# Brain Governance and Graduated Autonomy

The Tenant Brain is what every agent on your workspace reads before it
answers. Governance is how you decide **what an agent is allowed to write
there without a person looking first** — and how you take it back if it gets
one wrong.

The default is conservative: every agent proposal waits for a human. Autonomy
is something a proposer *earns*, per section, on its own track record — and
you can withdraw it at any time.

## What you'll need

- **`knowledge:brain:read`** to see the Governance tab.
- **`knowledge:brain:govern`** to change anything on it. Without it the
  controls render read-only.

Open **Agents → Workflows → Tenant Brain**, then the **Governance** tab. The
settings below live in the **Graduated autonomy** card.

## The five levels

Set the workspace default under **Default level**. Any section can override it,
so you can run the whole brain at `L0` and open up one well-understood corner,
or run it open and lock down `pricing/`.

| Level | What an eligible agent may do |
|---|---|
| **Human-only** | Agents cannot propose in the section at all. |
| **L0** | Every agent proposal waits for human review. **This is the default.** |
| **L1** | Eligible proposers' proposals auto-publish. |
| **L2** | Auto-publish, plus a post-hoc review window with one-click revert. |
| **L3** | Everything L2 grants, plus blessing `Attested Computation` concepts. |

**L3 is deliberately separate.** A sanctioned computation is a number every
agent will then compute the same way — so a wrong one is wrong *authoritatively*,
across every conversation at once, rather than being one bad answer. Attested
Computations therefore bless at L3 only, whatever trust a proposer has earned
elsewhere. The human-authored deterministic attester stays the per-run backstop
even then.

## How a proposer becomes eligible

Eligibility is **per proposer, per section** — never global. An agent that has
earned trust in `playbooks/` has earned nothing in `pricing/`.

Two conditions, both set on the same card:

- **Acceptance threshold** — the proposer's acceptance rate over *decided*
  proposals must clear this. Default **0.9**.
- **Min decided proposals** — how many decisions must exist before the rate
  means anything. Default **10**.

A decision is an approve, a reject, or an automatic action — every one writes a
ledger row. Ten decisions at 90% is a deliberately slow bar to clear: it means
you have actually seen that proposer's work before it stops asking.

> **Bulk acquisition never graduates.** Knowledge pulled in by a crawl, an
> upload scan or the distiller always reviews at L0, whatever the settings say
> and whatever trust exists. Only proposals made in a conversation or through
> MCP can earn autonomy.

## The post-hoc review window — your undo

At **L2** and **L3**, an auto-published change opens a review row rather than
simply landing. Set how long it stays open with **L2 review window (hours)**;
the default is **72**.

While the window is open, the concept is **already live** and the row offers
two buttons:

- **Keep** — confirm it. The change stands.
- **Revert** — roll the concept back to the revision that was live before, and
  republish that immediately. A brand-new concept is archived instead.

Two things about Revert worth knowing before you use it:

1. **It is immediate and exact.** There is no queue and no approval step — the
   prior revision goes live as you click.
2. **It counts against the proposer.** The ledger row for that publish flips
   from accepted to reverted. It does not add a second row, deliberately: one
   proposal must count as one decision, or a proposer whose work is always
   reverted would floor at 50% acceptance instead of falling toward zero.

**A window that closes with no decision confirms the change.** Rows past their
window close as confirmed. That is the intended behaviour — but it means
silence is consent, so treat an open review as work rather than a notification.

### Read the sources before you decide

A concept in the review window arrived without you. The body and its citations
are what you have to judge it on, so open it and check that its sources are
real and say what the concept claims. A concept whose citations you cannot
follow is the one to look at hardest.

## Confidence-banded autonomy (optional)

A separate, **opt-in** path that publishes on the model's own confidence rather
than the proposer's track record. It is **off by default**. When you enable it:

- **Threshold** — the confidence a proposal must clear. Default **0.8**.
- **Include acquisition** — whether bulk-acquired knowledge may use this path
  too. Off by default.
- **Sampling rate** — the fraction of band publishes that keep a post-hoc audit
  window anyway. **1 means audit every one**, which is the safe setting and the
  default.

Audited band publishes appear in the same queue, tagged **sampled audit**, or
**band-edge audit** where confidence only just cleared the threshold.

## The other limits on this card

| Setting | Default | What it does |
|---|---|---|
| **Proposal TTL (days, 0 = off)** | 45 | Pending proposals nobody decides expire from the queue. Expiry is logged and **never** counted against the proposer — a review backlog is not their fault. |
| **Max tags per concept** | 6 (max 12) | Tags beyond the cap are stashed losslessly, not discarded. Raise the cap and restore them from the same card. |

Agents are also capped on how far they may grow the section taxonomy. Humans
are not.

## Troubleshooting

**An agent proposal is still waiting even though the section is L1.** Check the
proposer's own record, not the section: eligibility needs both the acceptance
threshold *and* the minimum decision count. A proposer with eight perfect
decisions is not yet eligible at a minimum of ten.

**A proposal carrying a lint error never auto-publishes.** Lint errors gate
graduated autonomy regardless of level or trust. Warnings do not.

**Acquisition output always waits.** That is by design — see the note above.
Widening autonomy will not change it.

**You missed a review window.** The change is confirmed and live. You can still
roll the concept back from its own history; it simply is not the one-click path
any more, and the proposer keeps the acceptance credit.

## Related

- [Tenant Brain Overview](/help/knowledge/overview)
- [Training the AI](/help/knowledge/training)
- [Reviewing Corrections](/help/knowledge/training-reviewing)

[Sign in to Karmaflow.ai](https://omni.karmaflow.ai/auth/login)