Benchmarks — Promptolian

9.2%

Standard tier

avg CR — all prompts

20.1%

Developer tier

avg CR — long prompts

69%

Tool schemas

CR — turn 1 (JSON → DSL)

92%

Tool schemas

CR — 5-turn session avg

89.9%

Fact preservation

across all tiers

Domain Study

Results by domain

200 English prompts (150 short + 50 long) across 5 professional domains. Engine v4 — fully deterministic, no external API. Token counts via tiktoken cl100k_base.

Domain	Standard CR	Pro CR	Developer CR	Fact preservation
💹Financen=40	10.8%	11.5%	12.9%	90%
💬Social / Marketingn=40	8.6%	9.4%	9.7%	90%
💻Programmingn=40	11.0%	11.7%	11.8%	90%
⚖️Legaln=40	8.9%	10.3%	11.1%	89%
🏥Medicaln=40	6.7%	6.8%	7.1%	89%

Developer tier applies domain-specific abbreviation packs (MRR, ARR, NDA, IP, SLA, BP…) + spaCy clause pruning + parenthetical removal. All processing is local — no external API calls.

Tier Comparison

What you get at each tier

Numbers, URLs, names, and entities are always protected across all tiers.

Standard · Free

~9%

avg CR (up to 16% on long prompts)

Symbol substitution (§EXP, §NOT, →code…)
Phrase simplification
EN · ES · FR · DE · IT
Rule-based, instant, deterministic
No external API calls

Pro · $9/mo

~10%

avg CR (up to 18% on long prompts)

Everything in Standard
Telegraphic framing removal
Math operator substitution (≥, ≤, ≠)
Verbose phrase compression
Numeric abbreviation (1000 → 1k)

Developer · $29/mo

~11%

avg CR on prompts · up to 92% on tool schemas

Everything in Pro
Domain packs (finance, legal, medical, code)
spaCy clause pruning
Context engine (multi-turn delta pruning)
Tool schema DSL compiler (69–97% per session)
100% local — no external API calls

Live Examples

From the benchmark set

Real prompts from the dataset, showing original and Developer-tier output.

💹 Finance 71% smaller

Original · 41 tokens

Analyze the monthly recurring revenue growth for Q1 2026. Revenue increased from $2.4M to $2.9M, a 20.8% quarter-over-quarter gain. Project the annual recurring revenue trajectory for the full year.

Developer · 12 tokens

ANLZ MRR growth Q1 2026: $2.4M→$2.9M (+20.8% QoQ). Project ARR full year.

💻 Programming ~33% smaller

Original

You are an expert Python developer. Please review this function and return only the code with the bug fixed. Do not include any explanation.

Standard

§EXP py developer. « this FN →code BUG fixed. §NOT include any ? .

💬 Marketing ~45% smaller

Original

Summarize the key benefits of our product for a social media post. Be very concise. Return as bullet list. Important: focus on the unique value proposition.

Pro

∑ key benefits for social media post. →short. →list. !!: focus on unique value prop.

Agent Layer

Tool schema compression

JSON tool definitions are the largest token consumer in agent workloads — often larger than the user prompt itself. Promptolian compiles them to a compact function-signature DSL and caches them across turns.

69.4%

Turn 1 CR

JSON → DSL compilation

97.5%

Turn 2+ CR

session cache reference

91.9%

5-turn session

average across all turns

JSON schema — 111 tokens

        {"type":"function","function":{"name":"search_web","description":"Search the web for up-to-date information on any topic","parameters":{"type":"object","properties":{"query":{"type":"string","description":"The search query"},"num_results":{"type":"integer","default":10},"language":{"type":"string","default":"en"},"safe_search":{"type":"boolean","default":true}},"required":["query"]}}}
      

DSL — 34 tokens (69% smaller)

        search_web(query, language='en', num_results: int=10, safe_search: bool=True)

          # Search the web for up-to-date information…

        # Turn 2+:

        TOOLS:[search_web,get_user,create_order,…]

        # → 30 tokens (97.5% CR)

Tool	Raw JSON tokens	DSL tokens (turn 1)	CR turn 1
`search_web`	111	34	69.4%
`get_user`	98	30	69.4%
`create_order`	133	33	75.2%
`run_code`	119	41	65.5%
`query_database`	139	40	71.2%
`send_email`	144	42	70.8%
Total (10 tools)	1,220	373	69.4%

Type elision: 61% of parameters have their type annotation dropped because it is inferrable from the name alone (user_id → str, include_orders → bool, limit → int). Enum fields are rendered inline as active|inactive|suspended instead of a full JSON array.

Methodology

How we measure

All measurements are reproducible. The experiment code is in the repo.

Token counting

All token counts use tiktoken cl100k_base (OpenAI's tokenizer, used by GPT-4 and Claude approximation). Older experiments used a words×1.3 heuristic — clearly marked.

Compression rate (CR)

(1 - compressed_tokens / original_tokens) × 100. Measured on the prompt sent to the LLM, not the response.

Fact preservation rate (FPR)

Regex extraction of numbers, URLs, named entities, and quoted strings from the original. We verify each appears in the compressed output (or its decoded form). FPR = retained / total × 100.

Developer tier (NLP pipeline)

Pro output is passed through: (1) domain pack — 60+ domain-specific abbreviations detected via keyword scoring, (2) spaCy dependency parse — non-restrictive relative clauses and low-information adverbial clauses removed, (3) parenthetical pruner — non-numeric parentheticals ≥4 words removed. Fully local, no external API calls.

Prompt benchmark: domain_experiment_v4.py · Tool benchmark: tool_compression_experiment.py · Engine: engine_v4.py · API: POST /compress-tools

Real numbers.No fabrication.

Token counting

Compression rate (CR)

Fact preservation rate (FPR)

Developer tier (NLP pipeline)

Real numbers.
No fabrication.