# Qwen 4 real-world work test kit — version 1.2

Prepared November 15, 2026 (America/Vancouver). Execution is authorized under a cumulative US$125 inference cap. The delivered guide and evidence index report actual outcomes and costs.
All business records, people, products, companies, prices and source documents in
the exercises are synthetic. The tasks simulate real work; they are not actual
Kingy financial records or product recommendations. Task 6 images are staged UI
screenshots, not evidence from a commercial product.

## Run one assignment
1. Read operator/PROTOCOL.md. Choose the exact endpoint/version for each model
   family: Qwen 4, Astra, Sol, Opus, Fable. The executed identifiers/providers are
   recorded in operator/EXECUTION-SETTINGS.md; recheck availability before reuse.
   Extracting this package makes no API calls.
2. Create a fresh isolated workspace containing ONLY candidate/COMMON.md and ONE
   candidate/task-NN/ directory. Keep operator/, build/, grading files, other models'
   outputs and answer keys outside its filesystem and context. A zip alone is not
   a security boundary if the agent can browse the host.
3. Attach COMMON.md and that task's PROMPT.md plus its inputs. Paste the prompt
   exactly; use the same files for every model. Do not expose any answer key.
4. Save the full response, generated files, timing and tool trace. Stop the timer
   at completion or at the common 30-minute limit. Grade the original before revision.
5. Send operator/REVISION-PROMPT.txt verbatim in the same session, allowing 15 more
   minutes. Save and grade separately. Do not supply hidden answers or custom hints.
6. Use operator/grading.xlsx (Results, Rubric, Setup) and run-log.csv. Repeat in a
   fresh workspace/session for repetition 2. Planned count: 10 tasks × 5 models × 2
   repetitions = 100 initial runs, with up to 100 revision stages.

## What to open
- candidate/task-01: messy sales, refunds and ad exports.
- candidate/task-02: fixed purchase research packet with conflicting offers.
- candidate/task-03: runnable broken website (Python standard library, no npm install).
- candidate/task-04: emails, notes and a change request.
- candidate/task-05: fictional AI-release source packet and editorial brief.
- candidate/task-06: three UI screenshots and short product help.
- operator/answer-keys/: private ground truth for each assignment.
- operator/RUBRIC.md: exact scoring and critical failures.
- operator/run-order.csv: deterministic, rotated run order.

## Packages
candidate-inputs.zip contains only model-facing material. full-operator-kit.zip
contains the complete kit including answers and grading files; never give it to a model.
Tasks 7–10 add customer support, scheduling, invoice images and a long archive.
Task 2 is a controlled research-synthesis test. Optional live-web research is described
separately and must not be pooled with its scores.
