← Return to entry
Kernel · online
An AI that checksits own work.
BrainOS is a Python kernel that keeps agents honest. It names what counts as proof before the work starts, ties every claim to the source it came from, and refuses to close a task that doesn’t check out. The animation below walks the loop. The real kernel runs locally and is wired into Claude Code through a SessionStart hook.
trace
readyawaiting task
illustrated · proof-gating loop
8
CLI commands
cover the loop
cover the loop
0%
claims closed
without a source
without a source
every
Claude session
hook fires
hook fires
fail-open
errors never block
empty inject, exit 0
empty inject, exit 0
What you just watched.
An illustrated walk through the loop. The real kernel runs as the brainos CLI on your machine; this animation shows the steps a single task takes through it. No claim survives without a source it can point to, and the loop only closes when the proof bar it set up front is met.
01
scope
Name the task and what done looks like.
02
route
Pick where the answer should come from.
03
claim
State exactly what's being asserted.
04
source
Bind the claim to the origin it came from.
05
check
Test the claim against that source.
06
gate
Does it clear the proof bar set in step 1?
07
witness
Stamp a receipt anyone can re-verify.
08
close
Close — only if the work checks out.
What BrainOS does today.
Plain facts, all currently true:
- · A local Python kernel — single-user, runs on my machine; CLI exposed as
brainos(73 commands across 38 groups) - · A learning loop that runs — lessons become skills, and a skill only activates once a proof passes. Confirmed running on the local CLI
- · A design genome — 12 front-end pattern classes that turn a fuzzy visual goal into a recipe with stacks, a proof checklist, and a11y fallbacks
- · A render-verified code library — 16 cutting-edge UI components built under the same proof gate: 9 actually rendered in a browser and confirmed, 7 more reproduced from cited sources with render-proof pending. Every one traces to its source; nothing invented
- · Wired into Claude Code via a global SessionStart hook, installed since late May. The read side — injecting standing rules into a session — is enforced by the harness
- · Per-project context injection: it matches the working directory to a project profile and injects that project’s North Star, inherited rules, and forbidden actions
- · Inherited rules are proof-gated: recorded via
brainos learn, and only kept when a check confirms they helped - · Fail-open contract: any error → emits nothing, exits 0, never blocks a session
- · The loop: run → proof → lesson → skill → approve → activate — six steps, and nothing is promoted unless its proof passes
- · Honest limit: whether an injected rule has changed a real session’s outcome is the open question I’m still proving
