Orbit 00 · The Core · 관문

SENTINELBRAIN

Other AI remembers your code. This one remembers being wrong.

The house's internal engine: findings, fixes, dissents, and graded predictions, carried across every session. Orbit 00.

What follows is a real change from June 2026, replayed from its ledger. At the end, the gate is yours: land it, or kill it. Then see what the builder chose.

● substrate replay warming up…
01 · The finding · 발견

Nobody filed a ticket.
It read the code and raised this itself.

A review specialist working through LuminaNest produced this finding, which could not even be written to memory until it was grounded against the live git tree.

—

severity —
location —
—
raised —
02 · The proposal · 제안

A diff with an argument attached.

Rationale and risk are required fields. A change without a case for itself does not exist in this system.

—

Its rationale
Its own risk assessment
03 · Validation · 검증

Confirmed, — confident.

Typecheck passed. Applied cleanly in a worktree. Then a second model reviewed the claim.

Validator verdict

Also on the record: an earlier validator run on this finding died on its own schema check: logged as unable_to_verify, not deleted. Failures are rows here, not embarrassments.

04 · The dissent · 반론

An adversary was sent to kill it.
This is its best shot.

“Between them, the session could be invalidated (user logs out, token expires, admin revokes access).”
“The handler now has an authenticated id but cannot retrieve the user object it was told to fetch.”
Read the full attack, unedited except code redaction
—
attack result —
adversary confidence —
both sides stored forever
05 · The gate · 관문

No machine lands its own work here.
The gate is human. Right now, it's you.

24proposals died before this moment
—validator, for
—adversary's attack survived

This is a replay: history already happened. Your call gets compared to the builder's, nothing more.

The record · 기록

—

The builder's decision note
—landed commit
—applied in
[sentinel-brain]signed its own work
scoredits prediction, graded by reality
—

Its lifetime prediction accuracy: below a coin flip, published anyway. The architecture exists to move this number with evidence, not adjectives.

The difference · 차이

An advisor would have suggested this and forgotten by tomorrow. You just watched a system raise, argue, verify, attack, land, and remember its own work, and next session it starts from all of it.

The full loop · 순환

You stood at stage four.

↺ stage five is stage one's memory: every task starts from everything before it
The morgue · 부검

The 24 that never reached your gate.

Six shown; all enumerated. Companies publish their wins. The ledger keeps everything.

—rejected
—apply failed
—verification failed
Show all 24
Inside the core · 내부

A memory with its own constitution.

56 stores, one substrate

85 migrations. A restart loses nothing.

29 specialists, 5 managers

Each carries a calibration profile of its own accuracy.

Predictions before outcomes

99 written first, 59 scored by reality.

Trust is earned, never assumed

Zero autonomy promotions have ever fired.

Lessons enter canon by decision

Machine proposes; the builder promotes. Five active.

Claims verified at the boundary

Assertions re-fetched from origin before they leave.

The substrate, drawn · 궤도

Where all of it lives.

warming up the replay…
each spark is a recorded audit event travelling home · names read from the repository

Provenance. Every figure renders from a read-only export of the live substrate, stamped —, each mapped to the query that produced it. The export excludes five sentinel rows that would have inflated the system's age by twenty thousand days. Corrections run in both directions.