Foreshock / Docs / Overview / Introduction

Introduction

OVERVIEW·5 min read

Foreshock reads the room instead of the message. It measures the emotional and intent spectrum of language from geometry alone — no LLM classification, no black boxes — and points that same instrument inward at your organization and outward at the content reaching it.

What Foreshock is

Most safety and sentiment tooling asks a model to understand text — summarize it, reason about it, decide what it means. Foreshock never does that. It converts language to vectors once, then measures distance and similarity between those vectors and a set of open, inspectable reference points. Classification based on math, not opaque LLMs.

That one design choice is what makes everything else possible:

  • Guides escalations that need humans to humans. Foreshock doesn't try to replace judgment calls — it surfaces the ones worth making, with enough context (confidence, thread, history) that a person can act on them quickly.
  • Watches organizational health, burnout, and growth-phase transitions. The same engine that flags a hostile comment in milliseconds can, pointed inward at everyday chatter, show tension, hedging, or partisanship building over weeks — long before it surfaces as a resignation wave or a public blow-up.
  • Monitors bias and tonality in any written content. A blog post, a press release, an AI-generated reply — anything can be scored for its likely emotional impact before it goes out, on a graduated scale rather than a single pass/fail label.
  • Keeps communities safe and involved in their own governance. The phrases that define every sensor and shunt are open and editable by the people they protect. A community owns its own bias reports and its own firewall, instead of appealing to a vendor's hidden model.
  • Built for confidence-based automation. Nothing has to go straight to auto-action. Watch a rule in a telemetry-only mode until you trust its judgment on your own traffic, then flip it to active with a one-line, reviewable change.
  • Receipts are mandatory and mechanically-assured. Every adverse action can carry a receipt — what tripped, at what confidence, on what thread — generated by the scoring math itself, not written after the fact to justify a decision.
// The mandate

Open corpus, geometry only, no PII, on-prem ready. Foreshock never reads, reasons about, or stores your text — only vectors and similarity scores leave the system.

Who uses Foreshock

  • Researchers (social, economic, statistics) — an open corpus and a one-shot instrument for synthetic examination, with raw numbers exportable for further analysis.
  • Trust & Safety teams — community-editable shunts, a GitHub interaction envoy, and chat monitoring that never stores what was said.
  • Community managers — a garden of feeds and channels to watch a community's health over time, plus incentive tracking that surfaces the members quietly holding things together.
  • Marketers — score the likely emotional impact of copy, campaigns, or AI-drafted replies before they ship, across 75+ dimensions.
  • Artists & musicians — the one-shot Manual Spectrometer scores any text you paste in, lyrics included, purely as geometry — nothing is stored or judged by a person.
  • Speakers & strategists — check how a script or statement is likely to land, on outlook and motivation as well as raw sentiment, before it's delivered.

How to read the sidebar

The left sidebar is grouped the way you'll actually use the product:

  • Concepts explain the model underneath everything else — how Aldous scores language, and how shunts and Latent Concept Erasure work. Read these if you want the math.
  • Operate is the hands-on material — one page per feature, from the Manual Spectrometer to the GitHub Interactions Envoy.
  • Reference is the exact shape of the programmatic API: authentication, rate limits, requests, and responses.

Each page carries a small header — its kind, reading time, and when it was last updated — so you can judge relevance at a glance. The On this page rail on the right tracks your position within longer articles.

// New here?

Start with How Aldous scores language for the mental model, then jump into whichever feature under Operate matches what you're trying to do.

Conventions

Code blocks are copy-paste ready and assume an fsk_… API key in your environment. Anything destructive, rate-sensitive, or plan-gated is called out inline. Timestamps are ISO-8601 UTC unless noted.