Working standards, field studies, experimental designs, and results — from the project’s own record. What lands here has already happened, and cites where it happened.
Ratified standards for AI-assisted work, in order of release. Each was distilled from executed history and ratified on the record; full texts ship in the public baseline. Per-standard paper presentations — abstract, method, provenance — land here in the content pass.
| 2026-08-17 | Kernel Migration Procedure | v0.1 | ships with the next release |
| 2026-08-17 | Scholarly Source Convention | v0.1 | full text ↓ |
| 2026-08-17 | Vocabulary Extension Procedure | v0.1 | full text ↓ |
| 2026-08-17 | Onboarding Protocol | v0.3 | full text ↓ |
| 2026-08-17 | Diet Procedure | v0.1 | full text ↓ |
| 2026-08-17 | Ultracode Flight Protocol | v0.2 | full text ↓ |
| 2026-08-17 | Scan Procedure | v0.1 | full text ↓ |
| 2026-08-16 | Agenda Board Format | v0.1 | full text ↓ |
| 2026-08-16 | Fieldnote Format | v0.1 | full text ↓ |
Full texts are the shipped baseline’s files, verbatim and downloadable; a standard ratified after the newest anchored release ships with the next one.
A command is a small markdown file the agent reads — the procedure written down in plain language, shipped in the baseline’s .claude/commands/ folder and stamped like everything else. These are the six from the newest anchored release, verbatim and downloadable: read one to see exactly what a session is told, or drop it into your own instance’s folder and adapt it. The dictionary view — what each does, in one line — is the command library.
| /welcome | the scripted first contact — the standing check, the explanation, the two paths out | welcome.md ↓ |
| /begin | the trigger of the onboarding interview — the protocol is the script, the birth is its product | begin.md ↓ |
| /fractal | the baseline’s own opener — the living specimen of your project’s named command | fractal.md ↓ |
| /close | the close ritual orchestrated — checklist, checkers, ripple, commit and push | close.md ↓ |
| /fieldnote | deterministic capture — one line into the right ledger, then straight back to the work | fieldnote.md ↓ |
| /look | the observation window — boundaries first, read-only always, report and withdraw | look.md ↓ |
Verbatim copies of the shipped baseline’s files; refreshed with each release.
The first project born from FRACTAL’s kernel was birthed and run through Claude Code, then continued its working sessions in the ChatGPT desktop app — zero kernel changes, because the memory lives in the repo. The protocol of record states: “Client-neutrality moved from claim to observed fact.” In the founder’s words: “usually you’re vendor bound to get memory within projects.”
Two dependency-free checkers block red: one verifies the knowledge store’s invariants before every store-touching commit; one reads the corpus’s own prose claims about itself — versions, pointers, counts — against reality at every close. Both have caught real defects, and each catch becomes a mechanical rule. The tools refuse, never degrade. Where the checkers don’t reach, verification stays on humans — the machine-checked claims run in minutes; the rest is reviewed on the record, and drift is measured, not denied.
A fresh clone on a clean machine rebuilds the working system from the bootstrap document alone — verified, byte-parity, three green drills on the record. The record’s own words: “‘Laptop in the ocean’ is demonstrated fact.”
Every release ships as a signed tag with an independent OpenTimestamps receipt committed in the repository itself, so any clone carries the full proof chain; the earliest anchors are Bitcoin-final. This is the one claim a stranger can check without reading anything:
git tag -v beta-0.4 · ots verify Provenance/beta-0.3-public.txt.ots
The path the setup describes has been walked twice under stranger conditions: once when the first child project was born from an anchored release tag, and once in a simulated onboarding of the public mirror itself — a scripted non-programmer ran the quickstart verbatim through to a real birth, both integrity checkers passing, with no outside knowledge needed at any step.
FRACTAL does not abolish context limits — sessions read a compressed spine; the claim is owned, verifiable, vendor-independent memory, never infinite memory. No auto-updates into living instances — by design; sovereignty is the trade, named plainly. No adoption path for existing projects yet — the paved path births new projects. No UI — terminal and files, with one honestly failed experiment on the books. The network layers are unbuilt and trigger-armed, constitutionally barred from binding a build.
Whole episodes from the record, told with their evidence attached: the first child project’s birth from an anchored tag; the vendor-swap trial; the first kernel migration, executed under the child’s own protocol with every gate on the record; the close that survived its own client archiving the chat.
The first written studies land with the content pass; their raw material exists as flight records, protocols, and fieldnote ledgers in the baseline.
The claims worth making are the ones that can fail: whether structured context keeps agents right (a context-fidelity experiment with blinded scoring is running in the first child project now), whether it keeps them cheap (the tokens-on-new-context claim, designed to be measured against a bare-conversation baseline), and which engine at which capacity suffices for which use intensity.
Designs are published here when they freeze; results follow when they exist — including the ones that come back negative.
Nothing to report yet — by honesty, not by accident: the first experiment is still in the field. Results land here as experiments close, whichever way they land.
Editorial rule of this page: everything cites the record; nothing is written here first.