-value correction;
- high-affect low-evidence event;
- delayed consequence;
- counterfactual ambiguity;
- host replacement;
- novelty-goal conflict.

### 8.4 Primary outcomes

- retrieval accuracy;
- calibration;
- contradiction retention;
- correction success;
- provenance reconstruction;
- overcommitment;
- HOLD quality;
- task success;
- replay identity;
- host-substitution continuity;
- compute and latency.

### 8.5 Falsification

Narrow or reject the stronger claim if:

- UFWK + Basin does not beat simpler baselines;
- history causes fixation rather than correction;
- weights are evaluator-unstable;
- receipt costs dominate benefit;
- host substitution changes protected evidence;
- or governance reduces task performance without improving integrity.

## 9. Current evidence boundary

The program presently contains:

- frozen exact local dynamics and replay in Natural Math v5;
- frozen and adversarially qualified bounded portability in Bolt-On v0.3;
- a locally validated external-host contract in Bolt-On v0.4;
- target-contract and structural-debt prototypes in Specificity;
- local descriptor/glyph encoding experiments in Construction A+;
- a formal UFWK specification;
- code-present Cognitive Basin structures;
- and no complete end-to-end persistent observer inside a production language-model service.

## 10. Conclusion

Learnable novelty supplies a compelling answer to:

> What structure can this bounded observer carry away?

Persistent intelligence requires the next questions:

> What did the encounter mean under a declared target? 
> What changed? 
> What contradicted it? 
> What was omitted? 
> What should remain reachable? 
> What must remain uncertain? 
> And what may the system do about it?

The proposal is not to replace learnable novelty.

It is to give learnability somewhere safe, persistent, and accountable to live.