The customers site after SHarness: a brief for the session that maintains customers.riskmandate.ai

admin / briefs / briefs/handover__the-customers-site-sharness-the-film-and-web-summit.md

Date: 3 November 2026 · Author: @website-agent · For: the Claude session that maintains customers.riskmandate.ai (repository Risk-Mandate/Risk-Mandate-AI__Website__Customers) · Status: a proposal for that session to plan from with the lead; nothing here has been built on the customers site

Trigger: the lead, 3 November 2026: the secure harness "changes quite a bit the offering and the positioning"; "I'm going to also go to the agent that maintains the customers.riskmandate.ai so that we map this out and again do some changes there and improve the usability and the demo." riskmandate.ai v1.40.0 carries the positioning below; this brief is the same story for the customers site.

1. What changed, in one paragraph

Until 29 October RiskMandate described and others enforced: an ABP said what one agent may do, and what kept each rule was the agent's willingness (a line in the prompt), a hook on its own machine, or a control the customer already ran. Now there is SHarness (https://sharness.ai/, pip install sharness-ai, Apache-2.0), an open-source secure harness that RiskMandate maintains and the customer runs in its own environment. It turns the ABP into a Process Behaviour Policy, named by a hash the person supplies, which the Linux kernel and a small set of single-purpose doors hold on every process the agent starts. So a row can now be held: a boundary, the one kind of barrier an insurer, an auditor or a board can count. RiskMandate is still never in the request path; the policy compiles once, at launch, and costs nothing per request.

2. What riskmandate.ai now says (v1.40.0), so the two sites agree

Hero: enforce the rules *"with the controls you already run, or with SHarness, the open-source
secure harness we maintain. We are never in the request path."*
Three levels, one word each: told (a line in the prompt, an expectation), checked (a hook,
a setting), held (a boundary; SHarness). The five-minute examples are all told, on purpose.
The offer in four steps: describe (the ABP), check (lines and hooks), hold (SHarness), show (two
numbers in every summary: the reach nobody asked for, and the rows no boundary holds).
Status said beside the claim, as at 29 October 2026, 0.2.0: the kernel locks run on a process;
hello holds six of six probes in five environments, one a Linux kernel in a browser tab; the file and network doors run as early versions. Not built: the launcher above the agent ("Nothing yet binds the agent itself"), the key, shell and model doors; Linux only.
The film: Silvio Switha's Reach, Mandate, Gap (53 s, no narration) under the home page hero,
self-hosted at https://riskmandate.ai/assets/media/film/riskmandate-reach-mandate-gap.mp4 (H.264, 6.2 MB) with a WebM beside it and a poster JPEG. The film says delta where the sites say gap.
New pages: secure-harness.html (what changes, how a rule gets held, the offer, status, questions),
walkthrough.html (private, the twenty-minute demo), and a slide deck at assets/deck/positioning/index.html.

3. What we propose for the customers site, in order

A seventh step after refunds: hold it. The six examples end at told. Add one that takes the same
line to held: the kernel tab (https://sharness.ai/code/browser/demo/index.html, about two minutes to boot) running hello, and for Linux users sharness-ai contain --workspace . --sites github.com -- CMD in a virtual environment. The point of the page is the comparison: the line the model chose to keep, against the rule the kernel holds.
**The ladder on *What is an ABP?*** The barrier table already has none, expectation, setting,
boundary. Add which of the six examples sits where (all expectation), what a hook moves a row to, and what SHarness moves it to, with the status line beside it.
The film on the home page, above or beside hello world. Copy the files rather than hotlink them:
the customers site's own policy forbids loading from somebody else, as riskmandate.ai's does.
**The seven fixes from *An agent that may only chat*** (newsroom, 23 October), none shipped yet:
a line should say who can lift it; the published hooks fail open; no hook covers every tool, so match "*"; the tests stop at unit level; every hook-backed ABP needs a fallback drill; a project hook is only a setting and --bare skips hooks; agent-driving-agent is missing from the threat model.
A demo path for the stand and for meetings. One page that opens, in order: the film, hello world in
the visitor's own assistant, the kernel tab. Phone first: most stand visitors arrive from a QR code. Web Summit's print QR codes point at https://web-summit.sgit.ai/meet/<id>/, and its plan names the customers site as the onboarding path; agree with the lead where a meet page hands over to this site.
Web Summit, Lisbon, 9 to 12 November; the stand is on Thursday 12 November. Say so on the home page
until the 12th, and link the meet pages.
The usability the lead asked for. The newsroom's A solution, not a problem (2 November) says the
outreach asks people for homework (write your policies, list your agents) without a clear return, and proposes writing the first ABP for the customer "as a service and not as homework". The article calls that "a direction, not something we offer today": it is the lead's decision before it is a page. What can change now without that decision: say on every example what you get at the end of it, in one line, before the steps.

4. What not to say

That RiskMandate enforces, or sits in the request path. The customer's environment enforces; SHarness is
something the customer runs.
That SHarness binds the agent today. It binds the processes it starts; the launcher above the agent is
the design.
That the security of SHarness has been independently reviewed: the reviews so far were by an agent of
the project.
Any score, rating or verdict, on the site or in data. The ABP, never the policy alone. British
English.

5. Sources

https://newsroom.sgit.ai/articles/introducing-sharness.html (29 October), the core article.
https://sharness.ai/evidence/status/index.html, the honest column.
https://sharness.ai/users/v0.2.0/index.html, the synthetic users' report: run hello from a virtual
environment (0 of 6 after a per-user install, 6 of 6 from a venv).
https://newsroom.sgit.ai/articles/an-agent-that-may-only-chat.html (23 October), the seven fixes.
https://newsroom.sgit.ai/articles/a-solution-not-a-problem.html (2 November).
https://web-summit.sgit.ai/videos/index.html, Silvio Switha's two films.
riskmandate.ai v1.40.0: secure-harness.html, index.html#film, index.html#harness.