# Editor notes: LumenDesk 2.0 article (revision 2)

All entities are fictional test fixtures. The article uses only the files in `inputs/`. No browsing, no hands-on use of the product, and no contact with the vendor, DeskLab or any customer took place (editorial-brief.md, "Assignment").

The first submission is preserved unchanged in `output/original-submission/`. Changes since then are listed in `output/CHANGELOG.md`.

## Decisions an editor should confirm

1. **Lead with access, not price.** The assistant is open only to existing Team customers in Canada and the United States; new customers get a waitlist (S1-announcement.md, "Access"). For a reader deciding whether to try it now, that is the first filter, so it leads.
2. **USD 29 is the only price used as current.** USD 19 appears once, as a corrected error. See "Corrections reconciled".
3. **First-person "I" is used for advice and for my own arithmetic only** ("I would start with documents whose answers you already know"), following voice-sample.md. It never describes testing. The italic note under the headline says neither the author nor Kingy has used the product.
4. **No publication date is given in the article.** The packet does not supply one. The italic note dates the sources (October 16–18, 2026) instead.

## Corrections reconciled

| Point | Conflict | Resolution in article |
|---|---|---|
| Seat price | A launch-day screenshot showed USD 19; S5 suggests leading with USD 19 | S2-pricing.md, "Correction" (updated October 17) says the screenshot was incorrect and USD 29 is current. Article uses USD 29 and reports the correction. |
| Offline use | S5 says it "runs fully offline" | S3-data-policy.md, "Processing" says an external model provider processes requests and there is no on-device mode. Article follows S3. |
| Time saved | S5 says it "saves every team 60% of their work time" | S1-announcement.md, "Vendor test" reports only median first-draft time on 40 packets (50 to 20 minutes), excludes verification and says it does not establish end-to-end time saved. Article reports the vendor figure with those caveats, attributed to the vendor. |

The screenshot itself is not in the packet. The article relies on S2's description of it.

## Claims kept out of the article

All are from S5-internal-notes.txt, which labels itself "Unverified writer notes" and "leads requiring verification".

| Claim in S5 | Why excluded | What would be needed to use it |
|---|---|---|
| "Eliminates hallucinations" | No source supports it. S4-independent-review.md, "Method" records one wrong page citation in six documents, and S1 "Features" says the tool drafts answers. | Published accuracy evidence with method and results. |
| "GDPR certified" | No source supports it. S3 "Processing" says the policy does not claim guaranteed regulatory compliance, and S1 "Access" lists EU availability as only planned. | A named certification or attestation document from the vendor. |
| "Runs fully offline" | Contradicted by S3 "Processing". | Not usable unless S3 is superseded. |
| "Saves every team 60% of their work time" | Not supported; see table above. | Task-level results including verification time, ideally from outside the vendor. |
| "Lead with USD19 if punchier" | Contradicted by S2 "Correction". | Not usable. |
| Customer quote ("We fired our paperwork department") | S5 itself says there is no named source and no permission. | A named, consenting source and confirmation of the wording. |

Inference, not established: 50 to 20 minutes is a 60% fall in the vendor's median first-draft time ((50 − 20) / 50). The "60%" in S5 may be a garbled version of that figure. Nothing in the packet confirms the link, and the article does not use a percentage.

## Calculations in the article (mine, not the vendor's)

- Minimum monthly cost: 3 seats × USD 29 = USD 87 before tax (inputs from S2 "Current plans").
- With OCR add-on: USD 87 + USD 12 = USD 99 before tax.
- Overage example: 800 uploads − 500 included = 300 extra × USD 0.08 = USD 24. The 800 figure is an illustration, not a source figure.

The article marks the first as "by my arithmetic" and the third as "which I make". The citation after each points to the source of the inputs, not of the result.

## Assumptions and gaps in the sources

- **Is the assistant included in the Team price?** S2 lists Team pricing with 500 document uploads included, and S1 offers the assistant to existing Team customers, but neither says in so many words that there is no separate charge for the assistant. The article lists the Team price and avoids saying the assistant is "free" or "included at no extra cost".
- **Can a new customer buy Team now and gain access?** Not resolved. S1 says only that new customers can join a waitlist. The article says the announcement offers new customers a waitlist, not access, and goes no further.
- **Existing customers outside Canada and the US, or on other plans:** not addressed in the packet. The article does not speculate.
- **Waitlist timing:** S1 gives none. The article says so.
- **"So readers have nothing to check the figure against"** is my reading of S1's "No raw task-level results are published". Cut the clause if the editor wants only the vendor's wording.
- **"Both tests also used only procurement documents, so neither covers other kinds of paperwork"** is my observation from S1 "Vendor test" (40 procurement packets) and S4 "Method" (six procurement PDFs). Neither source draws that conclusion itself.
- **"The vendor's pricing document"** is my label for S2-pricing.md, titled "LumenDesk Team pricing — updated October 17, 2026". The packet does not say whether it is a web page, a PDF or something else.
- **OCR quality** is untested in any source. S4 says it did not test scans or OCR. The article says the outside review did not cover it and makes no quality claim.
- **Taxes and currency:** prices are USD and exclude tax (S2 "Current plans"). No CAD price is in the packet.
- **DeskLab's independence** is taken from the filename and title of S4 only. The packet gives no detail on who DeskLab is or its relationship to the vendor. The article calls it "the one outside source".

## Checks performed on revision 2

Script: `output/checks/check_v2.py`, run after the final edit; every line reported PASS.

- Article word count: 777 for the whole file; 764 without the H1; 729 without any headings; 716 without headings and citations; 695 also without the italic source note; 789 on a stricter count that splits hyphenated and dashed terms. All fall within 600–800.
- Headline: 61 characters (limit 50–70). Matches the article's single H1 exactly.
- Excerpt: 40 words (limit 45).
- Search description: 152 characters (limit 140–160).
- metadata.json parses as JSON and has exactly three keys: headline, excerpt, search_description.
- All ten citations (eight distinct filename and heading pairs) point to `##` headings that exist in the named input files.
- Every body paragraph containing a figure carries a citation, except the italic source note and the closing recap ("A trial makes sense…" and the bullet list), which repeat figures cited earlier.
- Every numeral in the article and metadata either appears in S1–S4 or is one of the declared derived values (87, 99, 800, 300). The script matches numerals only, so it cannot tell "24 hours" from "USD 24"; the USD 24 figure is covered by the arithmetic check instead.
- Arithmetic recomputed: 3 × 29 = 87; 87 + 12 = 99; 300 × 0.08 = 24.
- Article and metadata contain none of: "GDPR", "offline", "hallucinat-", "60%", "certified", "fired", "hands-on", or phrases claiming testing. "USD 19" appears once, in the correction paragraph, and not in the metadata.
- The original submission files exist in `output/original-submission/`.
- By reading, not by script: I re-read the revised article sentence by sentence against S1–S4 and found no statement that goes beyond its cited section, other than the labelled inferences listed above.

## Not done

- No independent fact-check beyond the packet; nothing in S1–S4 was verified against an outside source.
- No house-style check beyond voice-sample.md and the brief; no style guide was supplied.
- No legal review of the data-handling paragraph.
- No rendering check of the Markdown in a publishing system.
- Raw inputs were not modified (the tools cannot edit them).
