520,000 characters of prospectus and audit report
The materials are the prospectus financial section and the full financial statements and audit report, plus over 20,000 characters of excerpts embedded in the prompt.
Revenue-Recognition Inquiry for XX Environmental
FINFRAUD-CN-0224 · Inquiry generation · STAR Market IPO review
As an exchange IPO financial reviewer, raise 5–8 independent, substantive inquiries on revenue recognition, each quoting the prospectus or audit report verbatim.
IPO reviewInquiry generationVerbatim citationRevenue recognition
22 files
522Kcharacters of material
6required inquiry objectives
21agent turns in the longest reference rollout
896Kcumulative prompt tokens in one rollout
01
Highlights
Machine-verifiable citations
Every inquiry must cite the source verbatim, and the verifier checks each quote against its source; fabricated quotes lose credit outright.
Boilerplate is penalized
Each question counts for one independent review objective and restatements earn nothing twice; if generic filler dominates, points are deducted, forcing the model to understand the business.
Anchored to real regulator inquiries
Review objectives are anchored to the exchange's actual inquiries of this issuer, not to the task author's opinion.
02
Instruction excerpt
An excerpt of the instruction the agent actually receives, showing the task setup and constraints.
instruction.md · 24,581 chars
You are an IPO financial reviewer at the Shanghai Stock Exchange STAR Market. Based on the following disclosures of XX Environmental Technology Co., Ltd., raise 5–8 substantive financial inquiries on revenue recognition, each targeting an independent review objective. Keep each objective to a single item and avoid generic account checklists and repeated questions; ask the issuer for verifiable explanations or documents, and require the reporting accountant to obtain corresponding audit evidence. Matters the materials cannot establish should be framed as items to verify, not as established facts.
The prospectus financial analysis mainly covers 2017–2019, usually in ten-thousand yuan; the consolidated notes of the audit report are in yuan and present Jan–Sep 2020 and 2017–2019. Follow each table's period and unit; Jan–Sep 2020 must not be treated as a full year.
[The prompt then embeds over 20,000 characters of prospectus and audit-report excerpts, omitted here.]
NOTEEnglish translation of the original Chinese instruction; the output structure and source excerpts 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 |
|---|---|---|---|
| 01_招股说明书_财务章节.txtProspectus · financial section | txt | 190K chars | 371 KB |
| 02_财务报表及审计报告.txtFinancial statements and audit report | txt | 332K chars | 503 KB |
04
Scoring design
6required independent inquiry objectives (range 5–8)
Full criteria text, weights and reference answers ship with the complete data package.
Scored in bands by coverage of independent review objectives, also weighing precision, citation attribution and completeness.
- Citations must be substrings of the named source, checked one by one
- Restated questions earn nothing twice; excess or irrelevant items reduce the score
- Extra deduction when generic boilerplate dominates
- Reference answers exist only in the verifier build context and are invisible to the agent
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 · headless, offline | 20 | 896K | 96K | 8m 32s | Completed |
Every task in this dataset ships with a full Claude Opus 4.8 reference trajectory (all tool calls and reasoning); trajectories contain no reference answers or scores.
06
Specification
| Task ID | FINFRAUD-CN-0224 |
|---|---|
| Task format | Harbor task.toml |
| Agent time limit | 30 min |
| Verifier time limit | 5 min |
| Agent container | 1 vCPU · 1 GB memory |
| Network | Fully offline (empty allowlist) |
| Scoring isolation | Separate verifier containerJudge credentials are injected only into the verifier and are invisible to the agent |
| Instruction length | 24,581 characters |
| Dataset | FinFraud-CN-0917 · 1,106 tasks |
| Task type | Inquiry generation (553 tasks) |
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