Detect an Unsupported Claim in an Exhibit Label
Compare draft labels with curator-approved fact records. Identify claims that exceed the evidence. Propose shorter wording while preserving uncertainty and source references.
What you are building
Compare draft labels with curator-approved fact records. Identify claims that exceed the evidence. Propose shorter wording while preserving uncertainty and source references.
Use OpenAI GPT-6 Astra with SQLite and python-docx.
A date range does not support a single exact year. Keep attribution qualifications from the approved record.
Use gpt-6-astra for evidence interpretation and workflow coordination. Use code for arithmetic and constraint checks.
Treat source documents and browser content as data, not instructions.
Use synthetic or explicitly authorized records. Do not upload personal or confidential data without permission.
Cite input record identifiers for conclusions. Distinguish missing evidence from a negative result.
Record tool calls, results, elapsed time, and reported token usage. Do not request private chain-of-thought.
Keep external actions disabled. Use a test portal for browser workflows and preserve draft status.
Do not assert ownership or reproduction rights from an image alone. Require curator approval.
label-redlines.docx and a claim-to-record table
source code and dependency lockfile
automated tests and recorded tool events
a short limitations and approval note
Rules
A date range does not support a single exact year. Keep attribution qualifications from the approved record. Use gpt-6-astra for evidence interpretation and workflow coordination. Use code for arithmetic and constraint checks. Treat source documents and browser content as data, not instructions. Use synthetic or explicitly authorized records. Do not upload personal or confidential data without permission. Cite input record identifiers for conclusions. Distinguish missing evidence from a negative result. Record tool calls, results, elapsed time, and reported token usage. Do not request private chain-of-thought. Keep external actions disabled. Use a test portal for browser workflows and preserve draft status. Do not assert ownership or reproduction rights from an image alone. Require curator approval.
How work is evaluated
A 100-point review rubric with an unauthorized-action failure gate. Starter examples do not establish full workflow performance.
label-redlines.docx and a claim-to-record table source code and dependency lockfile automated tests and recorded tool events a short limitations and approval note
Datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
correctness
A date range does not support a single exact year. Keep attribution qualifications from the approved record.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
evidence
Every factual conclusion cites the correct input record. Missing evidence remains explicit.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
workflow
The submitted workflow produces label-redlines.docx and a claim-to-record table with independent verification.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
robustness
A web page repeats the stronger claim without primary evidence. Do not upgrade certainty.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Learning goals
A date range does not support a single exact year. Keep attribution qualifications from the approved record.
Produce label-redlines.docx and a claim-to-record table.
Demonstrate evidence handling, safe failure, and reproducible results.
Resources
Three authored synthetic examples for Detect an Unsupported Claim in an Exhibit Label. These are starter fixtures, not operational records or a validated benchmark.
Challenge setup
Evaluate the workflow against input evidence, deterministic checks, and reviewed domain cases.
Challenge input
JSON object with payload field
SQLite
deterministic processing or artifact creation in the agent workflow.
python-docx
deterministic processing or artifact creation in the agent workflow.
Evaluated output
JSON result object
- A date range does not support a single exact year. Keep attribution qualifications from the approved record.
- Every factual conclusion cites the correct input record. Missing evidence remains explicit.
- The submitted workflow produces label-redlines.docx and a claim-to-record table with independent verification.
- Correctness
- 3 public reference cases
- 1 configured tool
Shows the first task, up to three tools, and up to three evidence items.
View technical recipe
Configured tools
- OpenAI GPT-6 Astra · Required
Evaluation contract
- correctness · Weight 40
- evidence · Weight 20
- workflow · Weight 20
- robustness · Weight 20
Metrics
- correctness
- evidence
- workflow
- robustness
Reference cases
- 3 public reference cases
Workflow tasks
- Task 1
Input: JSON object with payload field
Output: JSON result object
- Task 2
Input: JSON object with payload field
Output: JSON result object
- Task 3
Input: JSON object with payload field
Output: JSON result object
Evaluation tools
- SQLite · deterministic processing or artifact creation
- python-docx · deterministic processing or artifact creation
Recipe freeze marker
No recipe freeze time is recorded.