Working architecture note | September 30, 2026
We've Been Looking at AI Drift Backwards
Exploratory drift can be useful. Unauthorized persistent-state drift is the governance problem.
WORKING / NON-CANONICAL
This note records a current design pivot, not an established scientific result. CF v0.4 RC1 remains the public canonical candidate. The September 30 v0.12 concordance is a local working synchronization record and does not supersede frozen project artifacts.
The pivot
Formation logic should constrain persistence, not possibility.
Fractalish previously emphasized drift mostly as corruption risk. That framing was too coarse. A system that cannot generate speculative, analogical, recombined, or wrong candidates has lost part of the machinery that makes search and imagination useful.
The new design target is high exploratory drift in proposal space and low unauthorized persistent-state drift in formed state.
Two kinds of drift
Exploratory drift
Candidate generation, speculation, analogy, recombination, wrong turns, novel continuations, and proposal-space variation. It may be useful and is allowed to be wrong.
Unauthorized persistent-state drift
Unsupported conclusions becoming durable, unverified claims entering state, authority increasing silently, corrupted state propagating, or revision occurring without a trace.
The target is not globally low drift. It is broad variation before promotion and strict accounting when a candidate attempts to become durable, authoritative, or action-bearing.
Proposal space and formed state
The membrane is the architecture.
Proposal Space
Permissive, speculative, probabilistic, divergent, creative, associative, and not automatically durable.
Formed State
Durable, governed, inspectable, scoped, receipt-bearing, contradiction-aware, revisable, retractable, and authority-bounded.
The distinctions remain typed rather than flattened into one confidence score:
These states are not asserted to form one total order. Generated is not established. Retrieved is not verified. Inferred is not established. Established is not authorized. Retained is not promoted. Fluency confers no authority.
Imagination and hallucination
The output channel matters.
The mistake was never necessarily that the machine imagined. The mistake was that we gave imagination the same output channel as knowledge and then blamed the imagination.
Imagination is the generation of possible structure. The working architectural thesis uses hallucination more narrowly: imagined, weakly supported, or unsupported structure presented with authority it has not earned.
This does not claim that all behavior currently called hallucination is imagination. It does not claim that Fractalish solves hallucination. It proposes a falsifiable separation between generation and the status, persistence, promotion, and use of generated material.
Current Fractalish v1 working flow
The generator may remain probabilistic. The transition into durable state is where governance becomes strict. Existing training corpora, transformers, embeddings, attention, gradient methods, retrieval, stochastic search, fine-tuning, multimodal learning, and useful model capability remain available.
Natural Math, CF-Logic, CRR, NLM, Basin, and R2R
Natural Math
Deterministic consequential bookkeeping, identity, state, provenance, replay, authority, receipts, and typed transitions within its declared boundaries.
CF-Logic
The warrant and claim-status layer: what the record is entitled to establish, hold, reject, branch, revise, or retract.
CRR
Consequential Retention-Recruitment: prior interaction, retained consequence, later recruitment, and altered later response. Retention is not recruitment.
NLM-BD001
A boundary discriminator, not a final definition of learning. It asks whether any additional NLM office is independently necessary beyond a fixed CRR-positive apparatus.
Cognitive Basin
Accumulated developmental terrain and durable executable history, not a synonym for a vector database.
R2R
Governed research-to-runtime promotion. External systems may contribute candidates without becoming continuity or authority custodians.
Hot assimilation, cold promotion. Receiving information is not establishing it, and establishing it is not authorizing an effect.
NLM Boundary Discriminator 001
Do not define learning first.
Freeze a CRR-positive deterministic base, then vary proposed additional offices independently: selective gate, provenance availability, capability or effect scope, improvement, utility, cost dividend, contradiction and revision capacity, generalization, and novelty relative to prior formation.
Learning requires another independently necessary operator.
Sufficiently scoped CRR is minimum learning.
Learning is an application-level compression over particular CRR behavior.
Status: WORKING / CREATOR CONSULT REQUIRED / NOT FROZEN. No outcome is privileged.
The chain may stop at morphology.
Morphology here means persistent organization of structure, relation, topology, coupling, thresholds, routing, accessibility, constraint arrangement, Basin configuration, or physical or software organization. It does not mean visible shape alone.
If subtraction leaves only this chain, the project stops there. It does not rescue learning, intelligence, cognition, memory, or consciousness by definition.
Actual condition is not the same as certifiable condition.
A causally active retained structure may exist even when its full history cannot be recovered. Missing provenance limits the certificate; it does not prove that the causal state is absent.
Scientific and engineering neighbors keep their authority.
Pedro Domingos's Tensor Logic: The Language of AI is external research proposing a language that joins tensor algebra and logic programming. It is not part of Fractalish and does not validate this work. A possible complementarity is narrower: a computational language may express or execute an inference; the Fractalish stack asks what that inference is entitled to become.
Modern machine learning, logic, databases, provenance systems, truth-maintenance systems, workflow engines, scientific measurement, and domain-native models retain authority in their own domains. Fractalish must earn any additional role through discriminating tests.
Research process
The roster varies, but major development rounds are routinely put before at least 10 independent AI systems as contributors, critics, adversarial reviewers, and implementation partners. Agreement is not the objective; disagreement, counterexamples, boundary failures, and attempts to break the work are preserved as part of the process.
Participation is recorded per round. Multi-model agreement is not evidence, and no system is represented as reviewing every conversational turn.
Try to kill it
What would make this framing retreat?
- Promotion governance adds cost without reducing unauthorized durable claims or effects.
- Proposal and formed-state separation cannot be made operational or audited.
- Native model calibration, provenance, or workflow controls explain the same gains more simply.
- CRR retention and later recruitment cannot be independently localized.
- NLM-BD001 cannot produce a discriminator beyond vocabulary preference.
Current boundary
This note does not claim solved hallucination, consciousness, intelligence, or learning; universal CRR; physical learning in CNTM; compatibility or collaboration with Tensor Logic; unanimous team agreement; or established science. It does not alter frozen Natural Math or R2R baselines.
The source set includes the September 30 CF Spine & Concordance v0.12 WORKING PDF, the CRR founding statement, the NLM comparison workbook, and the NLM team-response corpus. The concordance also names a Fractalish v1 closure TODO that was not present in the searched local corpus; that missing source remains HOLD.