The Trust Dial: Setting How Much Sera Can Promote On Her Own

The Trust Dial: Setting How Much Sera Can Promote On Her Own

Human review is the point of this system. It is also, honestly, where it breaks: when a passing note and a governance decision both sit in the same queue waiting for the same click, people spend their attention where it was not needed and still fall behind. The trust dial is the fix, and it is designed so that it can only ever be careful.

---

Three settings, per record type

Each reviewable record type carries its own dial, stored in Hub Settings under review_dials:

• auto: eligible items promote quickly once their objection window passes.
• tiered: promotion speed is derived from risk. This is the recommended setting for most types.
• always-human: nothing of this type ever promotes on its own. It waits for a person, indefinitely.

Set them per type rather than globally. A Knowledge Base draft and a decision candidate do not deserve the same treatment.

What "tiered" actually does

Tiered promotion uses what the system already knows about each item, and it can only ever make promotion slower or forbidden than the dial alone would allow. It never speeds anything past your setting.

• Low extraction confidence always requires a human. If Sera was unsure when she wrote it, a person reads it. No exceptions and no dial overrides this.
• High significance always requires a human. Above the significance threshold an item waits indefinitely, no matter what the dial says.
• Everything else gets a time-to-live derived from confidence and significance: a fast tier for high-confidence, low-significance items, and a standard tier for the rest.
• Nothing promotes faster than the minimum objection window. There is always a real window in which a person can object.

What can never promote, at any setting

Four categories are hard gates. Auto-promotion does not touch them at any dial value, and no setting exists that would let it:

• Canon Change Requests
• Policies
• Sensitive Review
• The CCOS ledger (the Transparency Ledger)

These are the records that change what the organization is or that carry private material. They wait for human consent, always.

---

How to choose a setting

The dial is a statement about where attention is worth spending, not about how much you trust the AI. Two questions decide it:

1. What would a mistake here cost? If a wrong record of this type would be expensive, embarrassing, or hard to unwind, set always-human.
2. What is eating the review time? If a type is high volume and low stakes, and reviewing it has become rubber-stamping, set tiered. Rubber-stamping is not review, and pretending otherwise is worse than a soft gate with a real objection window.

A reasonable starting shape: always-human for anything governance-adjacent, tiered for memory candidates and knowledge base drafts, and auto only for a type your circle has watched behave well for a while.

Review does not stop mattering

Auto-promotion is a soft gate, not an absence of one. The item is visible, the window is real, and an objection during it stops the promotion. What changes is that a person is no longer required to click on every low-stakes item for the system to keep moving.

And the review you do still do is worth more than before: your edits and rejections feed correction memory, so the same mistake stops recurring rather than being fixed one record at a time.

---

Where to change it

The dials live in Hub Settings, reachable from the Control Center. Like other Hub Settings, a change takes effect within one poll cycle, with no redeploy. Changing a dial is a governance decision worth recording, so treat it as a proposal to the circle rather than a private preference.

Key points

Per-type trust dials live in Hub Settings under review_dials and take one of three values: auto (promote quickly after an objection window), tiered (promotion speed derived from risk, the recommended default), or always-human (nothing promotes, ever). Risk-tiered promotion only ever makes promotion slower or forbidden relative to the dial: low extraction confidence always requires a human, and a significance score at or above the threshold always requires a human, indefinitely, regardless of dial. Nothing promotes faster than the minimum objection window, so there is always time to object. Hard-gate types are never touched by auto-promotion at any setting: Canon Change Requests, Policies, Sensitive Review, and the CCOS ledger always wait for a person. The dial is a statement about attention, not about trust in the AI: use always-human for anything a mistake would be expensive in, and tiered for the volume that is eating review time. Auto-promotion is a soft gate with an objection window, so review still matters, it just stops being a click on everything.