Early benchmarks: 30%+ input-token reduction from reusable reasoning — read the field note

Philosophy

Why a markup language for reasoning.

Four premises, one etymological note, one prior-art survey, five open questions.


Four premises

What Scholialang is built on.


Prior art

Why not just use what exists.

Scholialang sits in a lineage. It is not the first attempt at giving reasoning a structured surface. Here is how it compares — ● full coverage, ◐ partial, ○ none.

System Full vocabulary Composition rules Multi-agent semantics Human readable Machine parseable
Scholialang (this) 32 atoms, 11 ops, 6 primitives 17 validity rules <Review>, <Handoff>, <Question> XML-ish, designed for human eyes ~100-line parser
ReAct thought / action / observation only convention, not rules single agent fine in prose, not formal brittle line-based
Toulmin 6 roles (claim, grounds, warrant, backing, qualifier, rebuttal) argumentation schema human argument only no canonical encoding
AIF I-nodes + S-nodes graph constraints dialogue protocols add-on RDF, not really RDF / OWL
KQML / FIPA-ACL performatives, no reasoning atoms speech-act semantics multi-agent messaging LISP-like
Lean proof terms rich vocabulary strict type theory single prover formal, not readable fully machine-checkable

Scholialang sits in the middle: less formal than Lean, more structured than ReAct, multi-agent-native, and readable on its face.


Etymology

Why 'Scholialang'.

σχόλια — in Greek and Byzantine manuscript tradition, the structured critical commentary written in the margins of classical texts. The scholiasts annotated ancient manuscripts so later readers could trace the editorial reasoning.

The name was chosen deliberately as the opposite of neuralese — the feared private language of AI cognition.

Names considered, with a one-line reason each:


Unresolved

Open questions.

01

Timestamps.

Should every atom carry a wall-clock timestamp? Tradeoff: replayability vs. diff noise.

02

Signing.

Should atoms support cryptographic signing — multi-agent traces where provenance matters?

03

Streaming.

Should traces support append-only streaming semantics for live monitoring?

04

Grammar.

EBNF or a schema-first (JSON Schema) definition as canonical?

05

Natural-language compatibility.

Where is the line between a <Thinking> block and a structured sub-atom? Too much structure in Thinking defeats human readability.