What question did this study set out to answer?

The goal is to evaluate TASS as a compact, human-readable format for structured data extraction and its impact on inference cost.

June 1, 2026Open Access

Tokeniser-Aware Shorthand (TASS): A Stenography-Inspired Output Format for Reducing LLM Inference Cost in Structured Extraction Pipelines

Key Points

The goal is to evaluate TASS as a compact, human-readable format for structured data extraction and its impact on inference cost.
Validates TASS across eight application domains via byte-level measurements and live API benchmarks, involving 10 models and 400 API calls.
Compares TASS metrics against JSON and various established protocols, focusing on size, efficiency, and readability.
Calculates gas cost reductions in Ethereum transactions using TASS data format.
TASS averages 97 bytes compared to 189 bytes for JSON, achieving a −49% reduction across five production schemas.
At 1 million EVM transactions, TASS saves about $53,760 in gas costs compared to JSON.
LLM benchmarks show 0–42% gross output-token savings using TASS, with net savings of $9–$44 per million calls.

Abstract

TASS (Tokeniser-Aware Structured Shorthand) is a compact, human-readable text serialisation format using the syntax ~k: v — a tilde-prefixed single-character key, colon separator, and value, space-delimited. It requires no compiled schema, no external library, and is natively parseable by Large Language Models without fine-tuning. This paper (v3. 0) validates TASS across eight application domains through computed byte-level measurements and live API benchmarks (10 models, 400 API calls). All claims are labelled as computed C, measured M, or specification-derived S. Key findings: TASS averages 97 bytes versus 189 bytes for equivalent JSON across five production schemas (−49%), a reduction that holds at −37% after single-payload gzip and −50–53% after stream compression. TASS fits the LoRaWAN SF9 payload limit (115 bytes) where JSON fails in every test case. All five schemas fit a single 160-character GSM SMS as TASS; none fit as JSON. TASS (93B) is 33% smaller than NMEA 0183 RMC+GGA for a 10-field GPS fix. A 5-field blockchain transaction intent encoded as TASS (28B) fits Stellar's memoₜext hard limit (28 bytes) where JSON (72B) does not. At 1 million EVM transactions, TASS reduces calldata gas costs by approximately 53, 760 versus JSON (EIP-2028 pricing). In LLM output benchmarks, non-reasoning models achieve 0–42% gross output-token savings. Accounting for the ~113-token system-prompt overhead, net cost savings are approximately 9–44 per million calls depending on whether prompt caching is active. Reasoning models (DeepSeek V4-Pro, Kimi K2. 6) cost 9–10% more tokens with TASS due to think-chain overhead on symbol-table decoding. The paper includes direct protocol comparisons against Cayenne LPP (IoT/LoRaWAN), FIX 4. 2 (financial exchanges), NMEA 0183 (GPS), Bluetooth GATT binary (health wearables), and HL7 FHIR (healthcare). Financial exchange gateway protocols (FIX 4. 2) are explicitly out of scope; TASS applies to internal APIs, analytics pipelines, customer-facing alerts, multi-agent AI systems, blockchain memo fields, TinyML telemetry, and WebSocket state sync. TASS is the only evaluated format simultaneously satisfying: human-readable, compact enough for LoRaWAN SF9 / GSM SMS / QR code constraints, self-describing with a plain-text dictionary, parseable without a compiled library, and natively usable as LLM output.

Tokeniser-Aware Shorthand (TASS): A Stenography-Inspired Output Format for Reducing LLM Inference Cost in Structured Extraction Pipelines

Key Points

Abstract

Cite This Study