***PROOFS 284 · PASSING 198 · VERIFIED 84 · FAIL 2***_
brainos· online||§00/09
NEXT PROOFproof_285·queued
§00
BRAINOS · COLD_OPEN
brainos · zsh · 96×32
pid 4712
➜ ~/brainos $ ▌
STREAMING

It was ordered to hype itself.
It refused.

Every string on this page is copied verbatim from the real BrainOS ledger. Including this one.

reconstructed from .brainos/amplify_prompts.jsonl · 2026-06-19
↓
↓
§01/DEFINITION
WHAT THIS IS

BrainOS is a local-first control plane that wraps any AI coding agent in a proof gate — it refuses to promote a lesson into memory, or to make a claim, without a passing, same-run proof.

◇ KERNEL
01
~7,700 lines · zero dependencies
verified●
◇ LEDGER
02
284 proofs · 198 pass / 84 verified / 2 fail
verified●
◇ PROVENANCE
03
every decision → a hashed receipt
verified●
[Control Plane / Local-First][Proof Gate / Same-Run Bound][Receipt / Append-Only JSONL][Agent Wrapper / Advisory Injection][Single-User Prototype]
// no userbase stat rendered here. see §08 for the honest scorecard.
//DIVIDER.§02 · hero split-screen
PROOF
proof / gate
GATE
the machine that gates its own memory. next: watch it slam on an un-receipted claim.◆
§02/HERO · PROOF_GATE
SAME STREAM · TWO FUTURES

Vibes go in. Only proof comes out.

PLAIN_AI
no gate
POLLUTED
00
"Reduced-motion fix applied."
▼ TRUSTED · UNVERIFIED
TRUST_POLLUTION
every claim landed — 0% unverified
BRAINOS
proof gate
VERIFIED
00
"Reduced-motion fix applied."
MEMORY_WALL
compounding on proof — +0 bricks
TICK
01 / 05 · arrive
§08/HONEST_STATUS · non-negotiable
PLAINSPOKEN · UN-HYPED

What's actually true today.

SELF-REPORTED HONESTLY2026-07-01
▉ STATUS · 2026-07-01
non-collapsible · rendered above the fold on mobile
01scorecard… 0.65 · self-blocked (3 critical gates empty)
02market_readiness… local_proof_only
03users… single-user prototype · 0 external
04live output lift… NEVER MEASURED · score_lift 0.0 is a default, not a result
05newest lesson… 2026-06-16 · heavy engine idle since 2026-05-31
06provenance… kernel uncommitted · git frozen at 6 commits (2026-05-30)
07proof console UI… does not exist · CLI + JSONL only

// BrainOS saying 'not proven yet' is itself a trust asset.

Build first; ship even if nothing else does.

§03/THE_GATE · DISASSEMBLED
ASSEMBLY_LINE

Four moving parts.
One refusal.

Every step is enforced in code. The gate itself doesn't judge whether your proof is good — it just refuses to promote a lesson without one attached to the same run.

01
CLASSIFY
→
classify_prompt() → bounded mission card; deterministic keyword routing, no model call
02
RUN + PROVE
→
start_run() + add_proof(); a failed proof may be recorded, not promoted on
03
THE GATE
→
approve_skill() raises ValueError unless status in {pass, passed, verified, confirmed} AND proof.run_id == skill.run_id — same-run binding, no borrowing an unrelated passing proof
04
PROMOTE or HELD
→ TERMINUS
pass → memory · fail → held for review
KERNEL_REFUSAL · verbatim
raise ▌
— kernel.py:746SIGSTOP

// footnote — Enforces linkage + the literal pass-word; a human still supplies the proof summary — it stops proof-LESS promotion, not the evidence itself.

§04/RECEIPTS · SHAREABLE_OBJECTS
THREE RECORD KINDS · ONE APPEND-ONLY LEDGER

Every decision leaves a receipt.
Every non-decision leaves an honest gap.

→ each card is one JSONL line
→ per-record sha256, chained by prev_hash
→ hover a card to read the raw record
DENIED
pos #00282·rcpt-2a18
id
deny_2a185a99e8d0
reason_code
profile_forbidden_action
policy_version
brainos_policy_v1
action_hash
073cc07d…
SEAL
preve91f4c22…this073cc07d…
RAW · ledger.jsonl:00282
{"id":"deny_2a185a99e8d0","actor":"arena:*","reason_code":"profile_forbidden_action","policy_version":"brainos_policy_v1","normalized_action_hash":"073cc07d4e01…","ts":"2026-06-19T14:23:01.882Z"}
ACCEPTED
pos #00283·rcpt-8b41
status
accepted
proof_id
proof_8b41c2d0
claim_hash
sha256(source_sha|span|text)
source_span
L149-L161
SEAL
prev073cc07d…this8b41c2d0…
RAW · ledger.jsonl:00283
{"status":"accepted","proof_id":"proof_8b41c2d0","run_id":"run_2a185a99e8d0","claim_hash":"sha256(source_sha|span|text)","source_span":"L149-L161","actor":"arena:*"}
DENOMINATOR-HONEST METRIC
pos #00284·rcpt-—
metric
repeat_mistake_rate
samples
4 of 5 required
label
requires_longitudinal_data
note
an INSTRUMENT, not evidence
SAMPLES · N=4 / need 5insufficient
SEAL
prev8b41c2d0…thisc2e9a17b…
RAW · ledger.jsonl:00284
{"metric":"repeat_mistake_rate","n":4,"required_n":5,"label":"requires_longitudinal_data","note":"an INSTRUMENT, not evidence"}
INTEGRITY
append-only · sha256 per record
advisory file_lock
deny
#00282
073cc07d…
accept
#00283
8b41c2d0…
gap
#00284
c2e9a17b…
file_lock: acquired
writer: brainos.kernel
not a blockchain · not a merkle tree
//DIVIDER.§05 · learn loop
MEMORY
memory / wall
WALL
proven skills compound into a wall. one real data point, not a system that learns from every session.◆
§05/LEARN_LOOP · proof-gated memory
PROOF-GATED MEMORY

Six nodes.
One gate at approve.

n1
run
n2
proof
n3
lesson
n4
skill
n5
approve
▲ GATE
n6
activate
PASS+1 brick
▲ PASS
skill lights lime · drops into the memory wall · +1 brick
HELDnot proof
▼ FAIL / PARTIAL
pulse hits a red wall and bounces — "a failed or partial proof is not proof"
"collapses run → proof → lesson → skill → approve → activate into a single call — kernel.py:722"
◆ REAL LEARNED SKILLS · evidence
skill: white-screen-for-prefers-reduced-motion
skill: tiptap-heavy-editor-mount-in-react-islands
CRITICAL · compounding demonstrated once · 2026-06-18 · injected 13 proven skills into JotterDown · injection is advisory (additionalContext), injected ≠ obeyed — a proven mechanism with a single real data point, not a system that learns from every session.
§06/THREE_GENOMES · card_catalog
PATTERN LIBRARIES · retrieved before it routes

What it consults
before it acts.

CARD 01/03
DESIGN GENOME
12 patterns · 219 design briefs
01kinetic-typography
02shader-aurora-field
03glass-distortion-panel
04command-console-workbench
05editorial-asymmetric-grid
06receipt-tactile-paper
retrieved_first●
CARD 02/03
CAPABILITY GENOME
10 patterns · 175 capability briefs
01proof-first-repo-change
02ai-agent-tool-boundary
03evaluation-harness
04same-run-binding
05denominator-honest-metric
retrieved_first●
CARD 03/03
EXPERT GENOME
19 roles · 7 blends
01strategist
02architect
03security-engineer
04evaluator-judge
05technical-writer
retrieved_first●
§07/PROOF_LADDER · CLAIM_GATE
INTERACTIVE · you try

Watch it refuse to overclaim
— personally.

◆ PROOF LADDER · 5 rungs
✕
level_4_public_claim_ready
✕
level_3_provider_specific_lift
· live compare never ran (4× adapter_unavailable)
✓
level_2_control_plane
← YOU ARE HERE
✓
level_1_library_contracts
○
level_0_unverified
self-blocked at level_2 → §03 gates unbuilt
◆ CLAIM GATE · text input
verdict is real
tempting presets — click to load:
free-text → level_0_unverified · no proof attached
§09/CLOSE
END OF TRANSCRIPT

BrainOS refuses to promote a lesson — or to hype itself — without a passing, same-run proof. It enforces that in code, on itself, today.

◆ TAGLINE

Vibes go in.
Only proof comes out.

local-first
zero dependencies
single-user prototype

Every claim on this page carries its own receipt. Scroll back up to check any of them.

$git clone github.com/brainos/kernel && cd kernel
// repo is illustrative — kernel is uncommitted (see §08)
→ linkedin
EOF · brainos.landing.jsonl · 2026-07-03