Neuruh — Personal Intelligence Systems

Jeramie Neural Core, Built Runnable

The compound personal intelligence architecture as working instruments: the architecture ruling (no monolithic model), the layer assignment explorer, the 11-step ingestion pipeline, the hybrid retrieval pipeline with its 11 controls, the context compiler packet builder, the nine-adapter family, the tool layer, the evaluation suites, and the governed continual-learning loop with its rollback.

9 pages — read in full9 layer assignments 11 ingestion steps11 retrieval controls 9 adapters7 tool families, 7 evaluation suites

§2Layer Assignment — Where Does It Live?

The ruling made concrete: each kind of material has exactly one home layer, chosen for a stated reason. Facts stay out of weights (thread decision D-015); reasoning style may enter adapters (D-016); workflows become tools (D-017). Click a material.

Select a material.

§3Memory Model — The Memory Object

Every memory is an object with class, provenance, status, confidence, scope, and a decision receipt. Six classes (the Atlas renders their system diagram); seven statuses from raw to quarantined. An object in quarantine can never train anything (§12).

Status lifecycle

Promotion is a governed act with a receipt, not an automatic graduation. Superseded and rejected objects are preserved as lineage, never deleted into amnesia.

§4Ingestion and Canonicalization — 11 Steps

Run the pipeline. Step 1 is the Archive Refinery rule inherited from thread decision D-019: preserve raw bytes first; never hand-clean the landfill before ingestion.

Idle.

§5Retrieval Architecture

Query to answer-with-provenance in eight stages, with eleven controls governing what may surface. Run it, then toggle the controls.

Idle.

Retrieval controls (§5.1)

§6Context Compilation — The Packet Builder

The context compiler does not dump top-k chunks. It produces a multi-resolution packet: objective, canonical definitions, active constraints, relevant episodes, conflicting evidence, tools, and expansion handles. Toggle the packet fields and watch the packet assemble.

§7Adapters and Training

Nine adapters in the family, one core. SFT teaches approved reasoning and execution patterns — never every assistant response or unreviewed idea. Preference pairs teach selection: chosen against rejected, with reason codes. Click an adapter.

Select an adapter.

SFT example (§7.1)

Preference example (§7.2)

§8Tool Layer — Seven Families

Deterministic processes belong in code and services, not only in neural weights. Click a family.

Select a tool family.

§9Evaluation — Seven Suites

The suites that decide whether an adapter update ships or rolls back. Click a suite.

Select a suite.

§10Controlled Continual Learning

Production learning is gated (thread decision D-018): no direct online weight updates from unreviewed interactions. Nine steps from a live interaction to a versioned release, with rollback available at the end. Run the loop.

Idle.

§12-13Security, Governance, and Build Sequence

Governance (§12)Encrypt raw corpus at rest. Separate secrets and personally sensitive material. Least-privilege tool credentials. Consequence-based authorization. Log retrieval and tool calls. Keep model and adapters versioned and reproducible. Maintain source deletion and legal-hold policies. Prevent training on quarantined or uncertain-authorship material.

The 12-step build sequence (numbered 12-23 in the source, continuing its section numbering; renumbered here). Step 5 is the first deliverable that matters: hybrid retrieval with citations, before any adapter is trained.

0 of 12 sequenced (browser checklist only; this is a build specification, not a project tracker).