Spot the AI's mistake: puzzle #4 (draft)¶

Drafted 7 Oct 2026 for issue #227. Nothing has been posted. This was puzzle #1 until the owner moved it back (2026-10-07): its question is the older drone wording WITHOUT "My radio puts out 0 dBm" (added 2026-09-27, 9e90befd0), and the owner wants transmit-chain questions to state that input. The mistake itself does not depend on the input level. #1 is now the missing negative gate supply. Kept ready for a later week.

The owner's bar (2026-10-07): a puzzle must show an electrical mistake that survives web search, and a reader must be able to see it from the images alone. This one comes from a regular claude.ai chat that searched the web before answering. Puzzles are an extra weekly post: the puzzle midweek, its answer on Friday (see "Weekly puzzle" in marketing-block-1.md).

Source: C:/Users/hakic/magicon-mcp-ab/chats/run_link_budget/2026-09-22/try_4.md (outside the repo, so keep that folder), saved 2026-09-23; how it was asked is in notes.md there. A regular claude.ai chat, without MagicON, model Opus 5.5 "Extra" per the owner. That makes it "bare Claude", the owner's rule for puzzles (2026-10-07); the notes record the owner's words "another one without magicon". Whether claude.ai memory was on is not recorded. The answer's header, as pasted: "Searched the web, read 2 pages". The question is scripts/testing/mcp_ab/prompt_link_budget_designed.txt as committed in b3a88a81d. That file was reworded twice later, so the current file is not the question this chat got.

Decided: - The post says "an AI chat" and does not name the model. - No MagicON comparison. - The drone question has not been posted (posted-log.md); the coupler mistake is new.

Images (ready), built by docs/scripts/make-spot-the-ai-mistake-04.py: - Post 1, image 1 of 3: spot-the-ai-mistake-04-card.png. The puzzle: the chat's exact 1M803 line above TTM's four datasheet figures (no verdict), and "which number is wrong?". Added because, without the maker's figures, a reader could only find the mistake by looking up every datasheet. - Post 1, image 2 of 3: spot-the-ai-mistake-04-puzzle.png. The exact question, the chat's chain table with every cell in full (only the "Running gain" column left out, said on the image), and all five of its "where the numbers come from" lines, each with the source label the chat gave. - Post 1, image 3 of 3: spot-the-ai-mistake-04-results.png. The rest of the answer word for word: output power and P1dB, amplifier back-off, chain OIP3 with its working, typical vs minimum, OFDM, trace loss, supply and heat (try_4.md lines 12, 31, 33-41, 44, 46, 54-57). - ⛔ Why the image shows that much (2026-10-07): an earlier version cut the split row to "−0.25 extra …" (the chat wrote "−0.25 extra (beyond the 3 dB split)") and dropped the "Level after" column (−0.6 dBm transceiver setting, +8.0 dBm per arm, +28.0 dBm per PA). Two other AIs were shown that version. One "found" a missing 3 dB split and an over-driven AD9364, the other omissions the answer had covered; their own numbers were also wrong, checked with MagicON's link-budget engine. The cuts were ours, not the AI's. Never shorten a table cell again. Neither reviewer flagged the 60 W rating. A third pass then said the chain OIP3 was "never given": the answer gives 44.8 dBm (MagicON's run_link_budget: 44.77) and about 43.0 dBm with the transceiver (engine: 42.95). Hence image 2: show the AI's whole answer, not our selection. - Post 2: spot-the-ai-mistake-04-answer.png. The chat's 1M803 line next to TTM's datasheet.

Post 1: the puzzle¶

🔍 Spot the AI's mistake #4

I asked an AI chat, with web search on, to design a 5.8 GHz drone video transmitter: 1 W out, from a 12 V supply, every part on the board, with a full link budget.

It searched the web, picked the parts, and listed where its numbers came from.

The first image shows what it wrote about one part, next to the maker's datasheet. One of its four numbers is wrong, and it is the kind of number that decides whether a part survives. Which one? Its whole answer is in the next two images.

Answers on Friday.

RFEngineering #ElectronicsEngineering #EngineeringStudents #AI¶

[Images: spot-the-ai-mistake-04-card.png, spot-the-ai-mistake-04-puzzle.png, then spot-the-ai-mistake-04-results.png]

Post 2: the answer key (Friday)¶

🔍 Spot the AI's mistake #4: the answer

The coupler is rated 20 W, not 60 W. The AI described TTM's 1M803 hybrid coupler as having "0.25 dB maximum loss, 20 dB minimum isolation and 60 W average rating". TTM's own datasheet (Rev H, page 1) matches the band, the loss and the isolation. Its power column, "Ave. CW Watts", says 20.

So the AI searched the web and got three of the coupler's four numbers right. The one it got wrong is the rating that tells you whether the part survives, and it was 3× too high.

The moral: web search makes an AI's numbers better, not trustworthy. Check every power rating on the maker's own datasheet before you put real watts through a part.

(Setup: a regular AI chat with web search on, Sep 2026; the answer's header said "Searched the web, read 2 pages". Only the coupler line was checked here.)

RFEngineering #ElectronicsEngineering #EngineeringStudents #AI¶

[Image: spot-the-ai-mistake-04-answer.png]

Notes (not for posting)¶

The datasheet: TTM 1M803 (also sold as 1M803S), Rev H, page 1, from https://cdn.ttm.com/repository/products/wireless-xinger/3db-hybrid-couplers/1M803S/1M803S.pdf, read 2026-10-07. Frequency rows "5.0 – 6.0" and "4.8 – 6.0" GHz; "Isolation dB Min 20"; "Insertion Loss dB Max 0.25"; "Power Ave. CW Watts 20"; ΘJC 78.1 °C/Watt. The same rating was found by every web-on answer to the 40 W combiner question (run C:/Users/Public/magicon-bare/2026-10-07_183743: all 3 named the 1M803S as 20 W and rejected it).
The chat: try_4.md line 24, verbatim: "1M803 hybrid: a 5–6 GHz surface-mount hybrid with 0.25 dB maximum loss, 20 dB minimum isolation and 60 W average rating." The next line is its source label, "Microwave Journal". What that article says was NOT checked, so the post does not blame or name it. ⚠ Where the 60 W came from (read 2026-10-08): an old Microwave Journal "New Products" notice, https://www.microwavejournal.com/articles/3045-new-products, dated September 1, 2000 (read in the in-app browser; a plain fetch returned 403). Its Anaren Xinger 1M803 item says, word for word: "Average power is 60 W (max)." Its other figures match the chat too (20 dB minimum isolation, 0.25 dB maximum insertion loss). So the AI did NOT invent 60 W: it faithfully repeated a 2000 trade-magazine product notice and preferred it over the maker's current datasheet (TTM: "Power Ave. CW Watts 20"). The saved chat kept only the source label, not the link. That makes this a stale-source mistake, not an invented figure. Every quoted string on the puzzle image was checked word for word against try_4.md, and the question against git show b3a88a81d:scripts/testing/mcp_ab/prompt_link_budget_designed.txt.
Not claimed: that the design burns out. The chat's own table shows +30.8 dBm after the combining coupler (try_4.md line 10); the post says nothing about harm in this design. Not checked: the chat's other parts and figures (HMC408LP3E and HMC407MS8GE datasheets are on analog.com, which blocks our downloads; the Mini-Circuits loss figures live on web data pages). The notes for this sitting record that two other tries disagree about HMC407/HMC408 lifecycle status; not checked.
How it was found: two helper agents checked every part spec in ~38 saved claude.ai answers against maker datasheets (2026-10-07). This was the strongest web-on mistake; the other web-on find was a limiter recovery time (TGL2209-SM "115 ns" against Qorvo Rev H "< 30 ns"). Nine new web-on runs aimed at power ratings (hybrid, LNA survival, 5 W pad) found no clear mistake.
Class version: hand out the chain table and the five source lines, plus the five makers' datasheets, and ask which number is wrong.