Mario Alexandre  ·  March 26, 2026  ·  token-savings auto-scatter sinc-llm

The Screenshot That Proves the Auto-Scatter Hook Works

I keep saying I saved $1,588.56 in one week. I keep saying 21,194 prompts. I keep saying the exchange rate went from 4.2 to 1.6. I keep saying $42 in Haiku overhead. People ask: can you prove it?

Yes. Here are the real numbers from my measurement system. These come straight from the logs. They are not estimates. I will walk through each one.

sinc-LLM — what the hook injects before every prompt
x(t) = Σ x(nT) · sinc((t - nT) / T)

The Raw Numbers

// 7-day measurement period — Claude Code + scatter hook

user_prompts_total
21,194
assistant_responses_total
33,133
exchange_rate_measured
1.56 (declining from 4.2)
exchange_rate_baseline
4.2 (pre-hook week average)
output_tokens_7d
12,900,000
total_tokens_7d
7,140,000,000
actual_cost_7d
$967.01
projected_cost_at_baseline
$2,597.96
savings_7d
$1,588.56
haiku_scatter_calls
21,194
haiku_cost_7d
$42.39
cost_per_scatter_call
$0.002
savings_per_scatter_call
$0.08
roi_haiku_spend
38x
snr_before_scatter
0.003
snr_after_scatter
0.855
quality_observations
275

How I Measured Exchange Rate

Exchange rate means total assistant responses divided by total user prompts. It is a simple ratio. For the baseline (the week before the hook), I checked the logs from the prior week. Same workflow, same agent pipelines, no scatter hook. The average was 4.2 assistant responses for every user prompt.

During the 7-day test, the scatter hook was on. My 21,194 user prompts produced 33,133 assistant responses. That gives a rate of 1.56. The rate kept dropping through the week as the hook warmed up and I tuned the scatter template. I expect it will settle between 1.4 and 1.6 once fully stable.

How I Projected the Baseline Cost

The $2,597.96 projected cost is not a guess. I took the real 7-day token volume and applied the 4.2 exchange rate. Then I used current API pricing. The $967.01 actual cost comes straight from my real billing data.

The difference of $1,588.56 is real. It shows up in my Anthropic billing dashboard. I am not guessing from a small sample.

The SNR Measurement

SNR (signal-to-noise ratio) of prompts: 0.003 before scatter, 0.855 after. This is the ratio of useful content (on-topic information in all 6 bands) to noise (ambiguity, missing detail, missing bands). I measured this across 275 prompt-response pairs. I rated each one by hand for quality.

A score of 0.003 is almost zero signal. My raw prompts were almost all noise from the model's point of view. They were so vague and incomplete that the model had to ask clarifying questions just to move forward. A score of 0.855 is close to perfect signal. The model has almost everything it needs to answer correctly on the first try.

Why I'm Publishing This

Big claims need evidence. "I saved $1,588 in one week" sounds like a sales pitch. I want you to have the real numbers. I want you to understand how I measured them. And I want you to be able to verify them yourself if you use the hook.

The logging system that made these numbers is part of the open-source package. You can run it on your own workflow and get your own results. Leave a comment and I will send you the GitHub link.

// Production AI Engineering

Build AI systems that hold up in production.

sinc-LLM designs, audits, and stabilises production AI infrastructure: from vendor evaluation and cost accountability to incident controls and MCP architecture.

See what we do →