ChatGPT Work Report Looks Polished but Unverifiable? Build a Claim Ledger First

What you will make

You will make a claim ledger for a fictional weekly operations report. Each row will connect one report statement to its source note, show whether the wording changed, and mark what a person must check.

This is a troubleshooting guide for a common failure: the report looks finished, but nobody can tell which figures came from the source material. The inputs and output below are illustrative. No ChatGPT Work execution is claimed.

Check your access first

Open ChatGPT Work on the web or desktop surface available to your account. OpenAI describes Work as an agent for longer, multi-step work and finished deliverables. Its current file-creation guidance says you can attach source material, name the output format, state what must stay unchanged, and review the result before sharing. Availability depends on your plan, workspace settings, file type, and surface.

You also need permission to upload the files. Use fictional or approved material only. For this exercise, create these two text files.

source-notes.txt:

Weekly operations notes, 12 October 2026

1. The support team closed 42 tickets this week. The source system export was last checked at 16:00 on Friday.
2. Median first response time was 3 hours 20 minutes. The report owner said this figure excludes the two tickets marked waiting for customer.
3. The backlog was 18 tickets at Friday 16:00. Four tickets had no assigned owner.
4. The planned training session moved from Tuesday 13 October to Thursday 15 October.
5. The team lead asked for a short explanation of why the backlog rose from 11 tickets last week.
6. No cause for the increase was confirmed in the notes.

weekly-report-draft.txt:

Weekly operations report

The support team closed 42 tickets and kept median first response time to 3 hours 20 minutes. The backlog increased to 18 tickets, mainly because of a busy week. Training is scheduled for Thursday 15 October.

The phrase “mainly because of a busy week” is deliberate. The source notes do not support it. The report also leaves out the four unassigned tickets and the condition attached to response time.

Three steps

1. Attach both files and request a ledger

In a new ChatGPT Work task, attach source-notes.txt and weekly-report-draft.txt. Use this prompt:

Create a claim ledger for weekly-report-draft.txt using only source-notes.txt as evidence.

Return a table with these columns:
- report claim
- matching source quote
- source date or time
- status: supported, changed, unsupported, or missing
- review action

Rules:
- Copy each claim from weekly-report-draft.txt before evaluating it.
- Match every number, unit, condition, and date exactly.
- Do not treat a plausible explanation as evidence.
- Mark a claim unsupported when source-notes.txt does not state it.
- Add rows for material facts in source-notes.txt that the draft omits, including unassigned tickets and the response-time condition.
- Do not rewrite the report yet.
- If a source does not answer the question, write "not stated".

After the table, list the three highest-risk corrections and explain which source sentence supports each one.

2. Inspect the risky rows

Read the ledger against both files. Pay attention to the numbers, the phrase “mainly because of a busy week,” the training date, the response-time condition, and the four unassigned tickets.

Then use this follow-up prompt:

Audit the claim ledger against both attached files.

For every row, return PASS only when the claim, number, date, condition, and source quote match. Return CHECK when the row is unsupported, incomplete, inferred, or missing a material qualifier. List any sentence from source-notes.txt that the ledger did not cover.

Do not improve the report and do not invent a cause for the backlog change.

The audit is a review aid. It does not prove that the source notes are accurate.

3. Produce a corrected draft only after review

If the ledger and audit are accurate, ask for a clean draft that keeps the gaps visible:

Using only rows marked PASS, plus the source facts that a reviewer has confirmed, rewrite weekly-report-draft.txt as a short internal update.

Keep these rules:
- Keep 42 closed tickets, median response time of 3 hours 20 minutes, and the condition that the figure excludes two tickets waiting for customer.
- State that the backlog was 18 at Friday 16:00 and that four tickets had no assigned owner.
- State that the training moved to Thursday 15 October.
- Say that the cause of the increase was not confirmed. Do not write "mainly because of a busy week."
- Do not add a recommendation, owner, deadline, or performance judgement.
- After the draft, include a short "Needs confirmation" list.

Illustrative output

Report claim	Matching source quote	Source date or time	Status	Review action
42 tickets closed	“The support team closed 42 tickets this week.”	12 October 2026	Supported	Keep the number and quote
Median first response time was 3 hours 20 minutes	“Median first response time was 3 hours 20 minutes. The report owner said this figure excludes the two tickets marked waiting for customer.”	12 October 2026	Changed	Keep the waiting-for-customer condition with the figure
The response-time figure excludes two tickets marked waiting for customer	“The report owner said this figure excludes the two tickets marked waiting for customer.”	12 October 2026	Missing	Add the qualifier to the report
Backlog was 18 at Friday 16:00	“The backlog was 18 tickets at Friday 16:00.”	Friday 16:00	Supported	Keep the timestamp
Backlog rose mainly because of a busy week	“No cause for the increase was confirmed in the notes.”	12 October 2026	Unsupported	Remove the explanation and keep the cause unconfirmed
Training is Thursday 15 October	“The planned training session moved from Tuesday 13 October to Thursday 15 October.”	Tuesday 13 October; Thursday 15 October	Supported	Keep the moved date
Four tickets had no assigned owner	“Four tickets had no assigned owner.”	Friday 16:00	Missing	Add the issue to the review list

This table is editor-written and illustrative. It is not output from ChatGPT Work.

Acceptance checks

The workflow passes when:

Every claim in the report has a matching source quote or is marked unsupported.
The response-time condition remains attached to the 3 hours 20 minutes figure.
The backlog timestamp is not replaced with a vague reference to this week.
The unsupported cause is removed rather than softened into a guess.
The four unassigned tickets appear in the ledger or review list.
The training date is Thursday 15 October, not Tuesday 13 October.
The original two files remain unchanged.

If something goes wrong

ChatGPT marks the backlog cause as supported: ask it to quote the exact source sentence. The notes contain no confirmed cause, so the row should be CHECK.
The response-time condition disappears: upload the complete source notes again and require the quote and qualifier in the same row.
The ledger invents a report date: mark it “not stated” and do not add a date to the draft.
The output becomes a general summary: repeat the required table columns and ask it to stop before rewriting.

Limits and privacy

Work can create and edit files from instructions and source material, but the available file types, controls, and integrations vary. Upload only material your organisation permits. A claim ledger checks traceability. It does not validate the underlying business data or approve the report for distribution.

Evidence label

Documentation-based guide with complete fictional text files and an illustrative ledger. OpenAI documentation was checked on 12 October 2026. No ChatGPT Work execution, personal test, measured result, or business outcome is claimed.

Next lesson

For a broader checkpoint workflow, read Use ChatGPT Work to Keep a Multi-Step Task From Drifting.

Want one practical workflow each week? Join The Weekly Verdict.