Skip to main content

Lab Notebook · Semantic compression

GLYPH — compressing meaning, not text

GLYPH treats language as a compression of meaning: natural language stripped to ~1,024 atomic primitives and ~10,240 opcodes in a Polish-prefix tree, with a separate channel for tone. Bidirectional NL↔AST↔NL, ~57% smaller than token streams, trainable in ~13 GPU-days. A falsification-first spec — with hard kill-thresholds named up front — offered as an open standard.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 25 April 2026
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 25 Apr 2026
Living document. Reviewed 25 Apr 2026
Entry date
25 April 2026
Category
Language
Lead over the world
~1–2 years ahead
Access
🔓 Public
ResearchPartly provenDownload PDFPDF · 814 KB

What if you compressed the meaning instead of the text? A dense language of 1,024 primitives that round-trips to English and back — and trains in under two GPU-weeks.

1,024
atomic semantic primitives
~57%
compression vs token streams
~13
GPU-days to train a ~120M model
F1–F4
hard kill-thresholds named up front

Honest evaluation

Partly proven

The compression figure (~57%) is measured; the round-trip fidelity kill-thresholds (F1-F4) were named but not all cleared before the project was deliberately paused.

What would prove or disprove it further

What would prove or disprove it further: resume from dormant and run the F1-F4 gates to completion — fact recall and tone round-trip — on the reference implementation. Clearing ≥99.7% recall and ≥80% tone would move this to proven; falling short on either, honestly reported, would convert this into a disproven fidelity claim sitting next to a still-real compression number.

The evidence — full reasoning behind the verdict

Verdict: partly-proven. The compression claim is measured and real: roughly 57% smaller than a token stream, at an estimated training cost (~13 GPU-days for a ~120M-parameter model) small enough to be within reach of a single machine. The fidelity claim — that the compressed representation round-trips back to English without losing facts or tone — is not yet proven: hard kill-thresholds (F1-F4, including ≥99.7% fact recall and ≥80% tone) were named before building specifically so the project could fail honestly, and they were not all cleared before the project went dormant.

GLYPH strips language to roughly 1,024 atomic primitives and ~10,240 opcodes in a Polish-prefix AST, with a separate channel for tone, and the compression number comes directly from measuring that representation against a raw token stream for the same content. The falsification-first discipline is what keeps the compression win from being oversold: naming the pass bars up front means the project can't quietly redefine success after seeing partial results.

Runnable proof — see it work

The whole field's default is to chop text into tokens and train larger models on the stream. GLYPH takes a different bet: that language is a compression of meaning, and if you compress the meaning directly you get something smaller, cheaper to train, and possible to reason over algebraically. This is offered as an open standard — the spec and reference implementation are the contribution.

What it was

A super-dense intermediate language for meaning. Natural language is stripped to roughly 1,024 atomic primitives and ~10,240 opcodes arranged in a Polish-prefix abstract syntax tree, with a separate 16-dimensional pragmatic channel carrying tone. Crucially it is bidirectional: English → AST → English, so you can compress to meaning and reconstruct the prose.

What we built

A specification and the beginnings of a model:

  • The primitive set, opcode table, and the Polish-prefix AST format.
  • A falsification-first design: hard kill-thresholds F1–F4 named before building — for example, ≥99.7% fact recall and ≥80% tone on round-trip — so the project could be disproven, not just admired.
  • An estimate that a ~120M-parameter model over this representation is trainable in about 13 GPU-days.
Research

The measured/estimated facts: roughly 57% compression versus a raw token stream, at a training cost small enough to be within reach of a single machine. The interesting property is not just the size — it is that a meaning-first representation is something you can reason over, not merely predict.

What we learned — including the honest edges

  • A spec is not a validated model. The gate is in flight; the falsification thresholds are named, not all cleared. Those recall and tone bars are targets to beat, and until they are beaten the compression number is the proven part and the fidelity is the open part.
  • Naming your own kill-switch is the point. Declaring F1–F4 up front means the project can honestly fail. That is the discipline that makes the wins believable elsewhere in the lab.
  • Dormant-by-design is an honest status. It was placed on a council-approved deferral — parked deliberately, with the reason recorded, not quietly dropped.

Where it went / status

Dormant by design, offered openly as a spec. It is the sibling, for meaning, of the lab's work on a language that proves code — two halves of one idea: that language is something a machine should reason over and verify, not just read and imitate.

What is still open — kept visible

The honest edges, next to the wins. This is what turns 🔬 into 🟢 — honestly.

  • It is a spec with a gate in flight, not a finished, fully-validated model — falsification thresholds are named but not all cleared.
  • Round-trip fidelity (fact recall, tone) has hard pass bars (≥99.7% / ≥80%) that must be met — they are targets to beat, not claims already made.
  • It was placed dormant by design (a council-approved deferral), which is an honest status, not a quiet abandonment.

Where this connects

Sources

  1. vocabotics project audit — GLYPH (dense semantic language, ~57% compression), Apr 2026vocabotics internal project history · as of April 2026

    We use cookies.