ORIGINAL MEASUREMENTS / 3 OCTOBER 2026
The result.
And what it measures.
Ritoko saves a procedure your agent learned, replays its deterministic steps and keeps a local journal for each item. Here are two tests you can inspect: changing browser forms, and recovery after an interrupted customer batch.
01 / CHANGING BROWSER FORMS
RPA Challenge: 70 correct fields.
The saved workflow completed 10 spreadsheet rows with 100% success. The challenge moves its form fields after each submission. This is one measured run, not an average or a claim about every business website.
| Measurement | Result | What it includes |
|---|---|---|
| Challenge timer | 1,735 ms | The elapsed time reported by the challenge website. |
| Journal duration | 3,190 ms | The workflow run recorded by Ritoko, including its setup and checks. |
| CLI wall-clock | 3,559 ms | The complete CLI invocation, including work outside the challenge timer. |
02 / INTERRUPT, RESUME, RERUN
A crash leaves one item to check.
A local test back office accepts duplicate submissions. We kill the process after submission 5 reaches the server, resume it, then run the same file again. The test adds 700 ms to each submission so the interruption is visible.
| Stage | Journal state | Observed behavior |
|---|---|---|
| After the kill | 4 done · 1 unconfirmed · 5 pending | Five submissions have reached the server. The fifth has no confirmed result in the journal. |
| After resume | 9 done · 1 review | The four verified items are kept. The uncertain fifth is held; the five pending items complete. |
| After rerun | 9 skipped · 1 review | The server still has 10 submissions with 10 unique customer keys. |
This demonstrates recovery through the same local Ritoko installation and journal. It does not establish a universal exactly-once guarantee. Each real workflow needs a stable business key and a check that proves the destination’s result. An uncertain commit stays held until someone checks and resolves it.
What uses a model?
Your agent learns the task and helps repair a changed procedure. Unchanged deterministic browser, HTTP and MCP steps replay without a model call. Reading each new document with an agent or external service is a separate operation and may still use a model on every item.
Inspect, then reproduce.
The public measurement file records the runtime, timings and item counts. The repository contains the RPA workflow and setup instructions. Use the same inputs, timer boundary and runtime when comparing a new run; destination latency and session state can change the result.
Download the original measurement data ↗Set up Ritoko locally ↗