Increase time and attention for what matters
Reduce the drag of finding, reconciling, and re-explaining knowledge so Byron can spend more effort on judgment, relationships, creative work, and virtues.
Development roadmap · September 22, 2026
Local knowledge.
Grounded answers.
ByronBrain is a local-first, evidence-backed knowledge and reasoning system built to return time and attention to high-value work—and, ultimately, to the pursuit of virtues.
The direction
The early goals remain the test for every technical choice: usefulness, trustworthiness, and deliberate expansion.
Evidence before fluency
Answers cite sources, separate evidence from inference, expose conflicts, and abstain when support is inadequate.
Local first, expand deliberately
Legal-practice systems are the first domain. New sources, models, and cloud capabilities are added only through bounded, measured decisions.
What I want ByronBrain to do for me
This roadmap tracks useful outcomes in my life and work. The technical work is supporting evidence—not the destination.
01
Keep my knowledge under my control
My information stays local by default, its source and status remain visible, and the system does not quietly turn guesses into facts. Supporting work: policy gates, source-linked records, authority rules, integrity checks, and explicit abstention.
02
Find what I know without hunting for it
Bring in one clearly approved body of material, keep its context intact, and make the right pieces findable without disturbing the originals. Supporting work: bounded reading, hashing, evidence spans, source-linked claims, and safe failure tests.
03
Give me answers I can verify
Answer useful questions in plain language, show the supporting evidence, flag conflicts and uncertainty, and say “I don't know” when the support is not there. Supporting work: scoped retrieval, citations, authority and uncertainty labels, and fixed accuracy tests.
04
Help me turn knowledge into good work
Use what I know to support decisions, tasks, and writing while preserving my judgment and learning from my review. Supporting work: held-out evaluations, review history, voice comparisons, and workflow-specific model tests.
05
Become a dependable companion across my work
Coordinate with Chief, TrustAdmin, and later parts of my life without losing privacy, provenance, or the boundaries between them. Supporting work: versioned exchanges, one approved source at a time, verified backup and recovery, and a secure review interface.
What exists today
A real foundation is in place; several essential end-to-end capabilities are still deliberately unimplemented.
Trustworthy foundationWorkingCore records, boundaries, search, status checks, and safe abstention
Writing in my voice9 choices recordedStill awaiting comparison tests before they become standing rules
AI helpers4 comparedNone has earned a permanent job yet
Private knowledge onlineNoneThis site shows the roadmap only
Immediate priorities
These are the next concrete proofs that ByronBrain is becoming more useful—not merely more elaborate.
Make one real question work
Take one approved test source all the way to a useful, cited answer—and prove that unsafe or unsupported cases stop.
Decide which AI earns which job
Compare models on the actual kinds of help I want before trusting any of them with extraction, reasoning, retrieval, or drafting.
Make sure nothing valuable is fragile
Prove that the local knowledge store can be checked, backed up, and restored before it contains more of what matters.
Guardrails and later improvements
A capability is not an improvement if it weakens provenance, privacy, authority, or recoverability.
What the system will not do
- Silently promote model output into canonical knowledge.
- Crawl a source tree or mailbox without an approved bounded manifest.
- Upload sensitive content merely because a cloud service is available.
- Present derived material as source evidence.
Improvements that must earn their place
- Semantic retrieval and reranking only if they beat the FTS5 baseline.
- Graph storage only for demonstrated relationship queries.
- Additional models only through workflow-specific evaluation.
- Hosted evidence review only after stronger identity and synchronization controls.