Keep the denominator in an Orkas survey report

“Half the group wants the workshop” sounds like a useful planning result. It can also be the wrong sentence even when the division is correct. If three people said yes, three gave another answer, and two never replied, half describes the respondents. It does not describe everyone invited.

Before turning a small spreadsheet into a polished public page, make the population behind each number visible. A desktop workspace such as Orkas can help organize the source, analysis and writing tasks. The human reviewing the result still needs to check which rows each sentence counts.

One source table, two valid percentages

Imagine a team choosing whether to run an optional spreadsheet workshop. Eight people received an invitation. The question was: “Would you attend this session?” A blank cell means no reply has been recorded; it is not a no.

The counts reconcile: three yes, one maybe, two no and two unanswered total eight invitations. There are six recorded replies. That produces three useful ratios:

• 3 / 6 = 50% of respondents answered yes.
• 3 / 8 = 37.5% of invitees have a recorded yes.
• 2 / 8 = 25% of invitations have no recorded reply.

These are descriptions of this register. They do not tell us what the two missing respondents think, whether a maybe will become a booking, or how a larger audience would behave. A report can preserve those unknowns without becoming vague.

Give Commander an output contract

Orkas is an open-source, local-first desktop AI workspace. Commander plans work and coordinates specialist agents. For this task, a proposed division is to ask OfficeWorker to inspect the spreadsheet and reconcile the counts, then ask ContentWriter to draft a short explanation from the checked table.

Give that team a narrow brief. If you put the source in the project's Library, select that source for the task and keep the answer labels explicit. Do not rely on an earlier conversation to explain what an empty cell means.

Input: the eight-row invitation register above.
Blank means no recorded reply; never convert it to No.
Return: counts for Yes, Maybe, No and Blank;
        total invitations and total recorded replies;
        every percentage as numerator / denominator;
        a short summary and an unresolved-questions list.
Do not estimate the preferences of people who did not reply.
Do not turn interest into confirmed attendance.

This is a request the reviewer can assess, not a promise that a model will follow it perfectly. Keeping the calculation and the prose as separate deliverables makes a misplaced denominator easier to find. An impressive paragraph alone provides no way to inspect its row selection.

Review a claim beside its evidence

Before publishing, make a small claim table. The final column records the actual limit of the evidence, rather than another reassuring adjective.

The reviewer can now trace each sentence back to the source. If the draft says “50% of invitees,” revise that sentence. If it calls the two blanks refusals, correct the interpretation and recalculate the affected claims. Do not silently edit the source to make a generated sentence look right.

A publishable summary for this example is:

Six of eight invitees have replied. Three of those six respondents answered yes (50%); one answered maybe and two answered no. Across all eight invitations, recorded yes answers represent 37.5%. Two invitations remain unanswered. The register shows interest, not confirmed attendance.

Publish only the material the reader needs

A public HTML page can contain the summary, the fictional example table and the explanation of its limits. A real invitation register may contain names or contact details. Local storage does not make those details safe to publish. Prepare a separate public version, and check its title, tables, ending and links before sending it to a hosting service.

Orkas keeps project materials local by default, while model requests and optional connected or cloud services can send the data needed for those features. Its security explanation describes that boundary. Choose the material and model route accordingly. The core app is free; model usage and optional services may cost extra.

ShipPage is a separate hosting service, not an integration demonstrated here. Its documented API accepts HTML and can list a page publicly when public is true. That setting makes the prepared page discoverable; it does not perform the data review for you.

The useful handoff is a report whose numbers a reader can reconstruct. Keep the source and reviewed copy locally, and let each percentage carry the group it actually describes.