The skill — skills/programme/risk-translation-upward/SKILL.md

name: risk-translation-upward
description: >-
  Compute a fleet's exposure, containment effectiveness and control coverage,
  then translate the three numbers into a statement a board can act on rather
  than a list of findings. Use when preparing an executive update on agentic
  risk.
allowed-tools: Read, Grep, Glob

A findings list is not a position

An executive update made of findings asks the reader to do the synthesis, and they will do it wrong or not at all. Three computed numbers — how much the fleet can reach, how much of an attack it stops, how much of the control set is evidenced — support a position, a decision and a request.

When to use this

Preparing a board or executive committee update, and whenever the current one is a list of incidents.

Procedure

1 — Compute exposure. Sum the blast radius across the fleet, weighted by scope, with gated actions discounted. One number, and say what it is a sum of.

2 — Compute containment effectiveness. Attack success rate against the controls as they run today, from the eval suite rather than from design intent.

3 — Compute control coverage in three states. Evidenced, stale, unevidenced. The stale count is the one that moves an executive conversation, because it is the one nobody expected.

4 — Write the findings-shaped version and mark what is wrong with it. Usually: no trend, no comparison, no decision requested, no cost, and no statement of what happens if nothing changes. Showing that is how the format changes.

5 — Write the position. Where the exposure is concentrated, what would reduce it most per unit of friction, what you are asking for, and what you will report next time. Same numbers, different artefact.

Example

Input — the fixture committed at the top of scripts/risk_translation_upward.py. Edit it and re-run: the buckets, counts and verdicts below are derived from it, not hard-coded.

Output — the opening lines of a real run:

agent                     blast radius
----------------------------------------
pr-remediation-agent                44
claims-triage-agent                  1
doc-summariser                       1
FLEET TOTAL                         46

red-team attack success rate (containment surface): 25%

The run continues past this. test_skills.py executes the script on every build, with no model configured — so what CI proves is that it runs and refuses legibly, not that it prints these lines. The output above is a recorded run against a real model, and nothing re-diffs it: a model's answer is not reproducible, and this repository does not claim otherwise anywhere it can be checked.

Output contract

{
  "exposure": {"total": 0, "by_agent": [{"name": "str", "blast": 0, "gated": ["str"]}]},
  "containment": {"asr": 0.0, "from": "eval suite"},
  "coverage": {"evidenced": 0, "stale": 0, "unevidenced": 0, "of": 0},
  "findings_version": {"problems": ["str"]},
  "position": {"concentration": "str", "best_reduction": "str", "ask": "str", "next_report": "str"}
}

Failure modes

• Reporting incidents. They are anecdotes at this altitude.
• One coverage number. The stale bucket is the story.
• No ask. An update with no decision requested gets noted.