The last old Word files
About 100 old-format Word files in your other personal folders, read under their own scope.
Where ByronBrain stands · 24 September 2026
Everything built,
and what's waiting.
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. This page shows everything done so far. Your material itself stays on your computer.
What's been built
Everything in place, grouped by what it does for you. Open any group for the detail.
RulesYour knowledge stays under your control
Written rules for the agent, security and data, and a ranking of whose word counts: yours, then yours recorded, then ideas developed with you, then model proposals, then other authors' words.
- Model output never becomes a standing record on its own. It stays a proposal until you decide.
- Another author's words can never become yours, even when you highlighted them. A highlight shows your attention, not your agreement.
- Code lives in the synced folder; your material lives only in a separate runtime that isn't synced.
- Every reading job runs under a written scope that says what may be read, what may not, and how much.
ReadingYour material, read in place
1,747 of your files are in evidence, with their embedded dates and creators. Nothing was copied or changed.
- Reads Word (new and old), RTF, plain text in any common encoding, Excel, PowerPoint, OpenDocument, HTML and PDF.
- Scanned PDFs are read with the text recognition built into Windows: 1,639 pages across 255 files. Machine-read text is always labelled as machine-read.
- The nine folders of your own writing: 897 files, from journals and correspondence to creative and civic work.
- The Green Wizard project documents: 849 files. Credential files and browser profiles were found and kept out.
- Full inventories of the Green Wizard tree (52,320 files) and your personal folders (13,623 files), taken without opening a single file.
- Every citation can be checked again by re-reading the source, and a changed file is recorded as a new version, never overwritten.
ConversationsWhat you typed to AI assistants
542 sessions captured locally. Your typed turns are separated from everything the tools wrote.
- 1,245 of your turns recorded. Tool output, sub-agents and injected instructions are filtered out, so a model's words are never recorded as yours.
- Each turn is labelled by the project it came from, so practice sessions stay confidential or privileged.
- The claude.ai export is read: your turns, your project instructions as preferences, and the model's memories as the model's proposals.
- A ChatGPT reader is built and waiting for an export.
BooksWhat you mark and what you write back
Kindle notebooks from 15 books are read, with the three kinds kept apart: a highlight, a note, and a bookmark.
- A highlight is the author's sentence and shows only that you marked it. A note is your own writing.
- The local model reads the books for your interests and your convictions. The program checks every citation and tests each finding against the rest of your notes before showing it to you.
- Your rulings changed the rules: one strong source can stand, and statistics aren't findings.
FindingsFindings brought to you, answered on the page
Findings reach this site on their own, and your answers go straight back. No files to move.
- Each finding shows its evidence and the test it had to pass. A finding that failed its test is shown as a conflict, not hidden.
- Only your Google account can read or answer findings. The check is done by the database on every request, not by this page.
- Privileged and restricted evidence never leaves your computer.
- 21 findings answered so far, all collected with no export step.
FeedbackYour comments, read and put back to you
Every comment you leave is kept word for word, then read for what it tells ByronBrain.
- The local model sorts each comment into readings: a belief, an interest, a correction, a qualification, an instruction, a request, or context. Each one must quote what you actually wrote.
- You confirm each reading. Only a confirmed one becomes a standing record, and a request you confirm becomes a task.
- Words from one project are never used to describe another.
RunningIt runs itself
A launcher in the system tray starts at login. It stays out of the way while you work and does the heavy jobs when the computer is idle.
- Syncs findings every 15 minutes.
- Runs the daily job when the computer has been idle: Chief has been closed for 15 minutes, or there's been no keyboard or mouse for an hour, overnight included.
- A Windows notification tells you when more is waiting on you.
- The dashboard shows what needs you, what ByronBrain holds, your projects, and what it's doing. Chief has a link to the findings in its sidebar.
MapsYour work as a picture
A mind map of each project, drawn from its own records and never by hand. Every point on the map cites where it came from.
- Colour shows who does each step: you, a program, or the local model.
- On any step you can say "automate this", "I do this myself", "this is wrong", or leave a note. It's kept the moment you press it, and "automate this" becomes a task.
- An overview puts every project side by side with where the effort went.
SafetyNothing valuable is fragile
Every big reading job is backed up first, and each backup is verified.
- A restore refuses an unverified backup and takes a safety copy first.
- Four local models were compared on a ten-case extraction test. None has been given a permanent job yet.
- 328 automated tests pass.
The roadmap
Useful outcomes in your life and work. The technical work supports them; it isn't the destination.
-
Done
Keep my knowledge under my control
Rules, source-linked records, whose-word-counts ranking, integrity checks, safe refusal, and verified backups.
-
Mostly done
Find what I know without hunting for it
Nearly all of your written material is read in place. Still to do: the remaining old Word files, deciding which files are yours, and checking how well answers stay grounded in their sources.
-
Working on this now
Bring me what matters, and learn from my answers
Findings, feedback readings and process maps are working. Findings still cover your reading only; your conversations aren't grouped by subject yet.
-
Next outcome
Give me answers I can verify
Plain-language answers that show their evidence, flag conflicts, and say "I don't know." The path that finds cited evidence exists; the part that writes the answer doesn't yet.
-
Planned
Help me turn knowledge into good work
Nine writing-style rules you accepted are waiting on comparison tests before they become standing rules. Local models still have to earn each job.
-
Started
A dependable companion across my work
Chief is read locally and has a link to the findings. TrustAdmin coordination and further sources come later, one at a time.
At a glance
Files in evidence1,747Read in place, nothing copied or changed
Your typed turns1,245Separated from everything the tools wrote
Findings answered21All on this site, no export step
Local models4 comparedNone has a permanent job yet
Next, and not yet
What the program does next, in order, and what's deliberately parked.
Test the nine style rules
Compare drafts on examples held back from training before any of the nine rules you accepted becomes standing.
Decide which model earns which job
Expand the extraction test and add the routing, grounding, retrieval, drafting and rule tests before any model gets a permanent job.
Waiting by your choice or on something outside
- The ChatGPT and Amazon reading-history exports, parked on 23 September.
- Ongoing Kindle capture from the Amazon notebook site isn't built. Notebooks you share from the phone are read.
- 119 documents attached in claude.ai projects are listed by name only until you say who wrote them.
Known gaps
- The part that writes an answer, and the test runner that checks answers against held-back cases.
- Grouping your conversations by subject and spotting corrections in them.
- Images saved outside PDFs, 40 scanned pages even OCR can't read, one password-locked PDF, and 10 old Excel and PowerPoint files.
- An inventory of the law-practice archive, the Chief backup, and the software and media folders.
- A restore rehearsed on the live store. It's been proven on test copies only.
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.
- Anything more on this site only through a written cloud decision you approve.