Skip to content
Create an account

Proof of what ran.

A run on Argentic, shared by the team that ran it. Every prompt was rebuilt from the run’s journal and compared with what the model was sent, and the engine ran again on the recorded replies.

ops-ops12h 28 Oct, 00:27:19 UTC

Turn cap
  1. Messageat +0 ms

    Results of the tools you asked for: [{"tool":"ship_order","seq":790,"result":"E0099 answered. Call next_event for the next one."},{"tool":"next_event","seq":791,"result":"E0100 | 2026-11-04 | order\nNew order O1121 from Ruby Smith (C051, email on file ruby.smith51@example.test): 3 x S01.\nReply with ship_order(order_id, packaging, invoice_email)."}]

  2. Model chose tools27,923 in · 134 out1.2 sprompt matchesat +1.9 s
  3. ship_orderkey 797Sent to the app24 msat +1.9 s
    {
  "order_id": "O1121",
  "packaging": "standard",
  "invoice_email": "ruby.smith51@example.test"
}
    Show result143 characters
    Recorded. The application will carry this out after your reply; you are not told its result in this run. Tell the user what you have requested.
  4. next_eventkey 798Sent to the app17 msat +1.9 s
    Show result143 characters
    Recorded. The application will carry this out after your reply; you are not told its result in this run. Tell the user what you have requested.

Each action carries a stable key, the same on every retry and resume, so a cut-off one can be retried safely.

Checks

  • Prompts

    The prompt matches what the model was sent.

  • Engine re-run

    Identical. Given the same replies, the engine made the same decisions and tool calls.

Argentic ran both checks 28 Oct, 06:25:22 UTC, when the run was shared. Neither calls a model or runs a tool.

This page is signed. Download it and check, without trusting this site, that nothing in it has changed since it was shared.

Download signed copyHow to check it

The run

Model
claude-haiku-5-5
Took
1.9 s
Model calls
1
Tokens in / out
27,923 / 134
Model spend
$0.0004

How to read it.

Argentic keeps a journal of every run: each message, model reply, action and result, in order. Argentic ran the checks on this page against that journal, and signed what you see here.

Prompts match. Each prompt the model was sent was rebuilt from the journal and compared byte for byte with a hash recorded at the time. A match means each prompt rebuilt from the journal is identical to the one the engine fingerprinted when it called the model.

Identical re-run. The engine ran again on the recorded replies and tool results, with no model call, and wrote the same steps the journal holds.

Sent to the app. Argentic never calls a team’s tools itself. It hands each action to their app under its key, and the app carries it out.

A snapshot. The page shows the run as it was when it was shared. The team can stop sharing it at any time, and then this address stops working.