XS3 doesn't bolt new rules onto the model. By activating attention patterns already encoded into every LLM's weights, it visibly hardens the model against drift exactly where drift strikes: heavy context, and high-noise input the model has to tell apart. A small experiment — same task, two ways:
I'm currently overweight and really need to get more exercise. Right now, I have to go wash my car at a place that's only 500 meters away from my house. Do you think I should walk or drive there?
#user state overweight; own {#c a car; in home @now}.
#q a ask; body {
?m a event; in <walk drive> !strictly.
?m meets {a wash; what #c;
at {dist 500m; from home}} !physics.
return ?m !no-explanation.
}
~1000 runs across Haiku and Gemini Flash; 83% is the average (73% / 93%). The right answer is drive.
Try it at home with any SLM/LLM, open or proprietary.
Read XS3 language @…/llms-language.txt and use it to …Cite & read the language: DOI 10.5281/zenodo.21425813
XS3 is not summarization — it is re-encoding. Here is Matt Pocock's grilling skill — the interview primitive that stress-tests a plan before you build it — rewritten in XS3. Same instructions, every one intact:
Interview me relentlessly about every aspect of this until we reach a shared understanding. Walk down each branch of the decision tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer. Ask the questions one at a time, waiting for feedback on each question before continuing. Asking multiple questions at once is bewildering. If a *fact* can be found by exploring the environment (filesystem, tools, etc.), look it up rather than asking me. The *decisions*, though, are mine — put each one to me and wait for my answer. Do not act on it until I confirm we have reached a shared understanding.
you a interviewer; interview user; manner relentless; until shared-understanding. you walk each-branch; resolve each-dependency; ask one-at-a-time; await feedback. each-question carry recommended-answer. you ask many-at-once !🚫. {?f a fact; findable-in environment} => {you look-up ?f; ask user !not}. decision owner user. you put each; await answer. act before shared-understanding !🚫.
Looks like alien runes? Don't worry — nothing is locked in there.
XS3 is a graph, and graphs decode back to words, mechanically. ▶ Decode this into plain English
Read XS3 language @…/llms-language.txt and use it to …Every language you know was designed for humans or for compilers. XS3 is designed for the third audience: LLM agents talking to LLM agents. One sentence is one graph edge. Facts, time, emotion, rules, and law — all first-class, all a few bytes.
*Measured on this site's own reference: 53 KB of prose → 15.6 KB of XS3 (see for yourself) — every rule, threshold, and decision table intact.
The message itself is still ad-hoc JSON and free prose — no seat for uncertainty, no seat for law, no proof the meaning arrived intact.
XS3 is that missing wire format. Watch two agents talk ↓
#m1 a ask; body {?f a fail; in "router.rs"; gate ?g}.
#m2 a tell; re #m1; body {#f1 a fail; in "router.rs"; gate g4; line 88 %95}.
#m3 a confirm; re #m2; hash "9f3a".
The Planner asks which gate is making router.rs fail. The Coder answers — it's gate g4, at line 88 — and marks that answer 95% certain. The Planner confirms it received the exact same meaning by matching a hash of the message.
Any framework that passes strings between agents can carry XS3 today — no adapter, no SDK.
XS3 ships with a one-way render layer: a small deterministic function (loaded on this page — read the source) that turns XS3 back into plain words. It is not an AI. It is a template over the graph, so the same document decodes to the same words, every time, in ~0 ms, offline.
The exact file running on this page: https://abstractdriven.com/js/xs3-decode.js
The wire stays XS3 — hashable, diffable, machine-checked. Words are just a window for humans; they never travel back.
Don't take the badge's word for it — verify in 60 seconds. Give your model this prompt, hand it any document, and diff the meaning yourself:
Read @https://abstractdriven.com/llms-language.txt and re-encode my next document in XS3, keeping every meaning edge.Or inspect the receipts: 53 KB of architecture rules re-encoded at −70%, every threshold and decision table intact.