Research

Ideas with their evidence attached.

Six tracks, the public notes on each, and the nested implementation cases that pin those claims to source. Citations stay on the page. Borrowed benchmarks do not.

Tracks

Start from the subject, not a paper pile.

Six research areas, from AI assistance and durable memory to evidence selection and integration. Start with practical applications, then inspect the sources, implementation status, and limitations.

Put these ideas in context: explore ABI, Abbey, WDBX and Quasar, or choose a starting point.

Topics

Precise language. No borrowed benchmarks.

If a number is not produced by a repository test or a documented artifact, it does not appear here.

Memory architecture

Current

Episodic records with WAL, MVCC, causal DAGs, and content addressing. Memory is a substrate, not a chat log with embeddings glued on. Applies to WDBX, ABI.

Retrieval

Partial

WDBX implements exact and layered-HNSW retrieval with a pluggable scoring seam; evidence-weighted retrieval remains unimplemented. Applies to WDBX.

Distributed systems

Experimental

Reference cluster replication and read repair are in source. They do not establish production sharding or hosted authority. Applies to WDBX.

Model orchestration

Partial

Deterministic persona routing, template completion, and source-defined MCP tools. Model execution and quality require separate evidence from the selected runtime path. Applies to ABI.

Privacy and provider boundaries

Partial

Browser workspace documents and vault notes use localStorage. Configured website model requests and Quasar generation have separate provider and data boundaries. Applies to Quesar, Abbey, ABI.

Provenance and trust

Partial

Content addressing, signatures, and causal history expose record commitments and relationships. They do not establish the truth or evidential reliability of a record. Federation evidence is separately authorized. Applies to WDBX, Abbey.

Synchronization

Planned

Synchronization requires separate implementation and signed-device qualification. The browser vault uses localStorage; the native CloudKit availability plugin does not implement vault synchronization. Applies to Quesar, mobile.

Interoperability

Research

Research scope: consistent product, persona, and claim-provenance vocabularies across integrations, with implementation and UI semantics evaluated separately. Applies to Quesar.

Collection

Notes, overviews, and implementation guides.

OVERVIEW · overview · Implemented · SEPTEMBER 2026 · 1 min read

AI & agent behavior

Evaluate assistant workflows with explicit behavior, provider and review boundaries.

OVERVIEW · overview · Implemented · SEPTEMBER 2026 · 1 min read

WDBX memory & retrieval

Build document recall and decision-history prototypes with explicit source ownership.

OVERVIEW · overview · Implemented · SEPTEMBER 2026 · 1 min read

SEA evidence selection

Prototype project recall or document assistance without placing every available record in a prompt.

OVERVIEW · overview · Implemented · SEPTEMBER 2026 · 1 min read

GPU & compute

Assess whether a target Apple workstation can use the optional Metal DOT path.

OVERVIEW · overview · Implemented · SEPTEMBER 2026 · 1 min read

MCP integrations

Evaluate a local integration from an MCP client into the ABI tool catalog.

OVERVIEW · overview · Implemented · SEPTEMBER 2026 · 1 min read

TUI & operator tools

Give developers a local diagnostics and interaction surface during evaluation.

CORE ARCHITECTURE · research note · Implemented · JUNE 2026 · 3 min read

WDBX: A Weighted-Backtrace Memory Store for Traceable Retrieval

Build document recall and decision-history prototypes with explicit source ownership.

RESEARCH · research note · Implemented · JUNE 2026 · 3 min read

Sparse Evidence Attention for Bounded Context Assembly

Prototype project recall or document assistance without placing every available record in a prompt.

CORE ARCHITECTURE · research note · Implemented · MAY 2026 · 2 min read

WDBX Graph Weights for Traceable Neural Retrieval

Build document recall and decision-history prototypes with explicit source ownership.

SAFETY · research note · Proposed · APRIL 2026 · 2 min read

Policy-Locked Tool Use in Multi-Agent Systems

Evaluate assistant workflows with explicit behavior, provider and review boundaries.

ENGINEERING · research note · Proposed · MARCH 2026 · 2 min read

Latency Budgets for Real-Time AI Orchestration

Assess whether a target Apple workstation can use the optional Metal DOT path.

RESEARCH · research note · Proposed · FEBRUARY 2026 · 2 min read

Backtrace Confidence Signals for Hallucination Reduction

Prototype project recall or document assistance without placing every available record in a prompt.

SCALABILITY · research note · Implemented · JANUARY 2026 · 2 min read

Vector Index Maintenance Under Continuous Ingestion

Build document recall and decision-history prototypes with explicit source ownership.

ETHICS & SAFETY · research note · Proposed · DECEMBER 2025 · 2 min read

Human Approval Gates That Operators Actually Use

Evaluate assistant workflows with explicit behavior, provider and review boundaries.

CORE ARCHITECTURE · research note · Proposed · NOVEMBER 2025 · 2 min read

Chunk Provenance in Long-Context Retrieval Systems

Prototype project recall or document assistance without placing every available record in a prompt.

ENGINEERING · research note · Implemented · OCTOBER 2025 · 2 min read

Offline-First AI Workflows for Sensitive Data

Evaluate assistant workflows with explicit behavior, provider and review boundaries.

SAFETY · research note · Proposed · SEPTEMBER 2025 · 2 min read

Prompt Injection Drills for Agentic Systems

Evaluate assistant workflows with explicit behavior, provider and review boundaries.

ROUTING · research note · Implemented · JUNE 2026 · 3 min read

Multi-Persona Routing Under Uncertainty: Policy Weights and Request Classification

Evaluate assistant workflows with explicit behavior, provider and review boundaries.

IMPLEMENTATION GUIDE · implementation guide · Implemented · SEPTEMBER 2026 · 2 min read

Evaluating Metal acceleration and CPU fallback

Assess whether a target Apple workstation can use the optional Metal DOT path.

IMPLEMENTATION GUIDE · implementation guide · Implemented · SEPTEMBER 2026 · 3 min read

Connecting an MCP client to ABI

Evaluate a local integration from an MCP client into the ABI tool catalog.

IMPLEMENTATION GUIDE · implementation guide · Implemented · SEPTEMBER 2026 · 2 min read

Using the local diagnostics dashboard and agent REPL

Give developers a local diagnostics and interaction surface during evaluation.

Implementations

Nested cases. Each one names the source it depends on.

These sit under /research/implementations. They are implementation readings, not extra benchmark claims.

ai · wdbx · sea · gpu · mcp · tui

A six-layer architecture for evidence-aware AI systems

Quesar organizes interface, orchestration, retrieval, acceleration, storage, and audit as distinct layers so that evidence and runtime claims can be traced to the component that actually supports them.

ai · wdbx

A private document pipeline with explicit fallback behavior

The private Quesar application implements bounded ingestion, workspace-scoped extraction and search, and consent-gated model access while making degraded retrieval behavior visible instead of silently changing providers.

ai · wdbx

Mobile Vault persistence across CloudKit and local fallback

The Quesar mobile Vault keeps one CRUD contract while selecting either the user's private CloudKit database or an encrypted local SecureStore fallback and exposing the selected backend to the application.

ai · mcp · tui

Quasar as a bounded repository-generation loop

Quasar couples an Anthropic tool loop to a narrowly constrained project filesystem, producing a local application through inspectable read, list, and write operations rather than granting an unconstrained shell.

ai · wdbx · sea · gpu · mcp · tui

Abbey capability claims as an executable ledger

Abbey centralizes capability statements in a typed registry whose status, evidence, dependencies, feature flags, and refusal behavior can be projected into documentation and checked for drift.

ai · wdbx · sea · gpu · mcp · tui

WDBX Specimen separates architecture, runtime profiles, and proof

WDBX Specimen combines a provenance-aware architecture specification with explicit browser and native runtime profiles, then uses conformance tests to make a bounded subset of the design executable.

ai · wdbx

Deterministic research exports with citation gates

The Quesar research export pipeline converts validated publication records into stable JSON and static HTML, verifies and copies declared attachments, and records content digests while enforcing unique identifiers, public source URLs, and canonical links.

Sources

Setup and claims, on this site.