Build notes

Aura Knowledge Roadmap

A product roadmap for turning Aura Knowledge into an agent-auditable publishing system, informed by runnable examples in Shubhamsaboo/awesome-llm-apps.

This page is for people inspecting how the garden is being built. It is not part of the main reading path.

Research report Source coverage Source repo Machine JSON
Reviewed 2026-06-17

Shubhamsaboo/awesome-llm-apps at ca3a3d3

P0 7

Foundation ideas

P1 7

Near term ideas

P2 2

Later ideas

Synthesis

Strengthen the research object before adding runtime agency.

The knowledge garden should evolve from static article plus agent packet into a validated research object pipeline: source intake, claim verification, deterministic artifacts, reviewer gates, and optional agent-native authoring tools.

  • Keep the public site static, readable, and inspectable; move runtime agent features into local authoring or separate services.
  • Use deterministic validation for schemas, paths, hashes, generated files, and graph integrity before adding LLM judgment.
  • Treat source-ledger entries as the only path from external material to public claims.
  • Make every agent-facing compression or UI surface a derived artifact with a canonical JSON source.

Sequence

Four phases, each with a concrete output.

01

next foundation pass

Make Evidence Packets Real

Every public claim has structured support, counterevidence, source quality, and validation rules that fail weak packets before publish.

  1. Claim Evidence Packet v2
  2. Evidence CI And RAG Diagnostics
  3. Claim Verification Workbench
  4. UX Governance Checks
02

near-term author workflow

Build The Authoring Workspace

Research plans, source captures, article drafts, artifact diffs, and reviewer results become durable files rather than disappearing in chat history.

  1. Research Workspace Builder
  2. Artifact Widget Catalog
  3. Trust-Gated Publishing Pipeline
03

after first public launch

Expand Intake And Retrieval

The garden can scout new sources, import papers and URLs, suggest graph relations, and expose tool-safe retrieval paths for agents.

  1. Always-On Source Ledger Scout
  2. Provenance Graph v2
  3. Source Ledger Importer
  4. Corrective Research Assistant
  5. Garden MCP Router
04

later workflow layer

Optimize Agent Consumption

Agents can consume compact packets, inspect evidence interactively, and regression-test briefs without turning the public site into a runtime app.

  1. Compact Agent Feeds
  2. Claim Evidence Canvas
  3. Agent Brief Evaluation Harness
  4. Agent-Native Packet Inspector

Product ideas

Buildable slices from the research pass.

P0evidence

Claim Evidence Packet v2

Extend claim records beyond source IDs into evidence snippets, source locations, support type, counterevidence, extraction metadata, and reviewer status.

Pattern and first implementation

Retrieval results and citation UI are backed by typed objects that carry source, evidence, confidence, and provenance.

First:Add an evidencePackets array to artifact.json and render one compact evidence card per claim in the article audit section.

P0governance

Evidence CI And RAG Diagnostics

Fail CI for orphan claims, empty evidence, missing counterevidence, stale source dates, low source diversity, and dangling graph edges.

Pattern and first implementation

A small failure taxonomy, relevance grading, and answer evaluation can catch structural problems earlier than prompt tweaks.

First:Add deterministic validators first, then allow optional LLM audit reports as non-blocking artifacts until the taxonomy is stable.

P0evidence

Claim Verification Workbench

Create an author-side workflow that walks each claim through required evidence, source quality, gap notes, counterclaims, and last-checked date.

Pattern and first implementation

Research agents extract claims, define required evidence, rate source quality, and synthesize findings with citations.

First:Add verification fields to claim objects and a report command that lists claims by verified, needs evidence, contested, or stale.

P0interface

UX Governance Checks

Make reading quality part of the publishing gate: mobile flow, keyboard access, source ledger scanability, claim marker clarity, and dark mode.

Pattern and first implementation

UX review can be encoded as reusable criteria for accessibility, hierarchy, interaction, and information architecture.

First:Add a docs checklist and optional Playwright assertions for no horizontal overflow, collapsed mobile audit rail, visible focus, and route availability.

P0authoring

Research Workspace Builder

Create a local authoring mode where research plans, source captures, claim packets, and final article artifacts are durable files in the repo.

Pattern and first implementation

A sidecar workspace captures plans, todos, files, research runs, and synthesis outputs as the agent works.

First:Start with CLI scaffolding for an article workspace before adding any live UI.

P0interface

Artifact Widget Catalog

Define approved static components for ClaimCard, SourceLedger, MaturityBadge, StemGraph, AgentPacketPreview, and RoadmapItem.

Pattern and first implementation

Generative UI works best when agents choose from bounded, typed components rather than arbitrary executable UI.

First:Document component contracts and render the roadmap page from a schema-validated JSON artifact.

P0governance

Trust-Gated Publishing Pipeline

Record agent roles, inputs, output hashes, review status, policy scope, and human approval before an article can move to published.

Pattern and first implementation

Agent participation is gated by trust checks, scoped roles, and hash-chained audit records.

First:Add optional provenance.json per article and make published status require a humanReview block plus generated artifact diff.

P1authoring

Always-On Source Ledger Scout

Monitor papers, product docs, protocols, GitHub repos, and newsletters for new evidence, counterevidence, and article seeds.

Pattern and first implementation

Scheduled agents collect, rank, normalize, and deliver high-signal briefs with dry-run delivery by default.

First:Create a dry-run scout that writes candidate source-ledger entries and article issue drafts without sending or publishing anything.

P1retrieval

Provenance Graph v2

Expand the graph from topic/article/claim/source into concept, entity, method, supports, contests, depends-on, mentions, and derived-from edges.

Pattern and first implementation

Knowledge graph RAG uses entity and relation extraction with multi-hop reasoning traces and source attribution.

First:Add reviewed edge metadata and show why each public graph edge exists.

P1authoring

Source Ledger Importer

Add a source CLI that imports arXiv, DOI, URL, GitHub, and newsletter items into the ledger with normalized metadata and candidate claim links.

Pattern and first implementation

Paper, newsletter, and URL chat examples show how to ingest external corpora for retrieval and summarization.

First:Start with URL and arXiv metadata import, then require manual approval before sources can support claims.

P1evidence

Corrective Research Assistant

Audit drafts for missing sources, weak counterarguments, ungrounded claims, and unclear research questions before publication.

Pattern and first implementation

Agentic RAG can grade relevance, rewrite weak queries, retrieve more context, and identify missing evidence.

First:Add a non-blocking draft audit report that groups suggestions by claim ID and severity.

P1retrieval

Garden MCP Router

Expose safe garden operations as tools: query artifacts, inspect graph slices, validate packets, create scaffolds, update source ledgers, and open PRs.

Pattern and first implementation

A router sends requests to specialist agents with only the MCP tools needed for that task.

First:Define tool contracts and dry-run semantics before adding any write-capable MCP server.

P1optimization

Compact Agent Feeds

Generate derived compact feeds for claims, sources, roadmap items, and graph edges beside canonical JSON and JSONL.

Pattern and first implementation

Compact serializations and context compression reduce redundant structured data while preserving retrieval hooks.

First:Add token estimates to agent feeds and experiment with a compact TSV or TOON-style claims table as a derived artifact.

P1interface

Claim Evidence Canvas

Add an optional evidence view showing claim coverage, source diversity, maturity, counterevidence, and graph health from artifacts.

Pattern and first implementation

Agents populate addressable dashboard state, charts, and panels instead of returning only text.

First:Start as a static summary on the roadmap or graph page, not a runtime dashboard.

P2optimization

Agent Brief Evaluation Harness

Regression-test agent.md and generated packets against expected retrieval queries and citation behavior.

Pattern and first implementation

A multi-agent loop generates scenarios, scores outputs against criteria, diagnoses failures, and keeps only targeted improvements.

First:Create a small static eval set before attempting self-improvement.

P2interface

Agent-Native Packet Inspector

Expose a local or server-side inspector where agents request graph slices, packet summaries, or review reports from garden artifacts.

Pattern and first implementation

MCP app examples link tool calls to sandboxed interactive resources and generated apps.

First:Keep this out of the public GitHub Pages site until the static artifact contract stabilizes.

Review

Cross-agent review treated runtime UI as a later layer.

Three independent research passes reviewed RAG/research memory examples, agent/MCP workflows, and generative UI/token-efficiency examples. The roadmap prioritizes durable evidence and governance before runtime agent interfaces.

  • Codex main agent
  • Ohm RAG and memory reviewer
  • Bacon agent workflow reviewer
  • Gauss generative UI reviewer