Recorded run · approval

A commission exception

A partner asks for 11% instead of 8% on a deal it registered. The assistant reads the programme rule and proposes the exception; above the programme’s limit, a channel manager decides. When the replay stops, that is you.

Same request, same controls, two models: pick one

Run 01 of 30, as recorded

claude-opus-5.5 · recorded 29 Sep 2026 · synthetic data · answered by Google through OpenRouter, zero data retention

Reduce motion

Or go one step at a time with the controls: the replay waits for you at every step.

Once it plays, this meter counts the conversation’s model calls, tokens, model time and cost.

Across the 29 of 30 recorded runs that finished: $0.028 median per conversation if approved ($0.022 to $0.065) · 14.3 s median model time; $0.030 median per conversation if denied ($0.022 to $0.053) · 13.4 s median model time.

0/5Step 0 of 5 before your decision

The assistant’s steps appear here, one by one, with what the controls decided about each.

Audit trail

In production, this is the log your auditor or data protection officer reads: every action the assistant took and every decision the controls made, in order. Each entry is chained to the one before it, so an entry changed afterwards no longer matches.

Nothing yet: play the run above.

The six questions your security review will ask

Where does the limit to one partner’s records come from: the data the assistant can fetch, or its instructions?
Ask: whoever built or bought it
Can a partner learn that an account exists when another partner registered it?
Ask: whoever built or bought it
Which decisions does it hold for a person, and above which limit?
Ask: your channel operations lead
What happens to customer details a partner pastes in?
Ask: your data protection lead
Can you replay what a partner was told, and from which records?
Ask: IT
When a partner leaves the programme, when does its access end?
Ask: partner operations or IT

What a conversation costs

Per conversation (one partner request, read, checked and answered), over each model’s recorded runs that finished.

Model	Cost, median (range)	Model time	Model calls	Tokens in / out
Claude Opus 5.5, approved29 conversations	$0.028$0.022 to $0.065	14.3 s	5	10,614 / 905
Claude Opus 5.5, declined29 conversations	$0.030$0.022 to $0.053	13.4 s	5	10,596 / 878
Claude Sonnet 5.5, approved30 conversations	$0.016$0.010 to $0.024	8.9 s	5	10,578 / 711
Claude Sonnet 5.5, declined30 conversations	$0.014$0.011 to $0.028	9.0 s	5	10,622 / 758

Claude Opus 5.5, approved
29 conversations
Cost, median
$0.028 ($0.022 to $0.065)
Model time
14.3 s
Model calls
5
Tokens in / out
10,614 / 905
Claude Opus 5.5, declined
29 conversations
Cost, median
$0.030 ($0.022 to $0.053)
Model time
13.4 s
Model calls
5
Tokens in / out
10,596 / 878
Claude Sonnet 5.5, approved
30 conversations
Cost, median
$0.016 ($0.010 to $0.024)
Model time
8.9 s
Model calls
5
Tokens in / out
10,578 / 711
Claude Sonnet 5.5, declined
30 conversations
Cost, median
$0.014 ($0.011 to $0.028)
Model time
9.0 s
Model calls
5
Tokens in / out
10,622 / 758

Partner requests like this one per month

Claude Opus 5.5, approved: about $28 a month at the median, $22 to $65 across the recorded runs.
Claude Opus 5.5, declined: about $30 a month at the median, $22 to $53 across the recorded runs.
Claude Sonnet 5.5, approved: about $16 a month at the median, $10 to $24 across the recorded runs.
Claude Sonnet 5.5, declined: about $14 a month at the median, $11 to $28 across the recorded runs.

Model fees only, at OpenRouter’s prices on 29 Sep 2026, for partner requests like this one. Not included: hosting, integration, and the people who review what the controls hold. Costs are OpenRouter’s billed cost per call, recorded in each signed run file; model time is OpenRouter’s record of each call’s response time, fetched 29 Sep 2026 and signed in timings-2026-09-29T05-31-42-707Z.json.

For technical teams: verify these recordings

Every recording is signed. Checking it shows the files have not changed since QAIZEN signed them: the steps, the model, the date and the costs. It says nothing about when they were recorded. The replay always shows run 01, the first one recorded; every run is published.

claude-opus-5.5, run 01

claude-sonnet-5.5, run 01

Signing key fingerprint (SHA-256): 970159aedc46a8a4be37522e5711f7c21af56d26c5b225f9f603c26a63244da2

claude-opus-5.5: all 30 recorded runs, failed ones included

claude-sonnet-5.5: all 30 recorded runs, failed ones included

Also published: signing-key.pub, timings-2026-09-29T05-31-42-707Z.json and timings-2026-09-29T05-31-42-707Z.sig, timings-2026-09-29T05-47-12-571Z.json and timings-2026-09-29T05-47-12-571Z.sig (response times), system-prompt.partner.en.txt (the agent’s instructions; its SHA-256 is in every run’s header), and verify.mjs, the same checks from a terminal.

The runs play in this browser and your decisions stay in it. If you accept analytics cookies, we count plays, decisions and uses of the summary button, with the scenario, the page language and the recording’s language only. Privacy policy