The width of the band is the test
The task does not ask for one number: it asks for a floor, a ceiling and where the spread comes from. The band has to follow from the disagreements in the material.
Long-Lived Assets of a Clinic Chain
YKCX-FIN-MEMO-939 · Asset-side forecast · capex and depreciation · flip point
The sell-side material covers only the asset side. From an 18-worksheet financial model, reconcile total new long-lived asset spend, put a floor and a ceiling on the closing balance and explain why the band is that wide, take a stance, and compute what would flip it.
Long-lived assetsDepreciationFlip pointAdoption stance
11 files · 18 worksheets
18worksheets
17scoring criteria
42agent turns in the longest reference rollout
2.53Mcumulative prompt tokens in one rollout
01
Highlights
Deliver the view with its fragility
After taking an adopt / adopt-with-reservations / reject stance, state the asset-consumption level at which the stance flips, or why it does not flip.
One sentence for a non-specialist
Beyond the workbook and the structured JSON, the memo must tell the client in one sentence which element matters most to the conclusion, and why.
Three-model mean score 0.4590
Three models answered under the same criteria for a mean of 0.4590, all below 0.52. The bar here is not arithmetic but explaining where the width of the band comes from.
02
Instruction excerpt
An excerpt of the instruction the agent actually receives, showing the task setup and constraints.
instruction.md · 2,307 chars
You are an investment adviser in a private bank's wealth management department. A client has received sell-side forecast material on a clinic chain that is opening new sites and wants an independent view on the asset-side forecast before adjusting the position.
Task
- Take an adoption stance — adopt directly, adopt with reservations, or reject — and explain the decision and the scope you keep or drop.
- Give total new long-lived asset spend across every forecast year, and a floor and ceiling for the closing balance in the final forecast year, explaining the width of the band.
- Give the asset consumption you accept for the final year, its gap against the stated figure, the effect of that gap on your stance, and the flip point if the stance would change.
- Tell the client in one sentence which element most affects the conclusion, and why.
NOTEEnglish translation of the original Chinese instruction; the block structure and field definitions of the three deliverables are omitted.
03
Attachments
Everything visible inside the agent container, with statistics generated from the task package. Original files ship with the full sample package.
| File | Type | Size | Bytes |
|---|---|---|---|
| 中弘智造_底稿.xlsxFinancial forecast and valuation model: fixed assets and long-term prepaid expense forecasts, financial statements, source data, assumptions (18 worksheets) | xlsx | 18 sheets | 357 KB |
04
Scoring design
17independent scoring criteria
16 scoring items1 penalty items
Full criteria text, weights and reference answers ship with the complete data package.
17 criteria: 11 binary plus 6 five-level items, with the heaviest weight on numerical accuracy.
- 17 criteria: 16 positive scoring items and 1 penalty items (negate)
- Each criterion is judged separately by an agent judge (individual mode), reproducible and auditable
- Judge credentials are injected only into the verifier; the agent container has no scoring channel
- Full criteria text and reference answers ship with the complete data package
Numerical accuracy5
Conclusion correctness3
Argument quality3
Professional standards2
Factual faithfulness2
Instruction following1
Compliance1
05
Model results
Trajectory statistics from frontier models in real agent harnesses. One turn is one model call; the context curve shows how state accumulates over a long-horizon task.
| Model | Agent turns | Tool calls | Cumulative prompt tokens | Peak context | Time | Result |
|---|---|---|---|---|---|---|
| Claude Opus 4.8Claude Code · Harbor | 18 | 1.53M | 107K | — | Completed | |
| GPT-5.6-solCodex · Harbor | 27 | 1.41M | 68K | — | Completed | |
| Qwen3.8-MaxClaude Code · Harbor | 43 | 2.53M | 122K | — | Completed |
06
Specification
| Task ID | YKCX-FIN-MEMO-939 |
|---|---|
| Task format | Harbor task.toml · schema 1.4 |
| Package version | 1.0.0 |
| Agent time limit | 3 h |
| Verifier time limit | 5 h |
| Agent container | 2 vCPU · 8 GB memory · 30 GB storage |
| Network | Only for agent harness installation |
| Tool set | filesystem · shell · python |
| Scoring isolation | Separate verifier containerJudge credentials are injected only into the verifier and are invisible to the agent |
| Instruction length | 2,307 characters |
| Difficulty | A3 (three-model mean 0.4590) |
| Reference solution | oracle reward 1.000 |
| Workpaper size | 18 worksheets · 2008–2012 history and 2013E–2017E forecast |
Contact
Want to run this task on your own platform?
Leave your details and we will send the full sample package for this task (instruction, attachments, verifier and reference trajectories) and walk you through volume and customisation for similar tasks.
- Full task package and verifier
- More tasks in this subdomain
- Integration and re-run support