---
id: "lib-2026-10-04-write-like-it-s-1866-llms-relearn-telegraphese"
type: "library"
title: "Write Like It's 1866: LLMs Relearn Telegraphese"
lang: "EN"
org: "Five Minutes Forward (Travis Smith)"
source_type: "Practitioner post"
date: "2026-10-04"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-04-write-like-it-s-1866-llms-relearn-telegraphese/"
original_url: "https://fiveminutesforward.com/post/2026-10-04-telegraph-test/"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Independent practitioner blog post with a reproducible public benchmark (50 passages, about 1,300 questions, paired McNemar tests), but self-run, not peer reviewed, and missing a key control."
summary_basis: "full-text"
original_minutes: "12"
tags: ["ai technology", "enablement", "token cost", "agent to agent communication"]
---

# Write Like It's 1866: LLMs Relearn Telegraphese

Tier B (Useful, limited): Independent practitioner blog post with a reproducible public benchmark (50 passages, about 1,300 questions, paired McNemar tests), but self-run, not peer reviewed, and missing a key control.

Summary based on the full text.

## In short

A blog post by Travis Smith on Five Minutes Forward (4 October 2026) with an open benchmark, the Telegraph Test. It asks models to write records in cablese, the compressed style telegraph operators used when words cost $10 each, and checks whether other models can still answer questions from those records. Output tokens drop 40-49% and other models read the compressed records as well as or better than plain summaries, as long as the compression happens after the content is settled.

## Key numbers

- 40-49%: fewer billed output tokens when models write records in lowercase cablese
- 0.99-1.10: recovery ratio of downstream models reading cablese records vs plaintext (1.00 = parity)
- about 10%: savings from codebook substitution, the other historical method
- 0.81: accuracy ratio when the model answers directly in cablese instead of plaintext

## What it reports

- Codebook substitution saves about 10%; instructing models to write cablese in lowercase saves 40.4% (gemma-4-31b), 48.9% (qwen3.8-27b) and 48.4% (GLM-5.3-Flash).
- Without the word "lowercase", models default to ALL CAPS and savings fall to 25.0% to 33.9%.
- Four reader models from different families answered 694 anchored questions per condition; recovery ratios ranged 0.99 to 1.10, with none favouring plaintext.
- Condition ladder: answering from a plaintext summary 75.8%, from a cablese record 82.5% (ratio 1.09), from a decoded record 81.6% (1.08).
- Answering directly in cablese drops accuracy to 74.4% against 91.5% for plaintext (ratio 0.81); expanding those answers recovers to 0.86.
- gpt-5-mini cannot turn off reasoning; cablese makes it think about 3x harder and its writes cost about double, so savings were only 17.7%.
- Storage loop: compress-then-expand costs one extra pass (~312 tokens), which only pays back through cheaper machine reads after one or two reads.

## How they got it

50 passages with about 24 short-answer questions each, graded by deterministic string matching at a 0.8 overlap threshold, no LLM judge. Scores are normalised against each condition's own plaintext control and compared with paired McNemar tests. Code, frozen ledgers and a notebook are on GitHub.

## Limits

- The author did not run the control of a plain "be as terse as possible" instruction, so the role of the historical register is an informed guess.
- Mostly small and mid-size open models; frontier models were not tested as writers.
- Question-answering on short facts does not test nuance, tone or reasoning quality.

## Why it matters for companies

For text that one system writes and another system reads, such as agent memory, handoffs and logs, a one-line instruction can cut output token spend by around 40%, and output tokens cost 3-5x input. The format stays human-readable, so audit trails survive. The method should not be used for text people read or for answers still being composed, and it fails with models whose reasoning cannot be switched off.

Original: https://fiveminutesforward.com/post/2026-10-04-telegraph-test/