Turn model drafts into reviewable state and files.
Connect model extraction, validation, source review, and CSV export through a seven-module contract.
Product definition, output design, and implementation · AI-assisted codingExpert leaderboard runner-up · Popularity leaderboard winner
2026 iFlytek AI Developer Competition, Intelligent Office Collaboration Skill Challenge, with 1,851 teams in the track.
View the competition page ↗One-minute example: separate requests from decisions
Illustrative scenario: a payment-page launch. These excerpts come from bundled simulated materials and a saved offline replay, with no real team information.
Chat: “The new payment page needs to launch this week—ideally a full rollout on Friday, or the campaign cannot proceed.”
Notes: “Confirmed: only a 20% rollout on Friday, not a full rollout.”
The minutes record a 20% rollout; the chat expresses a request for a full launch. The next step is to clarify whether the campaign requires a full launch, rather than combine these into a full-launch commitment.
When sources conflict, keep both references and verify timing, who confirmed the plan, and subsequent changes. Minutes do not automatically outrank chat messages.
View the full simulated input (Chinese)Engineering decisions: what I chose
- Fix the data contract
- The host model drafts seven modules. Python handles structure, unknown values, and bounded rules, producing card and meta review records.
- Preserve unknowns and sources
- Missing owners and deadlines remain “Not provided”. Citations support review; negation, conditions, and responsibility still need human checks.
- Use regressions to pin down behavior
- This page shows 18 existing targeted regressions, status cards, and CSV from the same revision. Test counts do not measure model accuracy or host compatibility.
Three actions in the saved card
The items below are excerpts from the saved reviewed JSON. Completion status is unspecified for all three; owners and deadlines follow the source material.
- Luo Sen · By end of day
- Provide the first backend field definitions
View source excerpt
Luo Sen will provide the first backend field definitions by end of day.
- Xu Yan · Before Thursday, 12:00
- Send operations the final payment-page screenshots
View source excerpt
Xu Yan will send operations the final payment-page screenshots before Thursday, 12:00.
- Meng Qiao · Not provided
- Integrate the backend fields
View source excerpt
Meng Qiao will integrate the backend fields; the deadline remains to be confirmed.
“I'll see if I can” is not a confirmed deadline
Chat: “I'll see whether I can finish the integration today, but the backend fields are not finalized yet.”
The saved integration deadline remains unspecified. Ask “Once the fields are confirmed, when do you expect integration to finish?” rather than assume completion today.
Return the result to the existing workflow
The same card exports to CSV with actions, decisions, risks, and open questions. This case demonstrates file export; Feishu and Notion field mappings do not demonstrate automatic platform writes.
Platform evaluation: seven passed, three lacked inputs
Of ten cases, seven passed; three were not evaluated because input files were missing from the environment. Cases cover extraction, aggregation, conflict boundaries, and masking.
This evaluation exposed an input-preparation problem. Next, check file availability, supply missing inputs, and retest to reduce unproductive searches.
View the full platform report ↗Check the output, then observe reuse
Next, observe whether a coordinator can complete the task, find omissions, and reuse the tool on another eligible occasion. Record which item they acted on, what changed, and why they sometimes did not use it.
Implementation and existing checks
Model extraction, structural validation, and human review have distinct responsibilities. Eighteen focused regressions passed, checking explicit program behavior in this revised version.
Exported files are opened in spreadsheet or document tools. Engineering acceptance should check unexpected inputs, content escaping, and export consistency, separately from semantic quality.
Data and evidence notes
Sources, version, and scope
Examples use bundled simulated material and the revised offline replay saved during presentation preparation, rather than a new model call or evidence of customer outcomes.
Source references support review; meaning, confirming parties, and commitments still need human verification. Structural validation checks formats and rules, without guaranteeing factual or semantic correctness.
Competition account: jamji; in-person awards on October 24, 2026. Platform evaluations and local regressions have their own models, materials, and versions.
← Back to selected cases · Next case: annual-report delivery →