Shunt Studio

What an editable shunt is

A shunt is one concept, defined by example. You write phrases that mean the thing you're watching for; Aldous embeds them and keeps their center point. Anything landing near that center trips the shunt — including wordings you never listed. This is not a keyword blocklist, which is exactly why paraphrases, typos, and novel insults still get caught. See Shunts & Latent Concept Erasure for the geometry behind that.

The more angles a phrase set covers — first/second/third person, whole sentences, bare token clusters — the wider and truer the net. That's what the phrasing coach on the right of the editor grades.

The studio workflow

Three panes, left to right:

1. Your shunts — the list of shunts in your account bank, plus starter templates (the built-in safety defaults, and a reward example) and a + New shunt button.
2. Editor — name (lowercase with underscores, 34 chars max), Intent (Reward or Discourage), Action, and a phrase composer (up to 32 phrases, 300 characters each). One idea per line; mix voices, whole sentences, and bare token clusters.
3. Phrasing coach — a health score out of 100, flags on weak or narrow phrasing, and a test playground where you paste a specimen to see whether this shunt would catch it.

Nothing goes live until you press Deploy — edits stay yours, in the browser, until then. Deploying re-embeds the whole bank in one pass and live-reloads it into the scoring gateway.

!!!SCREEN SHOT: The Shunt Studio editor with a shunt open, the phrasing coach showing a health score and flags, and the test playground populated with a probe specimen.

Live coaching

Wording lints run client-side and are always on. A second layer — live coaching — embeds your draft phrases as you type for real semantic feedback (does this phrase actually land near the centroid you intend?). It's a toggle in the header, since it costs a small amount of scoring capacity per keystroke pause.

Intent: reward vs. discourage

• Discourage dampens, receipts, or quietly logs what you want less of. Action options: minimize, receipt, log_only.
• Reward highlights what you want more of. Its only action is highlight — instead of suppressing a match, it surfaces it.

Starting from a template

The studio ships with the real Trust & Safety default phrase sets as starting points — racism, sexism, harassment — plus one reward exemplar, thoughtful_help, built from phrases like "You reproduced the bug and pinned the exact line — really helpful" and "Explaining the tradeoff instead of just saying no." Start a new shunt from any template and edit it to fit your community's own language.

Incentive shunt trackers: finding your unsung heroes

Reward shunts are the mechanism behind incentive tracking. A shunt trained on patient, specific, helpful responses — like thoughtful_help — doesn't flag anything to suppress; it finds the people who are quietly doing the good work of holding a community together, so you can recognize them instead of only ever hearing about the people causing friction. Pair a reward shunt with the GitHub Interactions Envoy or a chat channel and the highlight action surfaces those contributors in your graphs over time.

Limits

• 24 shunts per account
• 32 phrases per shunt, 300 characters each
• 34-character shunt names, lowercase with underscores

Where to go next

Shunts you build here apply anywhere your account scores content — including the Manual Spectrometer and connected sensors. For per-repository shunts that live in version control instead of the studio, see the GitHub Interactions Envoy.