Governer — the per-task risk score every gate reads

What it's for

If every check pushed back equally hard on every task, the harness would be unbearable on a one-line typo fix and not firm enough on a change to how payment works. Before real work starts, the agent answers in writing what could go wrong for a real person and how far a mistake would travel. That becomes one number, 0–100, and the rest of the harness reads it:

• the coding gate lets file edits through once the task has a score;
• the evidence gate requires proof from 20 up, and holds the agent to its written verification contract from 60 up;
• the agent's own depth: under 20 just build it; 20–59 light checks; 60–79 reflect before acting; 80+ every check, and it tells you plainly what's at stake first.

The score is computed on your machine by the plugin's copy of the author's scoring code (lib/scorer/), with no server and no key. It is written to this session's state with scoredBy: "agent". If scoring fails, the state says unscored — never a made-up number. On each message the harness may also note a quick keyword guess (for example "mentions billing"); it is labelled provisional, is only a hint, and never loosens any gate.

Is it set up?

It works out of the box with general risk tiers (payment and sign-in 100, onboarding 85, core feature 75, cosmetic 20, internal tooling 15, docs 0). Until you confirm your own table, every score is marked uncalibrated.

How to set it up

Run /alignment-harness:harness-setup → checkpoint "risk-table". It reads your repository for payment, sign-in, messaging and core-feature code (and, only if you say yes, your past sessions for what you most often corrected), proposes a table, and asks you to confirm each row. It is saved to alignment-harness risk-table path.

Dials (switches file, governer.*): gateStrength (block/remind/off), maxBlocksPerSession (3 — after that the gate only reminds), criticalMassFloor (80 — when two of the three "how dangerous is a mistake" answers are 8+), writeProvisionalBaseline, injectScoreBanner, applyToSubagents. The skill-prediction step needs an oracle of your own skill history (predictRequiredSkills.oracleCommand); without one, the governer skips it and says so.

Is it working?

• alignment-harness status — this session's score, who scored it, calibrated or not.
• alignment-harness doctor → governer: … N scores recorded, M distinct values. Scores should vary with the task. The same number every time means it isn't really being scored per task.
• Quick test: in a scratch repo ask for "fix a typo in a code comment" and then "change how a subscription is charged". The first should score under 20, the second 80 or more.