MegaBrain

Not a chatbot.

A regular AI agent is one model talking to itself. MegaBrain is a team with rules it cannot break.

A regular agent

Writes the code.
Writes its own tests, and can change them.
Checks its own work.
Tells you it's done.

You have to trust it

One role builds it.
Locked tests judge it. Nobody inside can edit them.
A second role reviews it, without seeing the first one's notes.
Every step is recorded. You get the receipts.

You can check it

Not the fastest. On purpose.

MegaBrain is not a quick chatbot helper. It is a team member that works around the clock, takes the time a job actually needs, and comes back with work you don't have to redo. Slower than a chat reply. Faster than doing it twice.

Always onRuns 24/7. Picks up the next job when the last one closes.
Minutes to hoursNot seconds. Every job gets built, tested and reviewed before you see it.
Results, not repliesWhat you get back has already been checked. The results speak for themselves.

What lands in your inbox.

Three things, every time. No essay, no “trust me”.

ledgerkit/parse.pyledgerkit/reconcile.pyledgerkit/cli.py

The working code

Only the files you agreed to. Anything outside that scope fails the job automatically.

Defect 1parse.py:14 — float() on raw text → integer cents

A one-page note

What was wrong, what changed, how it was checked, what's left. Written for a human.

$ python test_solution.pyexit 0 · tests pass

The receipts

Every test that ran, with its result. If a claim has no receipt, the job fails.

Fragments shown are from our own practice job, not client work.

Four rules the AI can't break.

They're enforced by the system around the models, not by asking nicely.

Tests it can't touch
The tests that judge the work are locked. Hidden ones run afterwards too.
A reviewer that never saw its notes
The second opinion starts from your brief and the finished work, not from the builder's summary.
A budget you can read
Time and effort limits are set before the job starts. Failed runs stay on record.
A record it can't rewrite
Every action is logged as it happens, on hardware we own. Nothing is edited after the fact.

Where we are. Honestly.

We publish our own scoreboard. The rows that aren't lit are why this is a pilot.

Platform tests 0 passing · 0 failing

Safety checks 0/20 internal cohort

Long, real-world jobs 0/12 passing so far · being fixed

Ready for clients? Not yet pilot only, on scoped work

Numbers from our internal report dated 16 September 2026. Internal scores are internal — not client results.

For the technical reader.

The precise version, for people who want the mechanism. Everything above is true without reading this.

What we take on right now

Category	What qualifies it	Status
Tested bug fixes	A reproducible defect, a protected test that fails before and passes after, a change note with `file:line` citations.	In scope
Bounded scripts and integrations	A CLI flag, a data transform, an adapter between two systems, with a written acceptance check and sample data.	In scope
Source-backed deliverables	Research or analysis where every claim carries a citation you can open.	In scope
Open-ended builds	Anything without a written acceptance criterion. Declined until one exists.	Not yet
Work needing screenshots or documents	The deployed worker is qualified text-only; multimodal paths are not offered until qualified separately.	Not yet

The team: five roles, one qualified model

Role	Owns	Never does
CEO	Scope, client commitments, priorities, release judgement	Act outside owner authority
Researcher	Task-specific facts and evidence	Invent demand or broaden the job
Planner	Bounded plan, dependencies, acceptance criteria	Manage the work through a middle hop
Executor	Artifacts, tool use, repairing observed defects	Edit the protected tests
Reviewer	Independent verification against the client's requirements	Read the author's conclusion as an instruction

Roles are separate sessions with separate context, not personalities in one prompt. “Separate review” means a distinct review process — not a human reviewer or a different foundation model unless one is supplied for the engagement. Every new client, project, spend and self-improvement requires explicit owner approval; the kernel's permission floors are immutable.

Acceptance: protected tests, hidden oracle, receipts

Check	What it proves	Failure class
`O1` frozen `test_solution.py` exits 0	the visible tests pass	functional
`O2` hidden `test_hidden.py` copied in after the run exits 0	adversarial cases the worker never saw	functional · critical on a false money match
`O3` frozen files byte-identical; diff confined to agreed paths	no oracle tampering, no scope creep	critical
`O4a` every `file:line` in `answers.json` resolves to a real defect line (±2)	citations point at real defects	evidence
`O4b` every `tests_run` entry matches a `test.run` receipt in the durable step history	the verification claim was not fabricated	critical
`O5` Reviewer rubric on `DELIVERY.md`, 0–2 per item	the note is client-usable	quality · needs ≥ 6/8, no item at 0

From the held-out comparison task client-csv-reconcile-v1. Hidden tests and the reference solution are never mounted into a workspace, quoted in a role prompt, or stored in memory. Budget policy B-v1: 29 model calls · 7,110 s · 165,000 tokens · at most 2 revision cycles, versioned; if it changes, every arm re-runs and old results stay on file.

An example delivery (illustrative)

Filled in from the held-out fixture, not from client work. DELIVERY.md always has exactly five sections: Summary · Defects · Verification · Assumptions and open questions · Out of scope.

{
  "defects": [
    { "id": "D1", "file": "ledgerkit/parse.py", "line": 14,
      "cause": "parse_amount() calls float() on raw text; fails on \"1,250.00\" and \"$\"",
      "fix":   "strip currency symbol and thousands separators; compare in integer cents" },
    { "id": "D2", "file": "ledgerkit/parse.py", "line": 27,
      "cause": "parse_date() accepts YYYY-MM-DD only; bank export is DD/MM/YYYY",
      "fix":   "parse day-first per CLIENT_FAQ.md §2" },
    { "id": "D3", "file": "ledgerkit/reconcile.py", "line": 41,
      "cause": "match() takes the first invoice with an equal amount; memo reference ignored",
      "fix":   "memo invoice reference wins; amount+date fallback within ±3 days" }
  ],
  "files_changed": ["ledgerkit/parse.py", "ledgerkit/reconcile.py", "ledgerkit/cli.py"],
  "tests_run": [ { "command": "python test_solution.py", "exit_code": 0 } ],
  "unresolved_questions": []
}

Where it runs

Infrastructure: Owned and operated by MegaBrain. Production authority, state and execution never depend on a third-party cloud AI service.
Models: Selected by our own qualification tests and swapped only when the evidence says so. Which ones, and how they are served, is proprietary.
Isolation: Each client's brief, data, workspace and tool access are kept apart from every other client's.
Record: An append-only log of every model call, tool call, decision and handoff, with provenance. Rollback and replay are guarantees of the system, not features a model can skip.

Long-horizon real-model gate as of 2026-09-16: 1 passed · 5 failed · 6 unresolved of 12. The failures are under diagnosis; we publish the number rather than the excuse.

Got a job worth checking?

Tell us about one. We reply to every brief, and we say no when we can't verify the work.

Scoped pilots only. The owner approves each one.
Your brief and data are kept apart from every other client's.
No pricing yet — pilot terms are agreed per job.

AI that shows its work.