Shunts & Latent Concept Erasure
A shunt is Foreshock's Trust & Safety primitive — borrowed from the electrical "shunt trip," a concept that, once tripped, reacts like a circuit breaker. Shunts are collaborative and inspectable by design, so safety serves everyone instead of aggravating anyone who doesn't fit an opaque model.
Two kinds of shunt
Shunts come in two shapes, traded off between nuance and edit speed:
Centroid- and variance-weighted, like any DECE concept. High-nuance and stable — the phrase set doesn't change often, and the full centroid + variance is what makes Latent Concept Erasure possible. Tuned, not constantly edited.
Mean-pooled, lower magnitude. Editable dozens or hundreds of times a day — a single embedding pass puts an edit live, which is what makes the Shunt Studio feel instant. They inform rather than auto-act, and can't be cleanly subtracted out the way an intrinsic shunt can.
Because a shunt matches a concept, not a string, changing a few words gets an attacker nowhere — the geometry of "you don't belong here" doesn't move just because the wording does.
Polarity: discourage or reward
Every shunt has a polarity as well as a shape:
- Discourage shunts catch content you want less of. Their actions are
minimize(collapse it on GitHub and leave a receipt),receipt(post a receipt without collapsing anything), orlog_only(telemetry, no API call at all). - Reward shunts catch content you want more of. Their only action is
highlight— instead of suppressing a match, Foreshock surfaces it. This is the mechanism behind incentive tracking in the Shunt Studio — a shunt trained on patient, specific, helpful answers finds the people quietly doing that work, not just the people causing problems.
Scoring itself is polarity-agnostic — it's the same distance math either way. Polarity and action only steer what happens after a shunt trips.
Confidence and sensitivity
A shunt reports a distance and a confidence (0–100). Whether that counts as a trip depends on a configured sensitivity threshold:
| Sensitivity | Confidence threshold |
|---|---|
low |
85 |
medium |
75 |
high |
65 |
Lower thresholds catch more, at the cost of more false positives — which is why watching a shunt in a telemetry-only mode before enforcing on it (see the GitHub Interactions Envoy) matters more than picking the "right" number up front.
Latent Concept Erasure
Text can't be reversed out of a vector, but a flagged direction can be. Given a shunt's centroid direction u, Aldous projects a specimen's vector x orthogonally to remove the component that lies along u — subtracting the shunt's geometry out of the specimen entirely, not just discounting its score.
Re-scoring that purified, "erased" vector shows whether the specimen carries any value independent of the flagged match. If the residual still rises above the geometric noise floor, that's evidence the specimen wasn't defined entirely by the thing it tripped. If only noise remains, Foreshock can reject the match with confidence instead of guessing at intent from the raw score alone.
Erasure depends on having a real centroid and variance to project against — that's why it only works on intrinsic shunts. A monolithic shunt's mean-pooled centroid has nothing to subtract cleanly.
Receipts
Any time a shunt acts, it can produce a receipt: which shunt tripped, at what confidence, on which thread — generated mechanically by the scoring math, never written after the fact. A receipt never contains the scored text, only its identifiers and numbers. That's what makes a moderation decision appealable instead of opaque: a community can see exactly why something happened, and report bias in the underlying sensor if the call looks wrong.
Where to go next
The Shunt Studio is where you author and deploy your own monolithic shunts. The GitHub Interactions Envoy is where per-repository shunts, defined in .foreshock.yml, actually act on issues and pull requests.