TERMIX · HUMAN-VERIFIABLE COMPARISON

Three real comparisons.
Do the work. Keep the evidence.

Enter your name, then start the timer to reveal the task. A calculator is allowed; AI is not. Reading, calculations and writing all count. Download each record before starting the next task.

LOCAL ONLYDrafts are saved when possible. Download the finished record. Timing recovered after a refresh or reopening requires review and does not automatically qualify as a control.

01 · MANUAL CONTROL

Complete the same task without an agent.

NOT STARTED

Task Participant

Start the timer to reveal the task

Loading…

00:00.0Reading, calculations and writing are timed. The timer cannot be paused.

Complete answer: calculations, conclusion and limitations

Actual cost Currency

Tools actually used I completed this task myself, without AI or a marketplace agent, and preserved the full answer.

Used AI or keeping an unfinished attempt? Save it as assisted practice. It will not count as an unaided control.

02 · BLIND PACKET

Hide which output used the Agent.

PREPARER

The preparer loads both raw JSON files. The page extracts answer content, removes transport metadata, and downloads a separate secret key. Answer wording can still reveal its origin: inspect the packet and disclose any remaining clues. Original files must be preserved. Do not send the key to the reviewer.

Agent raw output No-Agent raw output

03 · INDEPENDENT REVIEW

Score A and B before unblinding.

REVIEWER

Ideally, a person who did not prepare the files should review them. Load only the blind packet, never the secret key.

Blind packet Reviewer name

Reasons and limitations of these scores I reviewed both complete answers before seeing the separate mapping key.

What still needs a person

One operator must perform each manual run, a different reviewer should score the blind packet, and you must keep the secret mapping, outputs and review together for submission.

Three real comparisons.Do the work. Keep the evidence.

Complete the same task without an agent.

Hide which output used the Agent.

Score A and B before unblinding.

Three real comparisons.
Do the work. Keep the evidence.