# Follow-up test v2 (pre-registered 2 Nov 2026, before any run)
Why: the 22 Oct follow-up (exhibit 4) tested only ChatGPT, 5 of 14 wrong answers, with leading wording. In the Lyon case our
prompt even stated something false. The owner decided to redo it properly (2 Nov).

Scope: every scored wrong answer in the October consumer-app run (verdict overpay_12_50, overpay_50, no_fit, not_found,
price_not_open): 42 answers (ChatGPT 12, Gemini 16, Grok 14). The El Paso question (us-mexico-amount) is excluded for
everyone, per the audit. Each follow-up is typed into the ORIGINAL conversation, under the same app settings as on 22 Oct.

Prompt (identical for all, neutral: no claim that either plan is right, no price stated):
"Another source suggests this plan for the same need: <plan>, operator page: <url> (checked 22 October 2026). Please check
your recommendation and this plan on the operators' own websites. Which is the cheapest plan that fits my needs as I
described them? Would you change your recommendation? End with a JSON block (```json) with keys: operator, plan,
monthly_price, est_12_month_total, currency, changed (true/false)."
<plan> = the answer key's cheapest fitting plan for that question (sheet.json). For fr-tablet it is Lebara Forfait 50 Go
(owner correction of 2 Nov).

Scoring by code, with the same key and rules (KEY_v3.md):
- **corrected to right:** the new final pick is right.
- **changed, still wrong:** a different pick, still not right.
- **kept:** the same pick.
- **no answer.**
The new pick's verdict is reported as is, including when the assistant finds an even cheaper valid plan (then it's right).
Results: followup_v2/results.json, plus the report's exhibit 4 rewritten from them. The first full run is the result and
isn't redone.
Sheet: followup_v2/sheet.json (conversation link + exact prompt per answer).

## Deviation (2 Nov, before any run)
us-heavy and us-verizon-unlimited are held: the key's cheapest (US Mobile Unlimited Starter Annual) needs promo code
ONLY199, which the rules don't count, and for us-verizon-unlimited it runs on T-Mobile's network (doesn't fit "Verizon").
These 6 follow-ups run after the key is corrected. Plan names in the prompt are the plan's name without our notes.
Conversations are continued from each answer's share link ("continue" creates a copy in the same account, same context).

## Deviation update (2 Nov, after the first 36, before the 6 held ones ran)
- Key check (KEY_CHECKS.md §5): US Mobile Unlimited Starter Annual is a no-code new-line price and is eligible on all three
  networks, so the key for us-heavy and us-verizon-unlimited is unchanged and the 6 held follow-ups run with the
  original prompts.
- Key corrections applied before final scoring (KEY_CHECKS.md): ES simyo EUR 3 add-on line excluded (es-light re-keyed);
  fr-tablet excluded for everyone (owner decision 2 Nov). Answers that became right under the corrected key leave the
  follow-up set ("dropped").
- ChatGPT follow-ups ran on the account's current plan (Plus; reasoning "Gemiddeld"); the 22 Oct run was on Pro.
- r10-us-verizon-unlimited excluded for everyone (owner decision 2 Nov, after all 42 ran; KEY_CHECKS.md §6): a possibly
  cheaper Visible offer ended before we could verify it on Visible's own page.