Persistent observer architecture
Learnable novelty estimates the reusable structure a bounded observer can extract. Fractalish extends the question: what should the observer preserve, how should the encounter alter future accessibility, what contradictions and consequences must remain attached, and which actions remain impermissible regardless of informational value?
Native architecture: The Fractalish spine remains Natural Math → Specificity / Ageometrics → UFWK → Cognitive Basin → Bolt-On and governed host continuity. In neighboring terminology, Cognitive Basin forms part of a persistent-observer architecture.
Claim boundary: This page joins frozen software results, locally validated prototypes, code-present structures, formal specifications, and proposed integrations. Their statuses remain separate. No complete persistent observer operating end to end inside a production language-model service is claimed.
Central distinction: A metric scores an encounter. A persistent observer is changed by it, inspectably, reversibly, and with receipts.
The missing step after learnability
Learnability can say that an observer extracted reusable structure. It does not, by itself, say which target made that structure relevant, what evidence was preserved, what residue remains, what later accessibility should change, or what action remains blocked.
What learnable novelty contributes
From Entropy to Epiplexity separates structure learnable by a bounded observer from surprise that remains unlearnable under that observer and model class. Fractalish records this as neighboring external work, not a locally reproduced result. Claims: CLAIM-0083 and CLAIM-0084.
Intelligence from Learnable Novelty reports a reservoir-based closed-form estimator or approximation of epiplexity using a fixed bounded observer. Co-evolving observers and LLM substrates remain proposed future work. Claims: CLAIM-0085 through CLAIM-0091.
The reported Acrobot result is treated here as evidence that intrinsic learnability is not a substitute for a declared target. MNIST labels were absent from training but used for evaluation.
From the finite observer to the persistent observer
Yanbo Zhang's Age of Subjectivity and the subsequent work with Michael Levin place the finite observer at the center of complexity: structure is complex to the extent that a bounded observer can extract and reuse it. Their learnable-novelty estimator gives that insight a practical computational form.
Fractalish begins from a closely neighboring premise but follows a different systems question. Once an observer extracts structure, what should remain attached to it? Which target made it relevant? What evidence and residue were preserved or lost? How should contradiction alter the record? How should the encounter change future accessibility? And what prevents an intrinsic learning signal from acquiring authority over action?
We therefore treat learnable novelty as a candidate signal inside a larger persistent-observer architecture, not as truth, value, memory, or governance by itself.
Layer distinction: Objectivity and subjectivity belong at different layers. Evidence should be preserved as exactly and inspectably as possible. Interpretation remains observer-, target-, context-, history-, and protocol-relative.
UFWK line: Immutable evidence; revisable interpretation.
The Fractalish architecture
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
Governing frozen integer baseline. Bounded oracle and replay suites passed: 25/25 integer fixtures, 15/15 cluster fixtures, 200/200 replay configurations, and 10/10 bounded replay cases. Claims: CLAIM-0001 through CLAIM-0014.
Specificity has local acceptance and pytest evidence. Construction A+ is software-only descriptor encoding with a preserved small-batch collision: five runs produced two glyph IDs, not collision-free uniqueness. Claims: CLAIM-0015 through CLAIM-0023 and CLAIM-0046 through CLAIM-0051.
Frozen portable sidecar evidence reports replay, host substitution, adversarial rejection, and zero adapter-executed actions under bounded fixtures. Claims: CLAIM-0065 through CLAIM-0070.
The external-host contract reports 73/73 tests, 37 contract requirements, 28/28 rejection cases, host actions=0, and bolt-on actions=0. It is not production integration. Claims: CLAIM-0071 and CLAIM-0072.
The structured WeightReceipt, uncertainty envelope, routing projection, accumulation ledger, and evidence/interpretation split are specifications. Claims: CLAIM-0024 through CLAIM-0031.
Rule 110, NCA, MNIST, and reinforcement-learning findings are recorded as external reported results until reproduced in a declared Fractalish protocol. Claims: CLAIM-0085 through CLAIM-0091.
Eracii Duel is a Natural Math-informed applied-development lane led by Melissa Ellen Clow. It is not part of frozen Natural Math v5 conformance or qualification.
What remains provisional
The proposed integration experiment
PROPOSED INTEGRATION
The learnable-novelty soliton result motivates a future experiment: can transient learnable structures produce persistent, receipt-governed changes in later accessibility?
Run an untouched host, summary-only memory, vector-only memory, recency-only memory, and learnable-novelty-only arms.
Attach a neutral adapter that can observe and normalize events but cannot execute native host actions.
Write immutable evidence receipts, then derive separate interpretation records and WeightReceipts.
Test whether contradiction scars, recovery routes, target contracts, and HOLD change future retrieval usefully.
Move the host boundary and verify evidence identity, replay, and governance behavior.
Reject the stronger claim if simpler memory baselines match performance or if governance cannot prevent intrinsic-drive capture.
Negative results and falsification
Why this work exists
We did not begin with an AI theory and then attach a social vision. We began from the conviction that free knowledge and universal education are load-bearing requirements of a survivable post-labor transition. When we examined the systems that would have to carry that knowledge, we found drift, silent rewrite, broken continuity, and uninspectable authority.
The exact-state, receipt-governed, host-authority-preserving architecture grew from the need to make a future knowledge commons trustworthy.
Knowledge, education, rights, and basic provision must not depend on a machine-generated human-value score.
Wider context: Synaptient, Logientia, Entroresilience, UHI, HALO, and The Great Work remain motivation and civilizational context. The persistent observer does not technically prove those social programs.
Scientific neighbors
Fractalish does not claim absence of prior art. We have not identified another public program combining these exact layers, but that is an audit statement, not a proof of uniqueness.
Collaboration invitation
Useful collaboration includes information theory, reservoir computing, persistent memory, retrieval evaluation, causal inference, event sourcing, formal governance, accessibility review, and adversarial testing.
Evidence and downloads
Claim IDs, statuses, scope limits, evidence IDs, negative evidence, and source identifiers.
Filterable FMB ledger generated from sanitized public-safe data.
Learnable novelty reproduction protocol starts in a not-yet-run state.
Review table binding public status and numerical claims to public evidence IDs.